Deutsche Bank - Volatility Risk Premium

Provided for the exclusive use of harel@oportod.com on 2020-09-15T09:46+00:00.
DO NOT REDISTRIBUTE
Deutsche Bank
Markets Research
Europe Derivatives Strategy Date

20 April 2017
Derivatives Spotlight
Caio Natividade
Volatility Risk Premium: New caio.natividade@db.com
Dimensions Silvia Stanescu

silvia.stanescu@db.com
Vivek Anand
Today's Derivatives Spotlight delves into systematic options research. It is the
vivek-v.anand@db.com
first in a series of collaborative reports between our derivatives and
quantitative research teams that aim to systematically identify and capture
value across global volatility markets. Paul Ward, Ph.D
paul.ward@db.com
This edition zooms into the volatility risk premia (VRP), one of the key sources
of return in options markets. VRP strategies are popular across the investor Simon Carter
community, but suffer from structural shortcomings. This report looks to simon-d.carter@db.com
improve on those.
Pam Finelli
Going beyond traditional methods, we introduce a P-distribution that best
represents our projected future returns and associated probabilities, based on pam.finelli@db.com
their drivers. Other topics are also highlighted as we construct our P-
distribution, namely a new multivariate volatility risk factor model, our Global Spyros Mesomeris, Ph.D
Sentiment Indicator, and the treatment of event-based versus non-event based spyros.mesomeris@db.com
returns.
+44 20 754 52198
We formulate a strategy which should improve the way in which the VRP is
harnessed. It utilizes alternative delta hedging methods and timing.
Risk Statement: while this report does not explicitly recommend specific
options, we note that there are risks to trading derivatives. The loss from long
options positions is limited to the net premium paid, but the loss from short
option positions can be unlimited.
Figure 1: Volatility Risk Premium: Thinking Outside The Box
Source: Getty Images.
________________________________________________________________________________________________________________
Deutsche Bank AG/London Distributed on: 20/04/2017
20/11/2017 13:35:00
16:58:00 GMT
13:24:01
Note to U.S. investors: US regulators have not approved most foreign listed stock index futures and options for US
investors. Eligible investors may be able to get exposure through over-the-counter products. Deutsche Bank does and
seeks to do business with companies covered in its research reports. Thus, investors should be aware that the firm may
have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a
single factor in making their investment decision. DISCLOSURES AND ANALYST CERTIFICATIONS ARE LOCATED IN
APPENDIX 1.MCI (P) 083/04/2017.
0bed7b6cf11c
Provided for the exclusive use of harel@oportod.com on 2020-09-15T09:46+00:00. DO NOT REDISTRIBUTE
20 April 2017
1. Introduction
Options trading is popular among discretionary and systematic investors, as it
serves multiple purposes in institutional investment portfolios. Today's
Derivatives Spotlight is the first in a series of collaborative reports between our
derivatives and quantitative research teams that aim to systematically identify
and capture value across global volatility markets.
While we assume some technical knowledge, the report is aimed at both

novice and experienced investors alike. For ease of navigation we outline each
of the sections below:
 Section 2 looks at the primary drivers of options returns – our starting

point. We look at both delta hedged and "naked" option return streams,
and find that the former is largely driven by volatility while the latter is
associated with underlying asset returns. The patterns exhibited in the first
principal components point to the existence of a volatility risk premium
(VRP) - the focus of this report.
 Section 3 expands on the VRP and highlights standard measures to tame
strategy drawdowns. We highlight the benefits and shortcomings of cross-
market replication (3.1), market timing (3.2) and buying tail options (3.3),
and allude to new approaches - namely the P-distribution and alternative
delta hedging.
 Section 4 introduces our approach to modeling the P-distribution, the first
new route to improve how we capture value from options markets. It
represents our best estimate of future returns and return probabilities
based on their drivers. Section 4.2 describes how we estimate the mean,
and Section 4.3 provides in-depth coverage of our variance estimates.
These are based on a multivariate risk factor model (4.3.1 and Appendix I),
a regime-dependent multivariate GARCH model (4.3.2 and Appendix II) and
event-specific volatility (4.3.3). The higher moments of our P-distribution
are described in Section 4.4, while Section 4.5 provides a thorough
assessment of how our forecasts compare to a number of benchmarks.
 Section 5 describes alternative delta hedging – the second avenue we take
as we seek to improve VRP strategy returns. We introduce 5 methods from
industry and academia: grid-search (5.1), moving averages (5.2), break
outs (5.3), expected returns (5.4) and a modified version of Whalley &
Wilmott (5.5). All are tested under both simulated data (5.6) and real
market data (5.7).
 Section 6 focuses on building our new VRP strategy. We first fine tune the
signal (6.1), then add alternative delta hedging to the mix (6.2). Section 6.3
deals with timing via our Global Sentiment Indicator, long-term volatility
reversals, and near-term volatility changes. Section 6.4 compares methods,
while Section 6.5 concludes with an application to the S&P 500, Eurostoxx
50 and Nikkei 225.
 Section 7 concludes the report, while Appendix I covers our multi-variate
risk factor model and Appendix II provides a review of our Global
Sentiment Indicator. Our Bibliography is presented at the end.
We believe the VRP is a generic premia, whose dynamics and investment
implications extend beyond multiple markets and asset classes. As such, our
analysis covers all 4 asset classes: equity indices, commodities, currencies and
global Treasuries. We look at options in the S&P 500, Eurostoxx 50, Nikkei
225, Bovespa, EUR/USD, USD/JPY AUD/USD, USD/BRL, Gold, WTI, Corn,
Page 2 Deutsche Bank AG/London

20 April 2017
Copper, 10Y US Treasuries, 10Y Bunds and 10Y JGBs. While their liquidity
characteristics are distinctly different, each should represent a regional hub
inside the asset class and allows us to understand the volatility premia more
holistically.
Most of our market data is sourced internally. We proxy variance swaps using
strike-weighted baskets of European vanilla options. In some sections we use
a purer basket comprised of puts and calls whose strikes expand up to 2.5
standard deviations away from spot. In other sections we use a more realistic
basket based on delta strikes (10-delta, 25-delta and ATM). All options are
rebalanced at expiry. All results prior to Section 6 omit costs; we only apply
them at the end as the VRP strategy gets finalized.
As is often the case in quantitative research, some results are not as

encouraging as we initially hoped. Other results turned out to be better than
expected. In all, we believe that the final recommendations - based on
alternative delta hedging and timing - should improve the way investors
harness this rich source of risk-adjusted returns.
Deutsche Bank AG/London Page 3

20 April 2017
2. The Primary Factors

Understanding what drives returns is key to investing successfully in any Figure 2: Correlations between PCs
market. To the quant investor, the learning process often involves running
and asset returns, implied and
principal component analysis (PCA) on the returns of all available market
realized volatility
instruments and interpreting what the output – the core factors – means.
PC1 delta
PC1 "naked" PC2 "naked"
hedged
returns returns
As we delved into the world of options data, our approach was no different: returns
collect instrument returns and run the PCA. For any given market, we took a A.
RV IV
A.
RV IV
A.
RV IV
Ret. Ret. Ret.
set of relevant calls and puts – 10- and 25-delta, and ATMF – across relevant
E/$ 0.9 -0.3 -0.3 -0.1 0.7 0.8 -0.2 0.4 0.7
maturities – 1, 3, 6 and 12 months – and calculated 2 historical long-only return
streams1: delta hedged and un-hedged. Therefore, each of the 15 options $/Y 0.9 -0.1 -0.2 -0.1 0.8 0.8 0.2 0.4 0.6
A/$ 0.9 -0.5 -0.5 -0.4 0.8 0.8 -0.2 0.6 0.6
markets outlined earlier had 48 return series: 24 for delta-hedged options, and
$/BR 0.9 0.6 0.6 0.4 0.9 0.8 0.1 0.5 0.6
24 for “naked” option returns. We used up to 12 years of returns data, where
S&P
available. 500
0.9 -0.5 -0.7 -0.6 0.8 0.8 0.0 0.3 0.2
SXE 0.8 -0.4 -0.6 -0.4 0.7 0.6 -0.2 0.3 0.4
The PCA output2 set the base for our approach to factor investing in options Nik 0.9 -0.4 -0.5 -0.5 0.8 0.8 -0.1 0.2 0.4
markets. The key findings from the aggregate of all PCAs were that:
Bov 0.9 -0.4 -0.6 -0.5 0.7 0.7 -0.3 0.3 0.6
Gold 0.9 -0.2 0.4 0.0 0.8 0.8 0.0 0.4 0.6
 Spot moves are the chief return driver of “naked” option contracts,
WTI 0.9 -0.3 -0.4 -0.4 0.8 0.8 -0.3 0.4 0.6
explaining circa 63% of total variations in each market, on average. This is
Corn 0.9 -0.1 0.3 -0.2 0.6 0.6 0.0 0.2 0.3
no surprise; the delta exposure, and hence the delta return, dwarfs all
other “Greeks” in the tenors evaluated. Cop 1.0 -0.6 -0.6 -0.6 0.8 0.8 -0.6 0.6 0.6
UST 0.9 0.0 0.2 0.1 0.4 0.7 -0.3 -0.1 0.3
 Changes in volatility were the main source of delta-hedged option returns,
Bun 0.8 0.1 0.2 -0.1 0.6 0.6 -0.2 0.5 0.4
causing 54% of the variations on average. This is the main source of risk
for the volatility trader. JGB 0.7 0.0 -0.1 -0.2 0.6 0.4 -0.6 0.2 0.6
E/$: EUR/USD, $/Y: USD/JPY, A/$: AUD/USD, $/BR: USD/BRL,
SXE: Eurostoxx 50, Nik: Nikkei 225, Bov: Bovespa, Cop: Copper,
 Changes in volatility – implied (delta) vols, in particular – are also the Bun: 10Y Bunds. A.Ret.: annualized returns, RV: realized volatility
(1Y lookback, 3-month half-life), IV: implied volatility. Source:
second driver of naked option returns, accounting for 17% of the Deutsche Bank
variations in the sample data. Lower delta options are more sensitive to
this factor.
 The shape of the volatility smile and the steepness of the term structure
explain an extra 20% of the variations in delta-hedged option returns. In
other words, the volatility trader cannot ignore skew and slope dynamics
either.
 The returns from delta-hedged options are more diverse – the first 2
principal components explain 65% of the variations on average, versus
80% in the case of naked option returns.
1
We used Black & Scholes pricing as these are vanilla European options. We built each surface using SSVI
interpolation and extrapolation. SSVI stands for Surface Stochastic Volatility Inspired. We used the
approach introduced in Gatheral and Jacquier [2013].
2
We used up to 11 years of daily returns, fixed investment notionals in each option (so that return
volatilities were more similar) and estimated the principal components using a correlation matrix instead of
the covariance matrix.

20 April 2017
Figure 3 – an aggregate of the results for all 15 markets – display our Figure 3: Explanatory power and
conclusions visually. The chart at the top shows how dominant the first factor loading by principal
principal components are in each type of return streams, while the bottom
component of option returns
chart shows how they load at each point of the volatility smile, after
Explanatory power
aggregating by expiries. The quasi-homogenous loadings per strike in the 0.70 Delta-hedged 'Naked'
delta-hedged streams, and the strong but inverse looking loadings in the 0.60
“naked” streams support the argument that the primary factor of returns 0.50
heavily affects all options in the surface. As we correlated the historical PC1 0.40
values with more tangible variables, we found an 88% average correlation3
0.30
with underlying asset returns (PC1 of “naked” options) and a 72% average
0.20
correlation with realized volatility (PC1 of delta-hedged option returns).
0.10
Confirming the points above, spot and volatility are the main drivers of
0.00
“naked” and delta-hedged option returns.4 PC1 PC2 PC3 PC4 PC5
PC1 loading
0.08
Next, we analyse the historical developments of these first principal
0.06
components over time in search for potential biases. Bias is what we ultimately 0.04
look for when scrolling through markets with our “quant lenses”, as it lies at 0.02
the core of most systematic strategies. The stronger the pattern, and the more 0
that it can be validated, the more likely that we should succeed in capturing it.5 -0.02
'Naked'
-0.04
Delta-hedged
Take the PC1 of “naked” option returns, for instance. The high correlation to -0.06
-0.08 Option delta

underlying spot markets suggests that the alpha capture will come from -0.5 -0.3 -0.1 0.1 0.3 0.5
forecasting future spot values, just as we do in our delta-one portfolios. The Up to 12 years of data on option returns. See section above for
profit of a “naked” option trade depends on where the asset will be relative to which assets and strikes are included. Source: Deutsche Bank
its break-even at expiry, and the latter reflects a market-implied probability that
the option will expire in-the-money. As such, we should try to estimate our Figure 4: Cumulative standardised
own probabilities that the option will expire ITM, and emit signals based on PC1 values per asset class
where the probability spread is most significant. We call this the probability risk
premium (PRP), and it will be the subject of future research. FX Equities Comm. Treasuries
Today’s report focuses on the PC1 of delta-hedged returns, whose historical

values are aggregated by asset class in Figure 4. It has a distinct characteristic:
regardless of the underlying asset, or asset class, this principal component
trends down over time. This obvious bias suggests that buying delta-hedged
options is a losing proposition over time. And since theory argues that excess
returns should be zero if the “true” volatility is embedded into the option
price 6 , we conclude that the bias originates from implied volatility being 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
generally above what is realized. This bias suggests that selling delta-hedged The standardization went as follows: we took daily returns of each
option for a given market and divided by its 1-year volatility. We
options is a profitable trade over time, and we call it the volatility risk premium then estimated the PC1 loadings using a correlation matrix with 5
years of daily data, and calculated the most recent PC1 value from
(VRP). That said, the “hick-ups” in the series also show that buying options can the loadings. We repeated the exercise daily. We then averaged
the PC1 values for each asset class and plotted the cumulative
be attractive once in a while, which means our capturing process should be values. Source: Deutsche Bank
dynamic; we should not just blindly sell options.
3
Monthly non-overlapping returns, 11 years of data.
4
We did not extend the study to cover, for instance, the drivers of spot returns and of spot volatility. We
recommend Natividade et al [2013] (Section 7) and Natividade et al [2014] (Section 2) for the former; for
the latter, we recommend Mccormick and Natividade [2008], Natividade [2008] and Saravelos and Grover
[2017] in FX and Corradi et al [2006] and Engle and Rangel [2005] in equities.
5
Two examples from delta-one markets illustrate our argument: the PC1 of commodity futures returns,
and the PC3 of a given yield curve. The former exhibits positive and stable autocorrelation over time, which
validates trend following strategies in commodity markets. The latter represents curvature and often
exhibits strong mean-reverting properties, which validates systematic butterfly strategies in liquid IRS
markets.
6
Assuming other theoretical conditions hold.

20 April 2017
3. Exploring the Volatility

Risk Premium (VRP)
The VRP is normally defined as compensation for systematic risk, price Figure 5: Historical VRP distributions
pressures from investor supply and demand, and biased forecasts of future by asset class
empirical volatility.7 We measure it as the distance between current implied and
(3M Implied - 3M Future Realised) / 3M Implied vol
25-75% 33-67% 50%
future empirical volatilities. Therefore, it is no surprise that this metric is not 0.25
only positive8 in every asset class, but more expressive in markets with heavier 0.20
0.15
natural demand for crash insurance – such as equity indices.
0.10
0.05
Figure 5 shows the historical VRP distribution for the aforementioned pool of 0.00
15 assets, bucketed by asset class, using the 3-month horizon as reference and -0.05
-0.10
expressing the VRP as a ratio to implied volatility for better comparison across
-0.15
markets. While built differently, it correlates strongly to the inverse of the PC1 FX Equities Commodities Bond Futs
of delta-hedged returns as described in Section 2. And as we group all asset Up to 12 years of data. We use 15 markets (see Section 1).
Source: Deutsche Bank
classes together, we find that the VRP has been stable and positive in 61% of
the instances. It has also been more strongly pronounced in equities, which
reinforces the demand argument outlined above.
VRP strategies primarily aim to capture that difference between implied and
realized volatilities. Volatility and variance swaps are the standard instrument
of choice for the VRP harvester, though a similar P&L profile can also be
achieved through selling baskets of vanilla options with frequent delta hedging
and capital allocation that is inversely related to the strike level.9 We opt for
such baskets as they give us more delta hedging freedom, which will come
handy later. In this section, we use baskets comprised of 1-month options with
12 strikes that are equally distanced and range from -2.5 to +2.5 standard
deviations away from spot.10 All options were delta hedged daily and rolled at
maturity, and we omitted costs prior to the final section of this report.
Figure 6 shows the cumulative returns of our baskets above, standardized such
that they are equally volatile and therefore easier to visualize. As expected,
they are also similar to the inverse of the PC1s from Figure 4, even though we
used a different weighting scheme to group the individual options.
7
See Ilmanen [2011] for a general review. Other authors who defended similar arguments include Carr and
Wu [2009], Benzoni et al [2010] and Broadie and Johannes [2009].
8
That said, the VRP is far less pronounced in single stock equities than in equity indices and other asset
classes. As argued by Cosemans [2011] and Valenzuela [2014], such difference is attributed to positive
correlation risk premium - how index implied correlations over-estimate future realised correlations. Our
research also validates their argument – see Prasad et al. [2016].
9
This is, in fact, the way a market maker replicates a variance swap, as it generates a stable gamma
profile across different spot levels. For a recent reference, see Derman and Miller [2016].
10
We use the standard deviation of monthly returns as a measure. Some of these strikes may not be
tradable, but suit this exercise as we are looking for a good variance swap replication that is also
computationally efficient.

20 April 2017
Figure 6: Cumulative, volatility-adjusted pre-cost 1-month VRP basket returns

per asset
E/$ $/Y A/$ $/BRL
350 S&P SX50 Nikkei Bov
Gold WTI Corn Copper
300 UST Bund JGB
250
200
150
100
50
-50
-100
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
Labels: E/$: EUR/USD, $/Y: USD/JPY, A/$: AUD/USD, $/BRL: USD/BRL, S&P: S&P 500, SX50: Eurstoxx 50, Nikkei: Nikkei 225, Bov:
IBOVESPA, UST: 10Y US Treasuries, Bund: 10Y Bunds, JGB: 10Y JGBs. Source: Deutsche Bank
In general terms, the investor who sells an option is selling some form of
insurance. By selling the whole smile, we are selling market insurance much
more broadly; significant moves in any direction are likely to hurt us. Most
assets involved fall faster than they rise, and therefore “significant moves”
typically coincide with bearish markets. In 201211 we launched the DB Global
Sentiment Indicator (GSI), a variable that captures market risk appetite and
whose details are described in Appendix II. Figure 7 shows how it relates to
our VRP baskets: irrespective of the asset class, falling VRP returns are
associated with rising GSI levels, indicating higher risk aversion.
11
See Natividade and Chen [2012b].

20 April 2017
In essence, the VRP strategy is pro-cyclical; it underperforms when markets are Figure 7: 2M cross-asset VRP returns
risk averse – most commonly when macroeconomic volatility is high and
versus DB’s Global Sentiment
output is contracting. Needless to say, this is a shortcoming. It means that VRP
Indicator
strategies correlate positively with the static – equity, credit – premia that
dominate institutional portfolios and, as such, that they also underperform 2M VRP returns / Vol of 2M VRP
FX Equities
returns
when core portfolios are suffering. Adding insult to injury, VRP drawdowns are 2
C'dys Bond Futs
notable for their magnitude; the 4 months following the Lehman Brothers 1.5
announcement in Sep-08 wiped out the previous 6 years of returns in a typical
1
tradable S&P 500 VRP strategy12, and 8 years of returns in short USD and EUR
tradable swaption strategies.13 0.5 R² = 0.1423
There are 3 popular methods that seek to improve the profile of VRP returns: -0.5
cross-market replication, factor timing and buying tail options.

-1
-1.5
3.1 Cross-market Replication 2M changes in the
-2 DB GSI
-0.4 -0.2 0 0.2 0.4
Cross-market replication is popular in the quantitative investment community, The 2-month VRP returns comprise all 15 markets, aggregated by
asset class. The reader may be intrigued as to why we chose the
especially when dealing with generic delta-one factors such as Momentum 2-month time window. Shorter windows produced a sparser
relationship (partly related to Footnote 15), while longer windows
and Carry. Since the number of independent trades influences the Sharpe ratio had too few datapoints. Source: Deutsche Bank
of any strategy, the quant investor will seek to maximize breadth by replicating
the same signal across a large number of markets.14
While attractive, the idea has a shortcoming when applied on the VRP. As is
often the case with other base factor strategies applied to a limited asset pool,
the VRP signal has limited breadth because it captures a “global” driver that
explains the majority of moves in all instruments available.15 Investor aversion
– the abstract risk whose premium is captured by short volatility strategies – is
common across asset classes, leading to a strong link between option returns
even in markets where spot returns are not highly correlated. Figure 8 shows
exactly that: the VRP return correlations in the bottom triangle are, on average,
stronger than the spot return correlations in the upper triangle.
Cross-market replication improves the return profile – the expected shortfall

shrinks by almost one third in our backtest16 – but it is not enough on its own,
as we seek a better overall improvement.
12
We use a standard, tradable index for this comparison: the DB Equity US Volatility Carry Index
(DBGLSVEU Index on Bloomberg). This strategy sells 2M and 3M straddles on the S&P 500, delta-hedged
daily. Costs are included.
13
We use a standard, tradable index for this comparison: DB ImpAct (DBIP3BE Index and DBIP3BU Index
on Bloomberg). This strategy sells 3M10Y USD and EUR swaption straddles every week, delta-hedged at a
given frequency. Costs are included.
14
This follows from the Law of Fundamental Active Management,
where Sharpe  IC  breadth. IC, the information coefficient, represents signal forecast
accuracy and breadth stands for the number of independent investment decisions. See Grinold & Kahn
[1999].
15
See Baz et al [2015], and Natividade et al [2016a]. These references also show that in order to preserve
signal entropy, base factor strategies are best implemented in time series form - in other words, without
committing to an equal number of long and short trades.
16
We use the same VRP replicating basket described earlier in the section and data since 2002. Shortfalls
are defined as the average of monthly returns equal to or lower than the 5th percentile of monthly returns in
the sample. The average shortfall is estimated by averaging the shortfalls of all 15 VRP strategies,
weighted by volatility, whereas the shortfall of the average is estimated by grouping all VRP strategies
(weighted by volatility) and then calculating the monthly shortfall.

20 April 2017
Figure 8: Correlations between underlying markets – VRP returns (bottom triangle) and spot returns (top triangle)
E/$ Y/$ A/$ BRL/$ S&P SX50 Nikkei Bov Gold WTI Corn Copper UST Bund JGB
E/$ 1.00 0.23 0.72 0.50 0.39 0.29 0.24 0.48 0.43 0.51 0.37 0.51 -0.14 -0.32 0.03
Y/$ 0.55 1.00 0.17 -0.09 -0.24 -0.19 -0.51 -0.10 0.37 -0.04 -0.02 -0.20 0.47 0.32 0.41
A/$ 0.73 0.63 1.00 0.76 0.60 0.49 0.39 0.72 0.46 0.60 0.55 0.66 -0.02 -0.31 -0.12
BRL/$ 0.65 0.55 0.85 1.00 0.47 0.42 0.35 0.62 0.41 0.60 0.46 0.64 -0.20 -0.37 -0.16
S&P 0.65 0.52 0.82 0.79 1.00 0.91 0.78 0.71 -0.08 0.36 0.33 0.58 -0.20 -0.40 -0.27
SX50 0.53 0.46 0.67 0.71 0.71 1.00 0.74 0.70 -0.04 0.21 0.25 0.48 -0.09 -0.35 -0.24
Nikkei 0.61 0.67 0.68 0.66 0.72 0.52 1.00 0.51 -0.24 0.22 0.12 0.47 -0.25 -0.30 -0.40
Bov 0.52 0.60 0.66 0.65 0.67 0.51 0.52 1.00 0.26 0.51 0.34 0.55 -0.13 -0.42 -0.29
Gold 0.45 0.45 0.45 0.46 0.53 0.37 0.40 0.62 1.00 0.30 0.29 0.26 0.09 -0.09 0.09
WTI 0.58 0.51 0.52 0.54 0.62 0.39 0.61 0.57 0.42 1.00 0.24 0.65 -0.28 -0.50 -0.21
Corn 0.41 0.32 0.41 0.34 0.40 0.30 0.50 0.39 0.31 0.41 1.00 0.42 0.03 -0.18 -0.03
Copper 0.42 0.31 0.61 0.60 0.71 0.51 0.49 0.46 0.48 0.47 0.26 1.00 -0.23 -0.43 -0.12
UST 0.31 0.18 0.17 0.18 0.17 0.36 0.21 0.12 0.08 0.18 0.12 0.04 1.00 0.64 0.41
Bund 0.37 0.23 0.17 0.15 0.17 0.33 0.13 0.08 0.20 0.16 0.16 0.02 0.60 1.00 0.56
JGB -0.03 0.04 -0.20 -0.25 -0.24 -0.06 -0.04 -0.28 0.10 -0.14 -0.06 -0.13 0.10 0.25 1.00
Correlation of 2-month non-overlapping returns, Note that we use FX/USD and not USD/FX to make VRP and spot market correlations more comparable. 11-year history. Source: Deutsche Bank
3.2 Market Timing
Timing the VRP strategy is also popular and, if the backtests serve as an
indication, it can be successful as well. The premise is that if there is a strong
relationship between strategy returns and global risk appetite, and if we can
time the latter, we should be able to reduce the drawdowns in the former.
To test it, we created a linearly continuous leverage mechanism to the VRP

strategies above using the same GSI as before. The higher the GSI, the less we
sold of our options baskets and vice-versa; in other words, we de-levered
during periods of risk aversion and levered up when markets were calm. The
leverage ratio is defined as Lt ,t  h  2  GSI t  2 , and is therefore
constrained such that L  [0,2] . The leverage changes every 2 weeks, and it
has an average of 1.17
17
Unlike in prior applications, we apply the GSI leverage on a bi-weekly basis instead of daily. We
evaluated the sensitivity to changes in leveraging frequency and found that "high" frequencies (daily) and
very low frequencies (bi-monthly) yield worse results. Daily rebalancing incurs more signal noise and factor
reversal risk: when the GSI is at extreme levels, next-day strategy returns are at greater risk of going
against the timing indicator than next-week or next-month returns. We also knew that once we attach
transaction costs to this exercise, daily rebalancing becomes unfeasible. On the other hand, very low
frequencies, such as bi-monthly, are not adaptive enough. Of the frequencies tested, the highest risk-
adjusted returns came when using monthly rebalancing. We chose bi-weekly for illustration purposes
alone. These results do not include costs - we leave that to later in this report.

20 April 2017
History suggests this is an attractive idea. Figure 9 plots the effect from Figure 9: Backtested monthly
leveraging our strategy according to the GSI. The X-axis represents the original shortfalls – cross-market VRP
backtested monthly shortfall18 in the VRP strategy for each of our 15 markets,
strategy
labeled by asset class, while the Y-axis shows the new shortfall after applying
Shortfall without timing
the technique above. The sample covers 11 years of data. While we only -2 -1.5 -1 -0.5 0
0
capture a handful of crisis periods, the results are encouraging in that most of -0.2
the effect comes from taming our drawdowns. They also concur with our -0.4
-0.6
positive experience timing risk-sensitive, pro-cyclical signals in the past – -0.8
notably the FX Carry trade, as shown in Appendix II. -1
-1.2
-1.4
Timing =
3.3 Buying Tail Options No Timing -1.6
-1.8
FX Equities
-2
C'dys Bd Fut Shortfall with timing
Practitioners have also been fond of buying “tail options” to improve VRP
We define shortfall as the average monthly return using months
strategy drawdowns. The method is straight forward: sell volatility at the that qualified as 5th percentile of worst months in the sample for
each strategy. Source: Deutsche Bank
nearby strikes as per original strategy but buy volatility at distant strikes so as
to flip the gamma exposure in the event that spot moves significantly.19 Ideally,
if the spot market has fallen aggressively, and volatility has therefore risen, our
net gamma exposure should have gone from short to long. In principle,
therefore, this should improve VRP return drawdowns.
In order to test the idea, we modified our option baskets slightly. Instead of
only using 1-month options as earlier in this section, we used 4 expiries: 1, 3, 6
and 12 months. We also changed our strikes; instead of 12 fixed-percent-
distance strikes equally spaced in steps of ½ of a standard deviation, we used
6 fixed-delta strikes with distance defined in delta steps: 10-delta, 25-delta and
ATMF puts and calls. These revised maturities allow us to capture different
theta decay profiles, while using delta strikes allows us to better compare
cross-maturity strikes.
The goal of this backtest was to evaluate the sensitivity of VRP strategy returns
to using one set of specific strikes as hedge. In other words, we went short
one unit of every option described above except for those with a pre-specified
delta. In the latter case, we would go long one unit (of each) instead. The
benchmark in this case is to go short all options, with no protection. As before,
we delta hedged daily and did not include costs.
18
We define shortfall as the average monthly return using months that qualified as 5th percentile of worst
months in the sample for each strategy. In other words, the average return in the 7 worst months in the
sample.
19
A variation of this idea is to only sell delta-hedged straddles. In this case, aggressive spot moves will
also have a smaller adverse impact on the VRP strategy because much of the straddle gamma will have
been eroded.

20 April 2017
The results here are also encouraging. Our Sharpe ratios rose and, more Figure 10: Shortfall / average returns
importantly, our drawdowns improved once we balanced our short exposures (VRP with tail options) minus
with a few long positions.
shortfall / average returns (VRP
benchmark)
But the improvement also depended on which strikes we chose to go long; in
most instances, it is only the low delta puts – the tail – that helped. Figure 10 10p 25p St'dle 25c 10c
shows the ratio of average monthly shortfall to monthly returns in each of our E/$ 0.44 0.0 0.3 0.0 0.0
15 markets, expressed as a difference to the ratio achieved using the Y/$* 0.7 1.3 -1.5 -0.4 0.5
benchmark strategy. The more negative the number, the better – it means that A/$ -4.9 -6.8 1.1 53.6 5.9
a typical drawdown erased less return months.
BRL/$* -2.7 -2.7 3.5 18.7 3.0
S&P -2.4 -0.1 0.1 3.1 1.2
Take the Nikkei as an example. In the benchmark strategy, the average
SX50 -2.4 -3.2 -0.7 9.6 3.2
monthly shortfall erased 25 months’ worth of average monthly returns. If we
modified the strategy so as to be long 10-delta puts in the 4 maturities while Nikkei -7.0 -6.2 -4.0 34.1 8.8
still short all the other options in the surface20, the ratio would have fallen from Bov 0.1 9.6 1.3 -2.0 -0.9
25 to 18, an improvement of 7 months (hence -7 in the table). Parallel Gold -1.7 0.2 2.8 1.2 -0.3
examples can also be made with low delta puts in most other markets, WTI -0.4 2.3 0.0 0.1 -0.3
particularly where asset returns exhibit strong negative skew.21 While one may Corn -9.0 -0.3 -4.9 7.1 23.1
argue that the return profile should improve further once we optimize our Copper -2.5 -1.2 -0.6 3.9 1.7
choices for theta decay efficiency 22 , these preliminary results are already UST 1.7 3.8 3.6 8.1 -4.5
encouraging. Bund -0.8 -0.6 3.4 2.1 -0.1
JGB -0.7 -0.9 8.9 2.0 -0.6
3.4 A New Way of Thinking * Flipped USD per unit of currency for better comparison with
other markets. Source: Deutsche Bank
These methods are encouraging, but narrow in scope. They focus mostly on
controlling strategy drawdowns, and do so by either timing risk aversion, or by
hedging or diversifying against it. In all 3 cases, we use prior knowledge and
experience to define what risk aversion means and how it manifests itself
across markets. We assume, for instance, that risk barometer (the GSI) is
accurate, and that only the left tail of asset returns should be of concern to the
VRP investor. Further, true instances of risk aversion are scarce - only a
handful in our backtest window - which makes our results even more
dependent on these priors.
In the remaining sections, we seek to introduce methods that are distinctly

different from the ones described above. First, we will not assume that
volatility risk premia will always be positive; we will compare the forecasts
from our real world distribution – introduced in the next section – with market-
implied volatility so as to define the sign and quantity of our VRP positions.
Second, we will use our knowledge of the underlying asset returns – direction
and trendiness – to delta hedge better. We start with the former in Section 4,
as we look at the whole P-distribution instead of asset variance alone. We then
move onto delta hedging in Section 5.
20
For clarity: we went long 10-delta 1M, 3M, 6M and 1Y put options, allocating 1 JPY unit to each. At the
same time, we went short 10-delta calls, and 25-delta and ATMF calls and puts in the same maturities,
allocating 1 JPY unit to each as well. The capital allocation is therefore 4/24 into long positions, and 20/24
into short positions.
21
The versions that are long 10-delta puts outperformed except in a few instances - most notably in
Treasury futures, a typical safe haven asset, where the version long 10-delta calls outperformed. This
reinforces the idea of being long the area of the smile most likely to outperform in the case of risk
aversion. That said, a few cases are less intuitive – notably with straddles in JPY/USD and Bovespa calls.
22
We did not test theta decay optimisation thoroughly, but we tried a rough version that bought only 1Y
10-delta puts (4/24 units) and sold all other options (20/24 units). This version also outperformed the
unconstrained VRP strategy, but underperformed the version that also sold 10-delta puts of other expiries.

20 April 2017
4. The P-Distribution
As is the case in all derivatives markets, profiting from options trading requires
taking a view (of the future) that differs from what is implied by the market
price. But options markets carry a unique advantage – their prices contain a far
more granular view of the future than futures and forward markets do. By
mapping the volatility smile into a risk neutral distribution, one can estimate
the market-implied probability that spot will be at any level by horizon date, in
addition to estimating how volatile or leptokurtic the market “expects” the
returns to be. 23
This abundance of market-implied views is valuable to the quantitative investor

because it increases the number of uncorrelated trades available using the
same type of instrument. The investor is no longer constrained to a bullish or
bearish position versus the futures price; she can instead trade various options
strikes to express higher order views. She can even express a view that is – in
principle – indifferent to spot levels, as does the VRP harvester.
The most holistic way to identify option trading opportunities is, in our view, by
estimating a distribution of future asset returns and comparing it with the
market-implied distribution. In other words, comparing our P-distribution with
the market’s Q-distribution. This sets the base for how we will assess value
across strikes, expiries and markets. While today’s report will only focus on the
VRP – differences in the second moment of the two distributions – we will
introduce the whole approach as it serves as base for future reports. This
section, therefore, introduces our P-distribution.
4.1 P-distribution Basics
The P-distribution is our subjective assessment of what future asset returns will
look like. Estimating it is a challenge. We have to account for realistic financial
market assumptions while also staying computationally efficient. With that in
mind, there are 4 aspects to our approach:
 We use our delta-one portfolios to estimate expected returns - the first

moment of our P-distribution. Each portfolio contains some additional
predictive power about future returns across horizons. The portfolios were
built as part of our Quantcraft series: Trend, Carry, Value, Sentiment and
Macro Factor investing.
 We combine a multivariate risk factor model, in addition to a univariate
model, to estimate asset variance. This approach utilises both cross-
market and idiosyncratic information to forecast the second moment of
our P-distribution, and is in line with how quant investment portfolios are
built.
23
As with other derivatives, market-implied views are derived through arbitrage-free relationships that
most often reflect the cost of hedging the position and supply-demand dynamics, as opposed to traders’
subjective views of the future. This difference is often the source of opportunity.
Specifically to options markets, Carr and Madan [2001] argue that the optimal trade for an investor requires
3 ingredients: (1) her beliefs regarding future outcomes (the P-distribution), (2) her risk preferences (level of
risk aversion), and (3) market prices.

20 April 2017
 We account for volatility jumps by segregating between event and non-

event volatility. The former relates to known economic releases and policy
decisions in each country.
 We also account for the influence of global sentiment regimes when
calibrating our parameters.
 Finally, we apply an analytical, moment-based approach to estimate the
higher moments and ultimately complete our P-distribution. This approach
has the advantage of being computationally efficient.
Our returns process is defined as follows:
rt ,i   t ,i  Bi Ft   t ,i
Et 1 rt ,i    t ,i
Vart 1 rt ,i   Bi  t| t 1 BiT  t t 1
where:
 t ,i is determined exogenously by our delta-one signals. It is non-

stochastic but varies over time. It is described in more detail in Section 4.2.
 Bi is a vector of factor loadings: Bi = (β1, i, β 2, i,... β K, i)T , for i = 1:N assets
and k = 1:K factors. Ft is a vector of de-trended24 observed risk factor
returns. Both are described in more detail in Section 4.3 and Appendix I.
  t ,i  ht , z t ,i is a disturbance process such that  t ,i ~ 0, ht ,  .
i i
  t is a time-varying factor covariance matrix, our systematic risk. It is

described in more detail in Section 4.3.
 t is a time-varying diagonal 25 covariance matrix of error terms. It
represents asset-specific risk.
4.2 P-Distribution: the Mean
As the P-distribution should represent our "best guess" of future values of

spot, it must incorporate all our knowledge about what drives spot returns. So
far, our work indicates five categories of drivers – trendiness, valuations, carry,
market sentiment and macroeconomic developments. The respective signals
are grouped through the portfolios highlighted in Section 4.1.
For any horizon, our forecasted expected return,  , is computed as follows:

1. The sign of our forecast reflects the average of the standardised weights26
from each portfolio.
2. The magnitude of our forecast reflects both the magnitude of recent asset
returns and a propagation function. We use a variance ratio as the latter.27
Mathematically,
24
A necessary requirement for Et 1 rt ,i    t ,i .
25
We assume the disturbance terms are uncorrelated.
26
We divide the current weight by its recently volatility. The absolute weights are bounded at 2.
27
We use an anchored long-term window to estimate the variance ratio term structure.

20 April 2017
 t ,i  f  rt ,i , rt 1,i ,..., r0,i   w

M
~
t , m ,i
m 1

where f  is an exponential moving average of absolute asset returns.
~ is the standardized time-t weight of asset i associated with portfolio Y .
w t , m ,i
m
In other words, our forecast reflects how bullish or bearish our delta-one
portfolios are now versus the past, and whether spot is likely to trend or mean-
revert over a given horizon.
4.3 P-Distribution: the Variance
The variance is the second step in estimating our P-distribution. It is the main
quantity of interest for the VRP harvester and hence a key focus of this paper.
Our approach must reflect real-world characteristics, while remaining
computationally efficient.
To model  t we apply a multivariate factor constant conditional correlation

GARCH model – CCC MGARCH.28 The building blocks are as follows:
 Common factor vs idiosyncratic risk: we discriminate between systematic

and asset-specific risk.
 A multi-regime approach: we use the aforementioned DB Global Sentiment
Indicator - GSI29 to exogenously incorporate different risk regimes into our
volatility model.
 Event-based jump risk: we also incorporate the impact of scheduled macro-
economic events into our variance forecasts.
4.3.1 Factor model
Factor models are commonplace in the quantitative investment community.
They carry two30 main advantages:
 Computational efficiency. In essence, factor models reduce dimensionality.

Instead of modeling the variations and co-variations of N assets
individually, we model those of K factors instead. By default, K < N. Factor
co-variances eventually map back into asset co-variances through the
latter’s sensitivity to each factor.
 Estimation robustness. In a world with increasingly integrated markets, a
factor model captures the common drivers – the factors – of cross asset
28
The fact that financial returns are not well described by iid normal distributions has been long
documented in the literature, starting as early as 1960s – see, for example Mandelbrot [1963]. Volatility
clustering – i.e. positively autocorrelated variances – is well established. The seminal papers of Engle
[1982] and Bollerslev [1986] enable the modeling of such volatility clustering. Jumps in the price and or/
the volatility of assets (around events) is another well documented phenomenon. Furthermore, multi-
regime GARCH models have been advocated in the relatively recent academic literature – see, for
example, Haas et al.[2004a, 2004b], Marcucci [2005], Alexander and Lazar [2006, 2009]. All these features
have been incorporated in the volatility model we propose and describe below. CCC-MGARCH stands for
Constant Conditional Correlation Multivariate Generalised Autoregressive Conditional Heteroskedasticity.
Factor models for modeling the volatility of assets, particularly of large portfolios, have been long
advocated in the academic and practitioner finance literature. Here we are combining the approaches
proposed in the seminal papers of Engle et al. [1990] on factor- ARCH, and Bollerslev [1990] on the
constant conditional correlation GARCH model, later developed to allow for dynamic correlation by a series
of authors, including, for example Engle [2002] or Tse and Tsui [2002].
29
See Natividade and Chen [2012b] for the methodology behind the construction of our Global Sentiment
Indicator (GSI). According to this regime indicator, we distinguish between three risk/market sentiment
states, namely: risk-seeking, intermediate and risk aversion.
30
Another, albeit secondary advantage is that a factor model can also incorporate spill-overs between
volatilities of different assets.

20 April 2017
variations better. As has been argued in the literature 31 , the inter-

dependence between the various volatilities may be due to an underlying,
or a set of underlying common factors. Sections 2 and 3.1 also illustrate
this argument; not only the common variations in delta-hedged returns
were highly linked to volatility, but those volatilities were highly correlated
across different markets as well.
We model systematic risk through K = 11 global risk factors, summarized in Figure 11: Factors used for
Figure 11. We divide them into 2 categories: macro and investable. The former Multivariate Variance forecasting
covers asset sensitivity to broad macroeconomic dynamics – namely growth
and inflation. The latter represents directly investable portfolios, either static Factor type Summary
(long-only) or dynamic (long-short). Our model is based on the Arbitrage US inflation
Pricing Theory (APT), an extension of the CAPM where multiple factors are (proxied by TIPS
used. Asset exposure to each factor is estimated using a stepwise robust vs USTs)
regression, a method that tackles collinearity and outlier effects. Our model Macro
Global growth
covers 80 assets across multiple asset classes and is updated daily. It is
(proxied by
covered in detail in Appendix I.
equities)
Having outlined the factors, we now relate them to our final asset variance Equities: EM
forecast. Recall from Section 4.1 above that the time t variance of asset i is
FX: USD/G10,
given by:
USD/EM
Vart 1 rt ,i   Bi  t| t 1 BiT  t t 1 Investable: Commodities:

  static energy, metals,
systematic risk idiosyncratic risk
agriculture
Rates: global 10Y
In the context of our factor CCC-MGARCH, we can write t as the following Treasuries
matrix product:
Cross-asset
Investable:
 t  Dt  t Dt dynamic
Momentum
where Dt  diag ht ,k  is a diagonal matrix of time-varying factor standard Cross-asset Carry
We use the terms “Macro” and “Investable” as per literature.
is a KxK factor correlation matrix.   diag h  is a

1k  K Source: Deutsche Bank.
deviations and  t t t , i
diagonal matrix encapsulating asset specific risk. 1i  N
: The estimation of our CCC-MGARCH includes the following steps:
 Step1: estimate the factor variance terms (i.e. the elements of Dt ) via a
GARCH-type model of choice – in our case, a multi-state model where the
states are exogenously given by our GSI. This is described in Section 4.3.2.
 Step 2: Repeat Step 1 for the residual variances (i.e. the elements of t ).
 Step 3: Estimate the correlation matrix  t 32 ; in the context of a CCC
model,  t   , i.e. correlations are assumed constant through time.
4.3.2 The exogenously-determined multi-state GARCH model
As the seasoned investor is aware, asset volatility follows multiple regimes
over time. Capturing regime-shifts quickly enough is key. In our view, the GSI
suits that task better than other standard choices such as historical returns
alone.
31
See for example Bauwens et al. [2005] – Chapter 5 in particular and references therein – , Corradi et al.
[2006] or Engle and Rangel [2005].
32
With the sole condition that this is a positive definite matrix, such that the positive definiteness of t
is ensured (i.e. variances are guaranteed to be positive such that volatilities are defined)

20 April 2017
The variance of the k-th de-meaned factor Fk,, conditional on information

available at t-1, is given by:
vart 1 Ft , k   ht , k  k   k Ft 21, k  k , st 1 Ft 21, k   k ht 1, k
The variance terms ht , k (k = 1,2…., 12) will represent the diagonal elements of
 t . We also use the same specification for the idiosyncratic variances,
namely:
vart 1  t ,i   ht , i  i   i t21,i  i ,st 1  t21,i   i ht 1, i
where st is an observable random variable. At any time t, we assume that st is

equal to either 1, 2, or 3, denoting low, medium or high risk regimes, 33
respectively, depending on the GSI value at t.34
For parameter estimation via maximum likelihood, as well as forecasting

higher moments and approximate distributions (see Section 4.4), the following,
equivalent expression for ht,k will be used:35
ht ,k   k   k Ft 21,k  1,k Ft 21,k I t11 ,k  2 Ft 21,k I t21,k   k ht 1,k

i 
where I t ,k is an indicator function, taking the value of 1 if st = i and zero
otherwise, with i  {1,2} .
4.3.3 Event versus non-event volatility

It is well established that asset returns can be “jumpy”, and we must account
for this feature if we want to make our variance forecasts more realistic. While
the standard approach involves a compound Poisson process, we opted for an
alternative method that targets macroeconomic events – the periods when
those jumps are more likely to occur. This bespoke approach is more
straightforward and computationally efficient in helping us estimate the P-
distribution as a whole. 36 Our approach follows Natividade et al. [2011], which
is in turn inspired by Bauwens et al. [2005]. We apply it to both factor and
idiosyncratic variances.
33
In Natividade and Chen [2012b], a Gaussian mixture of historical GSI values pointed to 3 as the optimal
number of regimes to describe the series.
34
Identification requires one of the k ,s t 1
parameters, with st-1= {1, 2, 3}, to be equal to zero (or
equivalently that α is equal to zero). In the equivalent specification used below, we implicitly assume
k ,3  0 . The same treatment applies to i ,s t 1

.
35
A similar expression will be used for the estimation of idiosyncratic risk.
36
See Cont and Tankov [2004] for a comprehensive review of the parametric treatment of jumps in
financial returns in continuous time. Using discrete time modeling, Guegan et al. [2013] propose a GARCH-
type model augmented with compound Poisson jumps in the returns. Their proposed approach is thus
related to the (continuos-time) jump diffusion model introduced in the seminal work of Merton [1976], and
also to the Bernoulli diffusion model (BDM) – see Honore [1998] and Ball and Torous [1983]. The jump-
diffusion modeling approach would have the advantage of modeling both (scheduled) events as well as
non-(scheduled) event-driven jumps. It would however have the disadvantage is that the estimation of
such models is not particularly straightforward, especially in our multivariate setting and given our aim of
forecasting cumulative variances and higher moments over various horizons.

20 April 2017
Figure 12: Macroeconomic events covered

US Eurozone Japan Australia
FOMC GE IFO Balance of Payments: Employment Change
Current Account
Fed Minutes PMI Composite Japan Buying Foreign GDP
Bonds
NFP HICP Japan Buying Foregin CPI
Stocks
GDP Unemployment Foreign Buying Japan Retail Sales
Bonds
ISM Manufacturing PMI Manufacturing Foreign Buying Japan Labour Price
Bonds
ISM Services ECB Policy rate decision CPI Policy Rate Decision
CPI France PMI GDP Brazil
Manufacturing
Ret. Sales France PMI Services Tankan Survey Policy Rate Decision
Cons. Conf. Germany PMI BOJ Policy rate decision CPI
Manufacturing
Home Sales Germany PMI Services
House Starts Italy PMI Manufacturing
Jobless Claims Italy PMI Services
Source: Deutsche Bank; Certain holidays (e.g. London. US) were also treated as events where applicable (i.e. market not closed). Country-
specific holidays were also excluded for each asset.
Our approach to forecasting the total variance of asset returns combines the
modeling of time-varying volatility from Section 4.3.2 with non-parametric,
event-driven jumps. The process involves the following steps:
 Step 1: we “clean” the returns data by removing the impact of event-

driven jumps. We estimate the impact of a scheduled event by dividing
asset variance on the day of the event by asset variance during non-event
days. Figure 12 summarizes the events we cover. These ratios will turn into
multipliers in the future once we have forecasted “clean” variance.37 Figure
13 illustrates how significant our calendar events are for a selection of
assets; the fact that the non-holiday events generally have ratios in excess of
1 validates our approach.38
 Step 2: we estimate the parameters of the GARCH process for (clean)
variance using the event-adjusted returns series obtained in Step 1.39
 Step 3: we compute forward h-step ahead “clean” daily variance forecasts
based on the parameters estimated in Step 2.
 Step 4: we re-insert the impact of scheduled events by using the ratios
from Step 1 as multipliers to the (forward-starting) daily variances from
Step 3. Finally, we integrate the forecasts so as to arrive at the final
forecast of cumulative, annualized volatility for a given horizon.
37
Factor models use US events in the modeling of systematic (factor) variance, whereas events pertaining
to the other countries are used in the modeling of idiosyncratic risk.
38
We note that the most significant impact across all assets is that of NFP announcements. Using high
frequency data for EUR/USD, Chen, Natividade and Wang [2011] also found this to be one of the events
with the greatest impact, second only to FOMC announcement. We note however that the event impact
estimation framework here differs in two important ways from that in Chen, Natividade and Wang [2011]:
1) the above variance impact is computed for factor variances rather than asset variances directly, with the
events impact for the latter subsequently computed based on the former as explained in the notes to
Figure 13; 2) daily (rather than high frequency data) is used.
39
Event-adjusted returns are equal to the observed returns divided by the square root of the variance
multiplier from Step 1 on event days, and left unchanged otherwise.

20 April 2017
Figure 13: Event influence on return variance

2.5 Variance multiplier (US NFP) Event = Non-Event
More volatile
during NFP
1.5
1
Less volatile
during NFP
0.5
0
S&P 500 SX5E Nikkei Bovespa E/$ $/Y A/$ $/BRL Gold WTI Corn Copper UST Bund JGB
2.5 Variance multiplier

More volatile
Equities FX Com Bd Futures Event = Non-Event

during NFP
1.5
1
Less volatile
during NFP
0.5
0
Jobless Claims NFP GDP ISM Fed Min ISM Serv FOMC House Starts Home SalesCons Confidence CPI Ret. Sales
Source: Deutsche Bank; We show the US events influence on systematic asset variance. The event impact at asset class level is computed based on the events’ impact on factor variance and our factor model as
follows: for each asset and each event, we compute the sum of (squared factor beta * factor variance multiplier)/ sum of squared betas. For each event, we plot the average variance multiplier within each asset class.
4.4 P-Distribution: the Higher Moments
As the reader may suspect, one is unlikely to find a closed-form distribution

that accurately reflects all the characteristics of asset returns. But we can get
somewhere close using a moment-based approximation. For this, we need two
ingredients:
 A moment-based approximation method.

 Tractable, computationally efficient higher moment forecasts.
Regarding the first ingredient, there is no unique recipe, but the Johnson S.U.
distribution40 is a convenient option. Not only it is leptokurtic, as are financial
returns, but its four parameters 41 can be fit via efficient, quasi-analytical
algorithms such as in Tuenter [2001].
A random variable X is said to follow a Johnson SU distribution if:

 x   JSU 
Z   JSU   JSU sinh1   ,
  JSU 
where Z is a standard normal variable, and sinh−1 is the inverse hyperbolic sine
function: sinh1 x   ln x  x 2  1  . The 4 parameters γJSU, δJSU , ξJSU and

λJSU may be estimated using the moment-matching algorithm described in
40
See the seminal paper by Johnson [1949] where these distributions were introduced.
41
While the four parameters do not have an intuitive interpretation, they map explicitly into the first four
moments of any target distribution. Here we are mapping the first four moments of h-period aggregated
GARCH returns.
Although flexible, the main disadvantage of this approach is that a Johnson SU distribution is not
guaranteed to exist for any set of mean, variance, skewness and (positive) excess kurtosis. That said, other
distributions from the same family can be fit in this case.

20 April 2017
Tuenter [2001]. Using the above relation between Z and X, the cdf of X can be
written as follows:
  x   JSU 
F x   FJSU  x     JSU   JSU sinh1   

   JSU 
Differentiating the above, we get the pdf of a Johnson SU:
 JSU  1  x   JSU 
2

f x   f JSU x   exp     JSU   JSU sinh1  
1
 
 2  
 JSU 2  x   JSU 
2
    JSU  
1   
  JSU 
where fJSU(x) is the Johnson SU fit to f, the pdf of h -period aggregated returns
R t+h,i.
One important advantage of our proposed modeling framework (i.e. the factor
CCC-MGARCH above) is that the higher moments of R t+h,i.- the second
ingredient in the approximation framework above – can be obtained in quasi-
analytical form for any horizon h.42 The proposed moment-based distribution
approximation method together with the computationally efficient method of
forecasting higher moments 43 for any horizon complete our P-distribution
modeling framework.
4.5 Assessing Accuracy – Mean and Variance Forecasts
Having described the mechanics of our P-distribution, we now evaluate its

(historical) accuracy in predicting the future relative to standard benchmarks.
We focus on 2 forecasts - the mean and volatility - as these will be most useful
for extracting the variance risk premium. The former should help us delta
hedge better, while the latter will help decide the direction and size of our VRP
positions.
4.5.1 Forecasting Accuracy: the Mean

We start by evaluating the directional accuracy of our forecasts of expected
asset returns. We focused on direction as opposed to intensity as the latter is
less relevant for delta hedging. At every evaluation date, we took the sign of
the mean of our forecasted return distribution and evaluated whether future
cumulative returns were of the same sign (our hit ratio). We evaluated forecast
accuracy according to 3 horizons: 1 week, 2 weeks and 1 month. These near-
term horizons were chosen so as not to deviate too much from traders’ typical
risk management constraints.
42
For certain (univariate) GARCH models, (quasi)-analytical formulae for the higher moments (of the
returns and variance processes) are already derived in the literature – see, for example, Alexander et al.
(2011). As Simonato (2013) explains, this reduces significantly the computational time. Here, we propose
applying his approach to our factor CCC-MGARCH modeling setting.
43
As specified earlier, quasi-analytical formulae for the higher moments can be derived in our framework.
These formulae will be used in our future research involving the P-distribution, but as they are outside the
scope of the present paper are skipped here.

20 April 2017
We used the same set of 15 representative assets as before, representing

equity indices, currencies, commodities and Treasuries. Our forecasts were
compared against 3 of our favourite44 trend-following benchmarks:
 Our naive trend signal, as introduced in Natividade et al [2013] but further
modified in Natividade and Anand [2016] to account for asymmetric noise
control.
 The Mann-Kendall signal, based on the correlation of sequential returns
from different lookback windows. This method, whose details are further
explained in Natividade [2012b], has been particularly useful for intraday
predictions – see, for instance, Natividade [2013c].
 The Oomen-Sheppard trend signal, based on the relationship between
beginning vs end values and peak vs trough values.45
The benchmarks used multiple lookback windows to learn from, and the
backtest encompassed 11 years. We addressed the likelihood that spot would
mean revert over a certain interval through a non-parametric impulse-response
function, which weighted the signal from each lookback window accordingly.
Figure 14 plots, according to each horizon and forecast method, our aggregate Figure 14: Accuracy (%) of forecasts
hit ratio – the percentage of instances when the direction of a signal is the of the direction of future returns by
same as the direction of future asset returns. Two observations stand out:
horizon
 All methods seemed to have outperformed a random walk when predicting P-distribution Naive Trend
the direction of future asset returns, regardless of the horizon, over the 53.5% Mann-Kendall Oomen-Sheph.
sample window. That said, the directional accuracy was not far distant
53.0%
from 50%, reiterating the challenges in time series forecasting.
52.5%
 Our P-distribution forecasts increasingly outperformed the benchmarks as
52.0%
our horizons lengthened. This is unsurprising; most of our signals are of
51.5%
low frequency. While our Sentiment and (some of) our Value signals decay
fast,46 the Trend, Carry and Macro Factor signals target longer horizons. 51.0%
This observation also concurs with our experience trying to explain asset 50.5%
returns at specific frequencies.47 50.0%

Hzn = 1W Hzn = 2W Hzn = 1M
These results suggest our forecasts should outperform the futures price and, to We used all 15 markets and 11 years of data. We used hit ratios
the extent that we can keep our delta positions un-hedged for longer, that they for accuracy – that is, the % of forecasts of the direction of future
returns over a given horizon that turned out to be correct. Source:
should outperform other standard methods. This will become more evident in Deutsche Bank
Section 5.7, when we look at the performance of different delta hedging

methods across asset classes.
4.5.2 Forecasting Accuracy: the Volatility

We now turn our attention to the accuracy of our empirical volatility forecasts.
44
Each looks at trendiness from a different perspective, and the first 2 have been successfully used in our
paper trading before. For a larger set of potential benchmarks, see Natividade [2012b].
45
See Oomen and Sheppard [2014]. We added this method for completeness given that it accounts for
trend magnitude in a way that the other benchmarks do not. That said, its estimates suffer from
discretization bias as we use very few observations to estimate the signal at any point in time.
46
It is worth mentioning that weighting our signals unevenly – giving more weight to faster signals – did
not improve the forecast accuracy. We would have expected better results on 1-week forecasts had our
signals utilised intraday data as well.
47
In Natividade et al [2014] we ran panel regressions that tried to explain short, medium and long-term
contemporary returns in each asset class using a series of market, macro-economic, technical and
fundamental variables. We found that the explanatory power rose as we lowered the frequency of one-
period returns (from weekly to monthly, then to quarterly, semi-annually and annually).

20 April 2017
As with the mean, we evaluate our forecasts under multiple horizons: 1 day, 1 Figure 15: Market turbulence periods
week, 2 weeks, 1 month and 3 months.48 All 15 markets were included. Period Event
Sep-/08 - Mar-/09 Lehman + GFC
We penalized forecasts according to the bias statistic introduced in Connor May-/10 - Sep-/10 EU Sov. Crisis
[2000] – also utilized in Alvarez et al. [2012] and Ward et al. [2016]. This metric May-/11 - Jun-/11 EU Sov. Crisis
measures the standard deviation of returns standardized by the volatility May-/13 - Jun-/13 Taper Tantrum
forecasted for the return period.49 If our forecast is accurate, this standard Sep-/14 - Mar-/15 Greek Debt Crisis
deviation should equal 1. If we have over-/under-forecasted vol, the computed May-/15 - Aug-/15 China Slowdown
measure will be less/greater than 1. Feb-/16 - Mar-/16 Oil Price Shock
Jun-/16 - Jul-/16 Brexit
We evaluated accuracy in 2 datasets: the long-term data and a subset that Source: Deutsche Bank
isolated crisis periods, as highlighted in Figure 15. In order to capture our

forecast accuracy going into these turbulent periods, we started our forecasts
one month prior to the dates shown.
Our tests focus on 4 aspects:

1. Parameterization: do parametric models outperform their simpler, non- Figure 16: Volatility forecasting
parametric peers? We tested the GARCH family versus EWMA, models
classical realized volatility, high frequency volatility and implied vols. Model Description Used in test
2. Factor modeling: does our factor model outperform its univariate RV Realized 1
volatility (daily
counterparts in predicting asset volatility? We tested both univariate data - lookback
and multivariate versions of our GARCH and EWMA models. = forecast
window)
3. Regime modeling: does calibrating our parametric models to
RV (HF) Realized 1
exogenous regimes help? We tested our parametric models with and volatility (30-
without regime calibration. min data -
lookback = 3
4. Event modeling: does separating event-related variance from non- days)
event variance help? We tested both cases. Univ EWMAs 7 EWMA 1
models (7
half-lives)
Figure 16 illustrates all the models used. Factor EWMAs 7 Multivariate 1
factor EWMA
We first tested the impact of parameterization. As per Figure 16, we compared models (7 half-
lives)
both CCC-MGARCH and univariate GARCH – which calibrate for persistence,
Univ GARCH GARCH(1,1) 1, 2
reaction and level – versus various EWMA estimates, in addition to past model
realized volatility using both close-to-close and higher frequency data50 and, Factor CCC- 3 Multivariate 1, 2, 3, 4
ultimately, implied volatility as well. None of these models incorporated MGARCH factor
GARCH(1,1)
regimes or events; these were addressed later. models *
IV Implied 1
Figures 17 and 18 show 5 conclusions: Volatility
Source: Deutsche Bank; * The 3 versions of this modeling
framework are: 1) no regimes, no events; 2) no regimes, with
 The GARCH models appear most accurate overall, outperforming their events; 3) with regimes, with events. The EWMA half-lives are: 5,
10, 11.2 (corresponds to the Risk Metrics ‘daily’ lambda of 0.94),
non-parametric peers. 21, 90 and 252 business days.
 Within both GARCH and EWMA families, the univariate and factor versions
perform similarly.
 The forecast accuracy of all methods worsens as the horizon lengthens.
48
While we used multiple forecast windows, we note that the most relevant forecasts are those up to 1
month. VRP strategies use short-dated options.
49
As the focus here is on the cross-sectional comparison between models, we used daily observed,
overlapping returns in order to improve the sample size at longer horizons, especially for our second data
sample. Furthermore, for robustness, we have also run our overall accuracy results on model rankings
using non-overlapping weekly returns and our conclusions remained unchanged.
50
We included “high frequency volatility” estimates as they are popular among market makers and have
been advocated in academia – most notably, Ait-Sahalia and Mancini [2008].

20 April 2017
 None of the methods accurately predict market shocks; they all under-
forecast volatility going into the periods outlined in Figure15. This is a key
finding, with implications for signal generation in Section 6.51
 That said, implied volatility outperforms the other methods, when
forecasting 3 months ahead, during those periods. This is likely due to the
options market over-forecasting volatility, in general, relative to the other
methods. 52 Note also that such conclusion is less clear over shorter
horizons; implied volatility does not outperform in 1-month horizons, and
we did not include it for shorter horizons.
Figure 17: Long-term forecast performance across models (all horizons and underlyings)
RV (HF) IV RV Univ EWMAs Univ GARCH

1.2
Forecast Accuracy (abs bias)
0.9
0.6
0.3
0
0 0.2 0.4 0.6 0.8
Forecast estimation error
0.3 Univ GARCH Factor CCC-MGARCH

Forecast Accuracy (abs bias)
0.25
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6
Forecast estimation error
Source: Deutsche Bank; Overall data set – Oct 2005 to Aug-2016 – used for computations. Y-axis: we show the ‘absolute bias’ (i.e. absolute value of the difference between the bias statistic and 1, which corresponds
to a perfectly accurate forecast). X: axis: we show the standard deviation of a 252- bd rolling window computation of the bias statistic.
51
The reader may rightly question what threshold distance from 1 would define statistically significant
under-forecasting. We did not delve into that topic, and therefore our statement is generic.
52
In our view, implied volatilities outperform going into turbulent periods not because the options market
is more accurate at predicting shocks, but rather because the time-homogeneous, over-forecasting bias
works in its favour in this case.

20 April 2017
Figure 18: Forecast accuracy during crisis periods according to forecast horizon (days)
Univ GARCH Factor CCC-GARCH Univ EWMAs Factor EWMAs RV (HF) RV IV

0.5
Forecast mean absolute bias
0.25
0
1 5 10 21 65 all horizons
Univ GARCH Factor CCC-GARCH Univ EWMAs Factor EWMAs RV (HF) RV IV
underforecasting
1.5
Bias Test Statistic (mean)
1.25
overforecasting
0.75
0.5
The X-axis shows the vol forecast horizons (in business days). Y-axis: top chart: the same absolute bias measure as defined in Figure 17 is now shown averaged across all assets, for each horizon bucket; bottom chart:
the value of our bias test statistic (averaged across assets for each horizon bucket) is shown. All computations use the turbulent market periods shown in Figure 15. Source: Deutsche Bank
We opt for the CCC-MGARCH framework – which is both parametric and

multivariate. First, we favour its parametric nature as it allows us to
approximate the P-distribution more accurately; as mentioned earlier, we need
an accurate P-distribution estimate to fully harvest the multiple premias
available in options markets. Not only that, but a well-calibrated P-distribution
also allows us to simulate asset returns in far more realistic fashion, whose
applications go well beyond the trading of options.
Second, we favour modeling risk through factors because while it looks more
complex, portfolio risk estimation is rather simpler. It is no less accurate than
univariate modeling and has the added benefits of computational efficiency
and transparent risk attribution.
We next turn our attention to the modeling of events and regimes, zooming
into CCC-MGARCH. Figures 19 and 20 illustrate the following findings:
 Distinguishing between event and non-event returns when predicting

future volatility only leads to a trivial improvement, whether in the long run
or in turbulent periods. A more granular assessment would also show
notable accuracy gains in FX and commodities, as in line with Chen,

20 April 2017
Natividade and Wang [2011], but these are offset by losses in Equities and
Treasuries.
 Regime modeling proved somewhat more fruitful: the CCC-MGARCH
model with regimes outperforms its uni-regime counterpart during
turbulent markets, for horizons of less than 3 months. A more granular
assessment would also show that this is especially true in equity markets.
Figure 19: Event modeling – impact on forecast accuracy according to forecast horizon (days)
0.12 0.3
L ong-term (2005-2016) D u ring crisis periods
0.25

0.1
0.08 0.2
0.06
0.15
0.04
0.1
0.02
0.05
0
1 5 10 21 65 all horizons 0
Factor CCC-GARCH (with events) Factor CCC-GARCH
Factor CCC-GARCH (with events) Factor CCC-GARCH
Source: Deutsche Bank; Neither model includes regimes.
Figure 20: Regime Modeling – impact on forecast accuracy according to forecast horizon (days)
0.12 0.30
L ong-term (2005-2016) Du ring crisis periods
0.1
0.25
0.08 0.20
0.06 0.15
0.04 0.10
0.02 0.05
0 0.00
1 5 10 21 65 all horizons 1 5 10 21 65 all horizons
Factor CCC-MGARCH (with regimes) Factor CCC-GARCH Factor CCC-MGARCH (with regimes) Factor CCC-GARCH
Source: Deutsche Bank; Note that both models include events.
In light of the results above, we stick to the originally proposed model: CCC-
MGARCH, with events and regimes, which will be evaluated for signal
generation in Section 6. We now move on to alternative delta hedging, the
next step in our quest to improve VRP strategy returns.

20 April 2017
5. Alternative Delta
Hedging
Alternative delta hedging is, arguably, as ancient as option markets. Market
makers will happily deviate from their hedging template if they have an edge in
doing so, and if their risk constraints allow. A tailored delta hedging scheme, in
essence, improves the positive aspects and reduces the negative aspects of
the “textbook” delta hedging approach. For the options seller, for instance,
delta hedging is usually a loss-making spot-trading strategy: "buy high and sell
low". Foresight of future spot conditions could reduce the loss by, for instance,
under-hedging53 if the asset is expected to reverse course. For the volatility
buyer, the converse is true. It is true that alternative schemes will increase
delta risk, but this side effect should be compensated by the value-add of the
new scheme.
The schemes we introduce here utilize our subjective assessment of what the
future will look like – specifically, on future spot dynamics. 54 They either focus
on the direction of future asset returns, or on how the asset is trending. If
correct, using this extra information should give us extra gains.
The trend-based schemes ultimately dictate how much volatility will be

captured through delta hedging. They assume that markets are not fractal;
over the same sample history, high frequency and low frequency returns can
be very different.
“True” volatility is a latent variable, and therefore what we observe depends on

the estimation frequency and method. For instance, if spot appears to move
sideways when evaluated in 1-week intervals, but goes through 2-day trends,
the annualized 2-day volatility will be higher than the equivalent 1-week
volatility. Therefore, if our strategy is to be short options, we should prefer
delta hedging weekly rather than daily as in doing so we will capture less
volatility and hence incur a lower delta hedging loss.
The trend-based schemes we introduce will, in essence, seek to capture more

or less volatility than what is otherwise observed by using close-to-close data
points. They do so either by grid-searching optimal delta estimation
frequencies or by using moving averages to get different delta quantities.
The direction-based schemes, on the other hand, use a distinctly different

approach. Rather than trying to capture more or less volatility, they use
53
That is, either hedging less often or hedging a level that still leaves her under-hedged.
54
The reader may question why we did not focus explicitly on volatility – including estimating the delta
according to the volatility forecasts introduced in Section 4 and delta hedging daily to it. We see the 2
topics – alternative delta hedging and hedging to realized volatility – as different in nature. The latter is
about achieving delta hedging “purity”, while the former is about adding “impurity” to achieve extra gains.
We refer the curious reader to Ahmad and Wilmott [2005] and in Derman and Miller [2016], who
thoroughly address the topic of delta hedging to “true” volatility. They show that delta hedging to the
“true” volatility may provide a terminal P&L that is already known at the start of the trade (the difference
between the premium priced using the volatility estimate and the premium currently quoted in the market).
The issues, as they partly highlight, are that (1) “true” volatility is often unknown, (2) we cannot delta
hedge continuously, and hence suffer from discretization error, and (3) the P&L variance can be quite
significant over the life of the trade.

20 April 2017
directional forecasts to decide whether to hedge or to run with the current

Figure 21: Grid-search preferred
delta exposure. They allow for more delta risk if – and only if – that risk is likely
delta hedging frequencies
to pay off. The delta exposures in this category are managed either through
V.Frequent Frequent Medium Freq
dynamic bands, break-evens and stop-losses, or by using the forecasts from Infrequent V.Infrequent EUR/USD
1.7
our P-distribution. 1.6
1.5
Sections 5.1 – 5.5 introduce each scheme, while Sections 5.6 – 5.7 show our 1.4
1.3
backtest results using both simulated and real data. For completeness, we
1.2
cover both long and short option strategies. 1.1
0.9
5.1 Grid Search 0.8
2002 2004 2006 2008 2010 2012 2014 2016
WTI
The grid search is the "brute force" of our alternative delta hedging schemes. 160
140
Every day, it takes our options basket and calculates what the risk-adjusted
120
returns would have been, up to the day before, from delta hedging that basket
100
at the frequency stipulated by each grid-point.
80
60
This method effectively “observes” the underlying market at different
40
frequencies and infers which one would have been best for delta hedging up
to time t  1 . It then decides how to delta hedge at time t .55 Hedging to
20
0
frequencies where spot is more volatile should produce better results when we 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
S&P 500
are long options, as we are capturing more volatility. The converse is true 2300
when we are short. Gridpoints represent different frequencies, and are divided 2100
1900
into 3 domains: delta, time and spot. We describe them as follows:
1700
 Delta frequencies: hedge if the delta of an option has moved by more than 1500
a fixed quantity since the last hedge. Delta anchors are popular among 1300
market makers, especially when dealing with intraday hedging, as they are 1100
akin to trading under a volume clock 56 ; this frequency domain avoids 900
700
trading during illiquid periods as both spot and delta stay constant, thereby
500
reducing market impact. The downside, however, is that option deltas can 2002 2004 2006 2008 2010 2012 2014 2016
be highly volatile going into expiry – an important issue in practical VRP 190
10Y UST
implementation, as it affects costs and signal-to-noise. The grid points we 180

used are 5-, 10-, 15-, 20- and 25-delta steps. 170
160
 Calendar frequencies: delta hedge now if the last hedge occurred after a
150
certain number of days. The results from this domain are easier to interpret, 140
as trendiness and volatility are normally observed as a function of time. 130
Calendar frequencies do not discriminate across liquidity pockets, but 120
these are much less observable in daily data, and daily is the highest 110
trading frequency we assume in this report. The grid points used were 1, 3, 100
2005 2007 2009 2011 2013 2015
7, 11 and 20 business days.
V.Frequent: every 1 day, 3 days, 5-delta or 0.1 standard deviations
(in1M spot moves). Frequent: every 7 days, 10-delta or 0.3
 Spot frequencies: delta hedge the option only if the underlying asset has standard deviations. Intermediate: 11 days, 15-delta or 0.5
standard deviations. Infrequent: 16 days, 20-delta or 0.7 standard
moved by more than a fixed quantity since the last hedge. This domain deviations. V.Infrequent: 20 days, 25-delta or 1 standard deviation.
captures the best of the previous two: it captures liquidity pockets and As an example, the grid search method recommended hedging
every 1 or 3 days, or every 0.1 s.d. changes, during time periods
does not suffer from unstable deltas near maturity. That said, it also does shaded in brown. Source: Deutsche Bank
not address the path dependency of an option’s gamma; this is less of a

concern for VRP baskets but more of an issue for individual options. We
tested 0.1, 0.3, 0.5, 0.7 and 1 standard deviations (of monthly spot moves)
as grid points.
55
We used a P&L lookback window of 250 time units – 1 year – in order for the results to be more
adaptive to changing market conditions. We used Sharpe ratios as the defining metric. As is standard with
grid search methods, we also applied nearest neighbour smoothing both historically and cross-sectionally
(within a domain) across grid points.
56
See Natividade [2013b].

20 April 2017
Figure 21 shows the time-varying optimal delta hedging frequency when short
a basket of options57 in four markets, assuming no risk tolerance constraints,
overlaid against the futures contract. Darker colours indicate that the grid
search favours higher hedging frequencies, while lighter colours show that the
method preferred lower frequencies (in delta, calendar and spot domains) at
that time.
The results are intuitive, and similar regardless of the asset class. Lower
frequency delta hedging has been more optimal because implied volatilities
normally overshoot the empirical; the VRP harvester would have been better
off just collecting the option premium and rarely trading the (money losing)
delta hedging leg. When volatility rises and the asset is trending, however,
higher frequency delta hedging performs better as gains from the delta hedge
leg offset some of the loss from the short options position.
5.2 Moving Averages Figure 22: Chosen moving average

windows per asset class
The moving average (MA) scheme is similar to the previous in that it also 16 S&P 500 10Y UST WTI EUR/USD
solves for an optimal parameter level, but different in that the parameter in 14
Optimal MA window(days)
question is the asset price that goes into calculating delta. We hedge the delta 12
daily, but estimate it using a historical moving average price instead of what is 10
currently observed in the market as reference. In essence, this method 8
assumes that it has a better understanding of future conditions than what's 6
"priced in”, and that delta hedging to it will lead to better returns. 4
The scheme makes certain assumptions about trendiness, volatility 58 and 0

2002 2004 2006 2008 2010 2012 2014 2016
optimal spot references based on the historical performance (up to 1 day prior)
of multiple moving average reversal trading strategies59. It also picks the best Source: Deutsche Bank
or worst MA as spot reference depending on whether we are short or long

gamma:
 If most MA strategy returns are positive, we assume the asset is mean-
reverting. If we are short gamma, the delta hedge computation uses the
MA level of the best MA strategy as spot input. On the other hand, if we
are long the option we delta hedge to the current asset price.60
 If most MA (reversal) strategy returns are negative, we assume the asset is
trending. If we are short the option, we calculate our delta hedge amount
57
We use the baskets introduced in Section 3: 1-month rolling expiry, 12 strikes set 0.5 standard
deviations apart from one another.
58
Changes in the delta, which maps into the P&L of the delta hedge strategy, will be less significant when
we use a moving average as input compared to the current spot price.
59
The MA lookback windows range from 1 day to 1 month; we chose not to use longer windows so as
control how much our delta estimate can deviate from the benchmark. The MA strategies are signal-
weighted; in other words, the more that spot rises above the moving average, the stronger our short
position. Signal weighting is important as delta hedging strategies are gamma-weighted, and gamma
intensity relates to the size of recent moves. We use an anchored window to calculate strategy
performance, and apply nearest neighbour smoothing as is standard in grid search optimisation. We chose
an anchored window so as to capture structural patterns.
60
If the asset is mean reverting, we expect the moving average to outperform current spot at predicting
the asset price in the near future. If we are short a call, and spot has been rising, we will buy less spot in
the delta hedge than the original because we will use a lower level of spot as input. This decision to under-
hedge will be correct if spot is indeed mean reverting and therefore subsequently drops, as we will lose
less from the previous trade. If we are long a call, we will delta hedge to the current asset price. If we
instead delta hedge to a moving average, we would end up selling less than the benchmark trade and not
capitalise as much when the asset price eventually drops, which is sub-optimal. We choose the best
performing MA as we assume it is the one that spot is more likely to revert to.

20 April 2017
 using the current asset price as input. If we are long the option, we delta
hedge to the current MA level of the worst performing MA strategy.61 Figure 23: Delta hedge procedure
under the break-out method
Figure 22 plots the moving average windows chosen to estimate the level of
spot (or futures) used in our delta hedge calculator, assuming that we are short
options. If the chosen MA window was zero, it meant the asset was trending in
the short-run. In that case, we assumed that current spot would outperform a
historical average when predicting future spot, and thus chose not to under-
hedge our deltas. The opposite applied if the asset was mean-reverting, and
the length of the optimal moving average dictated how under-hedged we
were.
A key highlight from Figure 22 is the stability of our results for the S&P 500, an
asset known for its short-term mean-reverting properties. In this case, delta
hedging to a 5-10 day moving average is generally preferred.
Currencies, on the other hand, tend to mean-revert under shorter windows,

which is also reflected in our choice for the 2-day lookback. As currency
markets often trend under shorter horizons than equities62, it is unsurprising
that the algorithm opts for smaller MA lookback windows – we should under-
hedge by less. (1) Establish the break-out thresholds as asset price +/- premium;
delta hedge daily. (2) Spot has broken to the topside: stop delta-
hedging, initiate rolling stop. (3) The stop-profit got triggered:
establish the new break-out thresholds and delta hedge daily
5.3 Break Outs again. (A) Establish the break-out thresholds as asset price +/-
premium; do not delta hedge. (B) The asset price has broken to
the downside: start delta hedging daily, initiate rolling stop. (C) The
rolling stop has been triggered: establish the new thresholds and
stop delta hedging.Source: Deutsche Bank
The break-out scheme is comprised of 2 variables: the options break-even and
a rolling stop-loss on spot. Well-suited for a market maker, it sits in between
the two smart delta hedging categories described above.
This scheme resembles trend following algorithms in that it uses a break-out

filter and a stop mechanism, be it stop-loss or stop-gain, just as classical trend
following systems do.63 The break-out bands, in this case, are equal to spot
plus and minus the price of an ATMF option whose expiry equals that of the
option we are delta-hedging. We interpret these bands as the option market's
best estimate of the current break-out ranges in spot.
It is also similar to the direction schemes we describe later because once the
asset breaks outside the aforementioned range, we assume that it will continue
moving in that direction, and therefore increase or decrease the delta hedging
frequency depending on our option position.
The scheme works as follows:

1. At inception of the option trade, or when the rolling stop has been
triggered, we calculate the spot range: spot +/- the premium of an ATMF
option that expires the same day as the option we are trading.
61
If the asset is trending, we expect the moving average to underperform the current spot level at
predicting the asset price in the near future. In reverse image to the Footnote above, if we are short a call
and delta hedge to the moving average we will end up buying less spot than the original and therefore not
capitalise enough as the asset continues to rise. In this case, therefore, we use the current asset price as
input to calculate the delta hedge. On the other hand, if we are long a call, we would prefer being under-
hedged as spot keeps rising, and delta hedging to the moving average gives us that. In other words, we
sell less spot than the benchmark delta hedge strategy suggests. We choose the worst-performing
moving average as it is the one that spot is the least likely to revert to.
62
These results are consistent with our work on impulse response functions shown in Natividade et al
[2014b].
63
See, for instance, Clenow [2013].

20 April 2017
2. While the spot level is inside the range, if we are long the option, we delta Figure 24: Signals from the Expected
hedge daily so that the gain from our long gamma exposure provides Returns scheme: long vs short the
relief against the loss from our short theta exposure. In other words, we
underlying asset
“gamma scalp”. If we are short the option, however, we do not delta
hedge and collect the returns associated with time decay instead. Long Short EUR/USD
1.55
3. If spot breaks to the topside of the range, we initiate a rolling stop trigger 1.45
set at 0.3 standard deviations below spot and roll it daily.64 The delta 1.35
hedge varies according to the option position: 1.25
o If we are long a call or a put: stop delta hedging65 and “let the delta 1.15
run”. 1.05
0.95
o If we are short a call or a put: cover the delta exposure immediately
0.85
and start delta hedging daily. 2002 2003 2005 2007 2009 2011 2013 2015
S&P 500
4. If spot breaks to the downside of the range instead, we also start a rolling 2250
stop level. The delta hedge varies: 2050
1850
o If we are long a call or a put: stop delta hedging and “let the delta
1650
run”.
1450
o If we are short a call or a put: cover the delta exposure immediately 1250
and start delta hedging daily. 1050
850
5. If spot retraces and hits the rolling stop, we cover the delta exposure
650
immediately and re-start the process above. 2002 2003 2005 2007 2009 2011 2013 2015
WTI
160
Figure 23 illustrates the process when applied to long call and short put
140
positions, under opposite market scenarios. 120
100
5.4 Expected Returns 80
60
The Expected Returns (ER) scheme explicitly incorporates our return direction 40
forecasts into the delta hedging decision, more so than any other scheme that 20
we are introducing. The idea is simple: if the delta hedge trade we need to do 0
2006 2007 2009 2010 2011 2012 2013 2015
today is in line with our spot view, we delta hedge; otherwise, we do not. In
10Y UST
other words, we will stay under-hedged or over-hedged if that suits our spot 180
views. 170
160
150
The better our forecasts can predict the direction of future spot returns, the
140
more that this delta hedging approach should outperform the others. That said,
130
it also causes our VRP strategy to deviate from the benchmark, which – in
120
extreme form – may lead to a significant breakdown in correlations with pure
110
VRP returns.66
100
2005 2007 2008 2009 2010 2011 2012 2014 2015
Our return forecasts come from the P-distribution introduced in Section 4. As Source: Deutsche Bank
described earlier, the forecasts combine all the individual signals from our
delta-one portfolios – Trend Following, Carry, Value, Sentiment and
Macroeconomic Factor investing. We are primarily interested in the direction of
64
The length used to calculate the standard deviation is equal to the number of days between the last
delta hedge date and today. The stop adjusts itself at every spot move but cannot become less stringent
than it was before. The 0.3 s.d. level was not optimised to this exercise; it was similar to a choice we used
before. See Natividade [2013c] for more details.
65
In cases where we are long the option and stop delta hedging, we do not remove the previous hedge.
We simply stay under-hedged from this point onwards.
66
In this case, we could end up capturing probability risk premium instead of the volatility risk premium.

20 April 2017
these returns, as opposed to the magnitude, as the former suits delta hedging
best.
Figure 24 plots flagship assets in FX, equities, commodities and Treasuries

against a background whose colours reflect the output from our spot models –
bullish (orange) or bearish (blue). One can see that the directional views from
our ER method do not flip often, a potential downside given that delta hedging
normally classifies as short-term trading.
5.5 Whalley & Wilmott Figure 25: Delta hedging bands –

Whalley & Wilmott method
Finally, we introduce a variation of the Whalley & Wilmott (WW) scheme. This 1.0 Delta bands
method has its origins in academia67, having been first introduced in Whalley & 0.9
Wilmott [1997]. It is a transaction cost model that defines a no-transaction 0.8
0.7
region for delta hedging. The costlier it is to trade the underlying, the wider the 0.6
band and therefore the lower the trading frequency. In our context, we modify 0.5
BSM delta
the model such that the delta bands reflect our views on the underlying asset 0.4
Wide bands
0.3
instead of trading costs. The bands are defined as follows: Tight bands
0.2
0.1
BSM delta
 3 exp  r T  t S
1

0.0
2 3 0.0 0.2 0.4 0.6 0.8 1.0
Bands   BSM   
2  
where  is a function of the transaction cost68, r is the risk-free interest rate,

t, T are time references, S is spot,  BSM is the Black-Scholes-Merton delta
and  is the BSM gamma.
The variable of most interest to us is  , a risk aversion parameter 69 that

defines the convexity of the bands. Instead of setting it as a function of our
cost-related utility preferences, as done originally, we define it according to the
spot views coming from our P-distribution. In essence, this method is a
constrained version of the ER method – we allow for some freedom in our delta
hedge decisions but ultimately also apply exposure thresholds. For instance, if
today’s delta hedging trade is in line with our view on future asset returns, we
apply very tight delta hedging bands. Otherwise, we set them wide.
67
To our knowledge, smart delta hedging is not a popular theme in academic circles. Delta hedging
discretization attracted a good deal of attention in previous decades, but the techniques introduced sought
to reduce trading costs as opposed to capture directional views or the non-fractality of asset returns.
The topic was initially addressed in Leland [1985], where the author revised the option’s implied volatility
to account for the costs of delta trading. Hodges and Neuberger (HN) [1989] followed through with an
optimal but computationally expensive band-like approach. Whalley and Wilmott [1997] refined the method
with an analytical formula, and Zakamouline [2006a, 2006b, 2007] further expanded on it using a double
asymptotic method that more closely resembled the optimal HN approach. We prototyped the 2 latter
methods; while the Zakamouline scheme seemed most appropriate for the cost problem, it added little
extra to WW while also costing more computational time. Sinclair [2008] provides a thorough review of the
topic.
The main references we found on smart delta hedging are Chen [2010] and Chen et al [2011], where the
author identified a jump-based delta hedging rule applicable to options on S&P 500 futures. Chen argues
that as the S&P 500 is likely to revert after large jumps, one should suspend rebalancing the hedge
portfolio after large jumps and only rebalance the next day. Finally, Sepp [2013] implemented a grid search
method similar to what we introduced in Section 5.1, but with the goal of reducing delta hedging costs as
opposed to alpha generation.
68
tc   | N | S
69
Refer to Whalley and Wilmott [1997], Sinclair [2008], Chen [2010] and Chen et al [2011]for details.

20 April 2017
Figure 25 illustrates the method in more detail, where the delta bands are
plotted against the original Black & Scholes delta.
5.6 Backtest Results – Simulated Data Figure 26: Simulated trending and
mean reverting price series – 5 paths
Having introduced our alternative delta hedging schemes, we now evaluate each
individual performances on both simulated and real market data. We are
6000
Simulated trending series
ultimately interested in the latter, but the former will be the backbone of our
5000
understanding of how different methods perform under various conditions.
4000
We focused on 2 types of simulated data – trending and mean-reverting. We 3000
simulated slightly over 4,000 datapoints in each 70 , using AR(1) return

2000
generation processes as in Kaminski and Lo [2007]:
1000
rt     rt 1      t 0
Dec-89 Dec-93 Dec-97 Dec-01 Dec-05 Dec-09 Dec-13
are daily asset returns,  t ~  (0,   ) is a noise parameter and

2
where rt 180
Simulated mean reverting series
  (1,1) reflects the mean reversion rate. Our trending price series are set 160
140
such that annualized asset returns and volatility both average 10%, and
  0.5 to indicate strong trendiness. Our mean-reverting series also have
120
100
10% annualized volatility, but average returns are 0% and   0.5 to 80
indicate strong reversal.71 Figure 26 shows 10 runs of our simulated trend and 60
40
mean-reverting series.
20
0
As for backtesting our options strategies: Dec-89 Dec-93 Dec-97 Dec-01 Dec-05 Dec-09 Dec-13
[1] We used the same basket of options introduced at the start of Section 3:
twelve 1-month options, whose strikes are set 0.5 standard deviations
apart from one another. We assumed a flat volatility surface such that all
options were priced at 10% implied volatility. Other operational details
were standard for simulation: zero interest rates, zero transaction costs,
individual delta hedging for each contract, and roll at expiry.
70
If each time unit represents a day, this would be equivalent to 16 years of data.
71
To make it easier for the reader:    A 1  2 , where  A  0.1 . Further,
   A 252 , where  A  0.1 in the trending series and  A  0 in the mean reverting
series.

20 April 2017
[2] We evaluated 4 strategies: [a] long the options basket in a trending Figure 27: Backtest scenarios –
market, [b] short the basket in a trending market, [c] long the basket in a simulated data
mean-reverting market, and [d] short the basket in a mean-reverting
market. Figure 27 provides a sketch of that. Trending Mean-
asset reverting
[3] We also imposed delta risk thresholds so as to respect the typical Long
constraints of a market making desk. As such, the strategy automatically
gamma
[a] [c]
delta hedges the moment that our exposure exceeds the threshold,
whether or not it is in line with the candidate method. The thresholds Short
used were 5-, 10-, 20- and 100-delta, where the latter represents the gamma
[b] [d]
unconstrained version. This resulted in 80 backtests: 4 strategies, 4
tolerance thresholds and 5 candidate hedging methods.
[4] All results were ultimately compared to the benchmark: the same options
strategy but delta hedged daily instead.
We assessed 2 P&L characteristics in each backtest. Our utility metric was the
marginal Sharpe ratio gain versus the benchmark; other metrics, such as
drawdown-related performance, did not suit this controlled environment. Our
chosen risk metric was the correlation between strategy and benchmark; we
wanted to penalize strategies that captured less of the VRP phenomena but not
penalize strategies that are simply more volatile. The benchmark strategy is to
delta hedge daily.
Figure 28 illustrates the backtest results on simulated data. Each sub-figure

aggregates the same data but via different buckets: returns by strategy and
market environment ([2]), by delta tolerance threshold ([3]), and by delta hedge
scheme. The Y- and X-axis display the utility and risk metrics defined above,
except for we display “risk” as 1 minus correlation for better visualization.
Delta hedging strategies closer to the top left should be preferred. Our key
findings are that:
 Alternative delta hedging outperformed the benchmark in 55% of the

instances and in 3 out of the 4 strategy environments described in [2]. The
benchmark strategy – delta hedging daily – ranks last among schemes in
risk-adjusted returns.
 In the large majority of instances, our backtest returns were heavily
correlated to the benchmark. 72 In other words, all methods and most
tolerance thresholds still allowed us to capture VRP returns just as we did
when delta hedging daily. A tolerance threshold of 20-delta mildly
outperformed others from a utility-versus-risk perspective. But as expected,
the tracking error rose as we loosened our delta tolerance thresholds.
 Our alternative delta hedging schemes performed best when the asset was
mean reverting and we were short options ([d]), followed by when the
asset was trending and we were long the basket ([a]). In both instances,
the methods under-hedged, thereby “letting the delta run”.
 The only bad environment for our methods was [c] – long the basket when
the asset is mean-reverting. This is no surprise; it is optimal in this
environment to delta hedge as often as possible, which means delta
hedging daily. The benchmark, which does only that, is therefore the best
“gamma scalping” method.
72
The correlation of daily log changes, using 4,049 units, was above +0.4 in most instances.

20 April 2017
 No scheme subsumed the others, as seen by the lack of clustering in

Figure 28. Grid Search (5.1) ranked the highest, which means that the
optimal hedging frequencies of the past also outperformed in the future.
Again, no surprise – we used a controlled backtest environment with
stationary returns, subdued noise and no regime shifts.
Figure 28: Backtested results by hedging scheme (left), delta tolerance (middle) and market environment (right)
2.5 2.5 Marginal gain/loss in Sharpe ratio vs benchmark 2.5
Marginal gain/loss in Sharpe ratio vs benchmark Marginal gain/loss in Sharpe ratio vs benchmark
2.0 2.0 2.0
1.5 1.5 1.5
1.0 1.0
1.0
0.5 0.5
0.5
0.0 0.0
5d 0.0
#1: G.Search -0.5
-0.5 10d -0.5 Trending, long
#2: M.A.
#3: B.Outs -1.0 20d Trending, short
-1.0 -1.0
#4: E.Rets -1.5 Uncon. M.R., short
-1.5 -1.5 MR, short
#5: W.W. 1 - Corr to benchmark
-2.0
-2.0 1 - Corr to benchmark -2.0
1 - Corr to benchmark
0.40 0.50 0.60 0.70 0.80
0.40 0.50 0.60 0.70 0.80 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80
5.7 Backtest Results – Real Data
Now that we have an idea of how the results should look like, it is time to
evaluate how they actually look. Using up to 15 years of data, we applied all
delta hedging schemes to each of the 15 markets – in other words, the cross-
asset pool described earlier. The backtest setup was similar to the previous in
that we used the same (1-month) options basket, the same rolling frequency,
the same evaluation criteria and the same benchmark – the returns from delta
hedging the basket daily.73 As before, the delta hedge applied separately to 2
combinations: long the options basket systematically, and short the options
basket systematically. As before, we did not add costs – this will be left to
Section 6.
Our results are presented in Figures 29 and 30. Figure 29 shows the central
finding from our (near) 600 backtest iterations: on average, the Expected
Returns (5.4) and Whalley & Wilmott (5.5) schemes outperformed the others and
the benchmark, whether we were long or short options. In addition:
 All schemes outperformed the benchmark when short gamma. We find
that intuitive as it resembles the results from scenario [d] in Figure 27:
short options in markets whose returns are mean-reverting – in this case,
over the short run. Here, the delta hedge strategy buys more in rising
markets and sells more in falling markets. The less we trade it, the more
under-hedged we are, and the less we lose relative to the benchmark. On
average, all schemes under-hedge the delta relative to the benchmark.
 Most schemes underperformed the benchmark when long gamma. This
result is similar to that of simulated scenario [c] – long options in (short-
73
As before, we only look at the delta hedge leg of the returns, thus ignoring the option mark-to-market
leg. The latter is the same for all strategies.

20 April 2017
term) mean-reverting markets. None of the schemes captures this reversal

pattern as aggressively as the benchmark, which increases the size of its
short (long) position every day that the asset rises (falls).
 The amount of outperformance and underperformance grew as the
tolerance threshold for delta exposure grew. This is in line with the
arguments above, and with our findings using simulated data.
 Contrary to the others, the Expected Returns (5.4) and Whalley & Wilmott
(5.5) schemes excelled when long and short options. Further, the
outperformance grew as the tolerance thresholds grew. In our view, this is
because the schemes are based on the directional forecasts of our P-
distributions, and therefore account for a much broader range of asset
returns than short-term price action alone. The results indicate that these
schemes can add value even when daily hedging already looks ideal.
Finally, the fact that these schemes further outperformed under looser
delta thresholds is in line with Figure 14, which shows that our return
forecasts improve over longer horizons.
 Unlike with simulated data, the grid search (5.1) scheme underperformed
in the new backtest. It shows that naïve calibration is often not enough
when dealing with real data; persistent patterns in asset returns can be
clouded by regime shifts and intermittent drivers.
Figure 29: Aggregate backtest results (short – left chart, long – right chart) according to delta hedging scheme,
expressed relative to the benchmark
Marginal gain/loss in Sharpe ratio vs benchmark Marginal gain/loss in Sharpe ratio vs benchmark
0.08
0.30 #1: G.Search #2: M.A.
0.06
#3: B.Outs #4: E.Rets
0.25 0.04
#5: W.W.
0.02
0.20
0.00
0.15 -0.02
-0.04
0.10 -0.06 #1: G.Search #2: M.A.
-0.08 #3: B.Outs #4: E.Rets

0.05
-0.10 #5: W.W.
0.00 -0.12
5-delta 10-delta 20-delta Uncon. 5-delta 10-delta 20-delta Uncon.
Maximum delta tolerance thresholds Maximum delta tolerance thresholds
Figure 30 shows additional results. As before, the Y-axis displays the

incremental Sharpe ratio from using the aforementioned schemes versus the
benchmark strategy. The X-axis shows how the returns deviate from the pure
VRP benchmark, as we subtract the correlation between scheme and
benchmark returns from one. The first column of charts shows each backtest
labeled according to delta thresholds, while the second divides our results
according to hedging scheme and the third partitions them according to
underlying markets. The rows represent individual asset classes.
The following observations stand out:
 The more we loosened our delta tolerance thresholds, the more that the
tracking error grew irrespective of the hedging scheme. This is the most
discernible pattern, as shown by the charts on the left and in line with the

20 April 2017
backtests on simulated data. However, we are still capturing the VRP – all
correlations to the benchmark74 remained above +0.5.
 Schemes from Sections 5.4 and 5.5 are located, on average, above others
and less often below the X-axis. The charts in the middle column show
that.
 The charts on the right show no obvious pattern. This is good; we do not
want our results to be biased by a small number of assets. It also shows
that we are capturing a feature of the VRP premia as a whole, and not
tailored to a specific market.
Figure 30: Aggregate backtest results according to delta hedging scheme, expressed relative to the benchmark. Y-axis:
marginal gain/loss in Sharpe ration vs benchmark. X-axis: 1 – correlation to the benchmark.
0.5
0.5 #1: G.Search #2: M.A. Equities 0.5
Equities
Equities S&P 500 Nikkei
5d 10d 20d Uncon. 0.4 #3: B.Outs #4: E.Rets
0.4 0.4 Eurostoxx
#5: W.W.
0.3
0.3 0.3
0.2
0.2 0.2
0.1
0.1 0.1
0.0
0.0 0.0
-0.1
-0.1 -0.1
-0.2
-0.2 -0.2
-0.3
-0.3 -0.3
0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45
0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45 0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45
FX FX FX
E/$ $/Y A/$ $/BRL
0.2 0.2 0.2
0.1 0.1 0.1
-0.1 -0.1 -0.1
-0.2 -0.2 -0.2
-0.3 -0.3 -0.3
-0.4 -0.4 -0.4

0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45 0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45 0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45
0.4 Commodities 0.4 Commodities 0.2 Cdy's

Gold WTI Corn Copper
0.3 0.3
0.1
0.2 0.2
0.1 0.1 0.0
0.0 0.0
-0.1
-0.1 -0.1
-0.2 -0.2
-0.2
-0.3 -0.3
-0.4 -0.4 -0.3

0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45 0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45 0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45
0.4 Bonds 0.4 0.4 Bonds

Bonds UST Bund JGB
0.3 0.3 0.3
0.2 0.2
0.2
0.1 0.1
0.1
0.0 0.0
0.0
-0.1 -0.1
-0.1
-0.2 -0.2
-0.3 -0.3 -0.2

0.00 0.02 0.04 0.05 0.25 0.45 0.00 0.01 0.02 0.03 0.04 0.05 0.00 0.01 0.02 0.03 0.04 0.05 0.05 0.25 0.45
0.05 0.25 0.45
Correlation of daily returns, 8-15 years of data depending on the market. Source: Deutsche Bank
74
Daily returns, 8 to 15 years of data depending on the underlying market.

20 April 2017
6. Building the Strategy

So far we have shown that there is value in forecasting asset volatility and in Figure 31: Signal-weighting schemes
delta hedging differently. We now evaluate whether these can be used to build Benchmark Wlsm
2.5
a better volatility risk premium signal and, ultimately, a better VRP strategy. 2
Wlsb
Wsb
Wsm
EURUSD (rhs)
1.5
1
Our tests covered the same 15 markets as before, and 1-month options. But as
0.5
this section deals with implementation, we modified the variance swap- 0
replicating options basket to be comprised of 10- and 25-delta puts and calls, -0.5
and ATMF puts and calls, all rolled at maturity. The weight of each option was -1
-1.5
inversely related to strike levels, as per standard variance swap replication. We
-2
also added transaction costs; we used our internal database to estimate -2.5
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
volatility bid-ask spreads for the key strikes, and applied linear interpolation to
Benchmark Wlsm
cover other points in the bid-ask surface. As was the case before, our 2.5
Wlsb Wsm
2 Wsb WTIUSDf (rhs)
benchmark is a strategy that sells the same basket of options every month, 1.5
delta hedged daily. 1
0.5
0
This section covers signal estimation and timing, in addition to implementation
-0.5
aspects of our alternative delta hedging (ADH) strategy and how it interacts -1
with timing. We finalise by presenting our results on selling option baskets on -1.5
-2
3 equity indices: the S&P 500, Eurostoxx 50 and the Nikkei 225.
-2.5
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
6.1 Fine tuning the VRP signal 2.5 Benchmark

Wlsb
Wlsm
Wsm
2 Wsb S&P 500 (rhs)
1.5
We first focus on the volatility forecasts from Section 4. As CCC-MGARCH 1
outperforms in forecasting future volatility, it may also produce better volatility 0.5
0
trading decisions than the benchmark – which always sells a constant amount.
-0.5
Here we assumed daily delta hedging but changed the size of our short (and -1
long) options positions to account for the forecasts. The weighting schemes -1.5
we tested are a function of the spread between current implied75 volatility and -2
-2.5
our forecasts of future realized volatility.76 They focused on direction (long or 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
short, short-only) and magnitude (binary weights, weight by signal intensity). Source: Deutsche Bank
The weights were built as follows:
 I   4C
1. We took the absolute spread s  and calculated the current
I
percentile relative to the past 1 year77 – our signal p.
2. We turned the signal into positions according to direction and magnitude:
 Long-short, binary-weighted: Wlsb  signs  .

 Long-short, magnitude-weighted: Wlsm  p  Wlsb  2 .78
75
We used variance swap strikes where available, typically sourced from the CBOE. Where unavailable,
we used a weighted average of 1-month implied volatilities across the smile so as to proxy the variance
swap strike.
76
It is worth noting that we also backtested variations where the original signal was based either on our
volatility forecast alone, or implied volatilities alone. While these schemes were more exposed to factor
momentum, the results were not significantly different in how they compared versus the benchmark.
77
We use a 1-year lookback window to make the signal more adaptive.

20 April 2017
 Short-only, binary-weighted:
if signs   0, then Wsb  0, otherwiseWsb  1.
 Short-only,magnitude-weighted: Wsm  p  2  Wsb

Figure 31 plots our weighting schemes over time in 3 markets – EUR/USD, S&P Figure 32: Performance rankings
500 and WTI – overlaid against asset returns. All 3 assets have typically fallen versus the benchmark
faster than they rose in the past, which results in most long volatility positions Sharpe ratio
coinciding with bearish spot markets. Bench LS M. S M LS B. SB
mark (Wlsi) (Wsi) (Wlsb) (Wsb)
That is not to say, however, that the long options positions paid off. This is best S&P 3 5 4 1 2
seen in Figure 32, which ranks the backtest results from different weighting SX5E 2 4 1 5 3
schemes according to their Sharpe ratio and shortfall-over-returns. The Nikkei 1 4 3 5 2
benchmark strategy (short-only, delta-hedged daily) ranks first. The Wsb (short- Bov 2 5 4 3 1
only, binary weighted) scheme, which uses the least amount of information E/$ 1 5 3 4 2
from our signal p , ranks second. $/Y 1 5 3 4 2
A/$ 4 3 5 1 2
In other words, neither the long volatility positions nor the long volatility $/BRL 1 4 2 5 3
predictions translated into overall outperformance. As alluded to in Section Gold 1 5 4 3 2
4.5.2, our forecast method – and all other benchmarks, including implied WTI 1 5 3 4 2
volatility – are not entirely accurate at predicting market shocks. Whether Corn 4 2 1 5 3
magnitude-weighted or binary-weighted, the long positions added more noise Copper 5 1 2 4 3
than value.
UST 1 5 2 4 3
Bund 3 2 1 5 4
The short-only weighting schemes outperformed the long-short, but still trailed
JGB 1 5 4 3 2
the benchmark. This is because if the volatility forecasts went above the
Rank 1 5 3 4 2
market implied, the exposure went to zero. These flat positions often translated of Avg
into missed opportunities to sell volatility and capture the VRP. Shortfall over annualised returns
S&P 3 5 4 2 1
Finally, the new weighting schemes led to a reduction in return correlations to SX5E 1 4 2 5 3
the benchmark strategy.79 By using our forecasts to reduce losses or turn them Nikkei 1 4 3 5 2
into gains, we ended up less exposed to the risk premium we aimed to capture
Bov 2 5 4 3 1
in the first place. VRP strategies are not alpha strategies; in our view, they
E/$ 1 5 3 4 2
should not be designed to produce positive returns in every market
$/Y 1 5 3 4 2
environment.
A/$ 4 3 5 1 2
$/BRL 2 4 1 5 3
These results highlight a challenge often faced by quantitative researchers:
Gold 1 5 4 3 2
better statistical results may not lead to better strategy returns. CCC-MGARCH
WTI 1 5 3 4 2
may outperform its peers at forecasting volatility, but it still suffers from the
Corn 5 2 1 4 3
same general shortcomings that make it underperform, on average, the naïve
Copper 5 1 2 4 3
benchmark strategy.
UST 1 2 5 3 4
We will use this forecast method elsewhere – it is a key input to our P- Bund 3 2 1 5 4
distribution, and as such will be used for deriving probability risk premium JGB 1 4 5 3 2
signals and any other use of our P-distribution technology. For volatility risk Rank 1 5 3 4 2
of Avg
premium, we opt instead for the original signal and therefore will hold a Sharpe ratios are ranked in descending order, while Shortfall-over-
constant short position on the options basket. returns are ranked in ascending order.LS M: long-short,
magnitude weighted. S M: short-only, magnitude-weighted. LS B:
long-short, binary weighted. S B: short-only, binary weighted.
Backtests since 2005, where available. Source: Deutsche Bank
78
We multiply the signal scores by 2 so that the absolute average historical weight is 1; this allows for
better comparison against the benchmark.
79
The 11-year correlations of daily returns versus the benchmark VRP strategy were 0.43 for the long-
short weighting schemes and 0.78 for the short-only weighting schemes.

20 April 2017
6.2 Adding alternative delta hedging Figure 33: VRP strategy performance
– benchmark versus alternative delta
Our alternative delta hedging methods are next. The results from Sections 5.7 hedging
are encouraging, but must be understood in more detail. Given its 1.5 Sharpe ratio (new D.Hedge)
outperformance, we opt for hedging scheme from Section 5.4 (Expected
Returns) but also cap our absolute delta exposures to a maximum of 20 delta 1.2
so as to reduce deviations from the benchmark. Whenever the threshold is hit,

0.9
we automatically hedge the delta back to zero.
0.6 FX Equities
The top chart in Figure 33 shows that a VRP strategy that delta hedges C'dys Bd Fut
according to the scheme above outperforms the benchmark in almost every 0.3
market. This had been alluded to in Section 5.7, and is attributed to the Sharpe ratio (benchmark)
predictive power of our P-distribution. 0.0
0.0 0.3 0.6 0.9 1.2 1.5
-10 -8 -6 -4 -2 0
0
But the more interesting finding lies in the bottom chart: on average the Average of top 5 drawdowns (benchmark)
-1
drawdowns under the new scheme are not worse, but they are not noticeably
FX Equities -2
better either. Just as with our volatility forecasts, our spot market forecasts do C'dys Bd Fut -3
not predict upcoming shocks. Alternative delta hedging acts as an extra -4
“income” generator; it improves returns but should not be viewed as a -5
drawdown control mechanism for VRP strategies.80 -6
-7
-8
We next evaluate whether these extra returns pollute our capturing of the
-9
volatility risk premia. We need to ensure that the marginal gains are not due to Average of top 5 drawdowns (new D. hedge)
-10
structural exposures to static or dynamic factors that a pure VRP strategy does Up to 15 years of returns depending on the market. Source:
not have. The long run correlations of daily returns between delta hedging Deutsche Bank
P&Ls81 are at 0.9482, which suggests we are still capturing the VRP. That said,
we need to understand what drives the difference.
In order to search for hidden exposures in each of our 15 asset-specific VRP

strategies, we performed an analysis of the attribution of returns. First, we
orthogonalised the new delta hedging P&L against the benchmark. We then
regressed the residuals against each of 3 explanatory variables: (1) the asset
returns, (2) returns from the positions on the asset in our delta-one portfolios
from Section 4.1, and (3) the returns from a (time series) momentum strategy
on the asset. In our view, these are most likely to incorporate any potential
time-homogenous exposures.
Figure 34 shows the time-varying aggregate explanatory power that these 3

variables have on the delta hedge P&L residuals over time, while Figure 35
shows the time-varying betas. The regressions use daily returns, and a 1-year
lookback window so as to be adaptive.
We highlight a few observations:
80
The ratio of annualised returns to top drawdowns improved in 13 out of the 15 markets under
alternative delta hedging versus the benchmark. However, that is mostly due to an improvement in returns
and not a reduction in the drawdowns.
81
We isolated the delta hedging P&L as opposed to the full strategy P&L in our calculations so as not to
bias our correlations upwards. The options mark-to-market is the same for both alternative and benchmark
strategies, it is only the delta hedging leg that changes.
82
We use 15 years of data in FX and most equity markets, and 10-12 years in commodities and
Treasuries.

20 April 2017
 The regressors explain, on average, 30% of the variations in the residuals Figure 34: Time-varying R-squared –
over time. The residuals are not just noise.
P&L residual regressed against the 3
 That said, there is no specific, time-homogenous factor exposure at the explanatory variables
aggregate level or specific to a given asset. All exposures are cyclical.
1 FX Equities
 Some asset class-specific patterns have also emerged. In Treasuries, the 0.9 C'dys Treasuries
0.8
alternative delta hedge strategy had a long bias, which reflects the multi-
0.7
decade rally in the asset class. Equities and FX historically loaded
0.6
positively to asset momentum, except during highly turbulent periods and 0.5
in the past year – both being instances when markets were notably mean- 0.4
reverting. 0.3
0.2
The cyclicality and adaptivity of these results further support the use of 0.1
alternative delta hedging as a replacement to daily delta hedging. As such, we 0

2002 2004 2006 2008 2010 2012 2014 2016
favour using it when harnessing the volatility risk premium. 1Y rolling window, daily data. Source: Deutsche Bank
Figure 35: Time-varying beta exposures – P&L residual regressed against asset returns, delta one portfolios and asset
momentum
5 FX Equities 5 Beta to asset FX Equities 5 Beta to asset direction FX Equities
Beta to delta one momentum
portfolios C'dys Treasuries C'dys Treasuries 4 C'dys Treasuries
4
4
3 3
3
2 2
1 2 1
0 0
1
-1 -1
0
-2 -2
-3 -1 -3
2002 2004 2006 2008 2010 2012 2014 2016 2002 2004 2006 2008 2010 2012 2014 2016 2002 2004 2006 2008 2010 2012 2014 2016
6.3 Strategy timing
Timing is the final aspect we will consider in this report. 83 Section 3.2
suggested that exogenous variables may be useful for conditioning exposure
during turbulent markets – periods when neither our volatility forecasts nor
alternative delta hedging helped. We now delve deeper into this idea.
6.3.1 Timing using the Global Sentiment Indicator

Timing a strategy implies increasing or decreasing the capital allocated to it
according to an algorithm. It can be based on endogenous characteristics of
the strategy – such as factor momentum or reversal – or defined by variables
that are exogenous. We considered timing through 3 indicators: our volatility
forecasts, implied volatility and the Global Sentiment Indicator. The first 2 were
transformed into 1-year percentile ranks for better comparison versus the third.
In each of the cases, we defined the capital allocation ratio as
Lt ,t  h  2  I t  0.02 , where I t is the level of one of the 3 indicators above
and L  [0,2] . The higher the indicator, the more risk averse markets were
likely to be, and therefore the lower the capital allocation.
83
Cross-market diversification and modifying the VRP options basket will be addressed in future reports.

20 April 2017
Figures 36 and 37 illustrate the results of timing the original benchmark VRP Figure 36: Sharpe ratio rankings –
strategy in each market according to the indicators above. While they VRP strategies timed according to 3
correlated heavily to one another, the GSI ranked top in most instances.84
indicators
Using the GSI timer, the drawdowns were cut by approximately 20% from the Implied Volatility GSI
original and the ratio of annualised returns to average drawdowns almost volatility forecast
doubled.85 S&P 500 3 2 1
SX5E0 3 2 1
But while these results are encouraging, they are not granular enough. In order Nikkei 3 2 1
to assess how consistent is the improvement across drawdown types, we Bovespa 3 2 1
looked at the ratio of returns between the GSI-timed VRP strategy and the E/$ 3 2 1
original benchmark in all drawdowns in the latter across all 15 markets. Such $/Y 2 1 3
analysis is important so as to reduce the sample bias highlighted in Section A/$ 2 3 1
3.4. $/BRL 3 2 1
Gold 3 2 1
Figure 38 plots the distribution of ratios across 2,185 drawdowns in the
WTI 2 3 1
aggregated benchmark VRP strategies, bucketed according to the magnitude
Corn 1 3 2
of loss. Numbers below 1 imply the GSI timer helped, whereas those above 1
Copper 3 2 1
imply the opposite. The findings are sobering: on average, timing only tames
UST 3 2 1
medium to large drawdowns – those above the 30th percentile of the aggregate
Bund 1 3 2
drawdown distribution. In other words, our proxy for global sentiment is
JGB 3 2 1
unlikely to capture smaller, idiosyncratic-driven shocks in each particular
Rank of 3 2 1
market, but it should help against stronger, macro-driven drawdowns in the Average
VRP strategy. That said, the wide variations inside each bucket also suggest Source: Deutsche Bank
that the GSI provides no guarantee of improvement.

Figure 37: Drawdown profile with
Figure 38: Distribution of drawdown ratios (timed vs un-timed strategy)
and without timing – VRP returns
according to drawdown size
-8 -7 -6 -5 -4 -3 -2 -1 0
0
GSI-timed returns divided by benchmark returns Average of top 5 drawdowns (benchmark)
25-75% 33-67% 50% -1
during benchmark drawdowns
Equities FX -2
1.8
C'dys Bd Fut -3
1.6 -4
-5
1.4
-6
1.2 Average of top 5 drawdowns

-7
(benchmark with GSI) -8

1.0
10-15 years of data depending on the market. Costs are included.
0.8
0.6
0.4
0.2
0.0
<10th 10-30th 30-60th 60-90th >90th
Drawdown size according to percentile buckets
84
Our volatility forecasts also outperformed the implied volatility in most instances, further confirming the
results from Section 4.5.2.
85
Average of the top 5 drawdowns in each VRP strategy.

20 April 2017
6.3.2. Timing using long-term volatility reversal

Another, equally pertinent question has to do with the direction of our timing
variable. We assume above that rising volatility, and rising risk aversion, should
be negative for future VRP returns. However, both implied and empirical
volatilities are notably mean-reverting over the long run,86 which suggests that,
at extremes, our actions should be opposite to what is implied by our timing
algorithm. In other words, it implies we should lever up the VRP strategy at
times of extreme risk aversion and de-lever when volatilities are at an extreme
low.
We partly addressed this hypothesis by looking at future VRP returns according

to current implied volatility levels. In order to compare the results across 15
markets, and hence across 15 VRP strategies, we divided future returns by
volatility and measured current 1-month implied volatility according to its
percentile rank from a 5-year history.87 We then aggregated all cross-market
standardized returns and volatilities into one set.
Figure 39 plots the distribution of future returns according to horizon (1 week,

1 month and 3 months) bucketed according to percentiles of current implied
volatility. The results only partly agree with the long run reversal hypothesis
described above. Average returns are V-shaped; they do not rise monotonically
as implied volatility moves into higher buckets. Further, the pattern becomes
more evident as we lengthen the horizon of future VRP returns.
That subsequent volatility-adjusted VRP returns are high at the top volatility
quartile is easier to understand, as the strategy is selling volatility when it is at
historical highs. But the high (future) returns when (current) volatilities are at
the bottom quartile are less intuitive. In our view, this is likely due to anecdotal
evidence that volatility stays at low levels for longer than it does at high levels.
Figure 39: Future VRP return distribution according to buckets of current implied volatility
Future 1W VRP returns over 1W 25-75% 33-67% Future 1M VRP returns over 1M 25-75% 33-67% Future 3M VRP returns over 3M 25-75% 33-67%
return volatility 50% return volatility 50% return volatility 50%
1.0 1.2 1.4
0.8 1.0 1.2
0.6 0.8 1.0
0.4 0.6 0.8
0.2 0.4 0.6
0.0 0.2 0.4
-0.2 0.0 0.2
-0.4 -0.2 0.0
-0.6 -0.4 -0.2

0.25 0.33 0.50 0.67 0.75 1.00 0.25 0.33 0.50 0.67 0.75 1.00 0.25 0.33 0.50 0.67 0.75 1.00
Current 1-month implied volatility Current 1-month implied volatility - Current 1-month implied volatility -
- percent rank (5Y history) percent rank (5Y history) percent rank (5Y history)
86
See for, instance, Francq and Zakoian [2010] for a recent reference.
87
In other words, we take today's implied volatility and calculate where it resides relative to the past 5
years. We do not use future data to estimate the current percentile rank.

20 April 2017
6.3.3 Timing using volatility changes

Another important implication from our analysis of long run reversal dynamics
has to do with whether this should be incorporated into how we devise the
VRP strategy. We have, without testing, opted against that. Increasing capital
exposure when volatilities are historically low goes against intuition, and doing
the same when they are historically high seems imprudent from a risk
perspective. We rather try to reduce strategy drawdowns than try to speed up
its recovery, as the latter leaves us exposed to potential structural breaks.
Further, this V-shaped pattern is not as easily visible when using shorter-dated
lookback windows to estimate the volatility percentiles, or when using buckets
of our GSI values.
We also looked for patterns in how VRP returns relate to changes in implied
volatility and the GSI. As before, we standardized and aggregated the changes
in future VRP returns across markets, and compared these to current
standardized and aggregated changes in the level of implied volatility and the
GSI.
Figure 40: Future VRP return distribution according to buckets of recent change in implied volatility and GSI
return vol 50% return vol 50% return vol 50%
0.6 1.0 1.2
0.8 1.0
0.4
0.6 0.8
0.2
0.4 0.6
0.0
0.2 0.4
-0.2
0.0 0.2
-0.4 -0.2 0.0
-0.6 -0.4 -0.2

0.25 0.33 0.50 0.67 0.75 1.00 0.25 0.33 0.50 0.67 0.75 1.00 0.25 0.33 0.50 0.67 0.75 1.00
Changes in 1M implied volatility (percent rank) Changes in 1M implied volatility (percent rank) Changes in 1M implied volatility (percent rank)
return vol 50% return vol 50% return vol 50%
0.6 0.8 1.2
0.6 1.0
0.4
0.8
0.2 0.4
0.6
0.0 0.2 0.4
0.0 0.2
-0.2
0.0
-0.4 -0.2
-0.2
-0.6 -0.4 -0.4

25.00 33.00 50.00 67.00 75.00 100.00 25.00 33.00 50.00 67.00 75.00 100.00 25.00 33.00 50.00 67.00 75.00 100.00
Changes in GSI (percent rank) Changes in GSI (percent rank) Changes in GSI (percent rank)
Figure 40 illustrates our results, as it plots the distribution of standardized

future returns bucketed by 1-month changes in volatility and 1-month changes
in the GSI. All charts suggest that future strategy returns tend to fall in the
upper quartile – when either the current implied volatility or the GSI rose the
most versus the recent past. The results are intuitive; significant jumps in
volatility and the GSI imply risk aversion, which is detrimental to future VRP

20 April 2017
returns.88 These findings indicate the need for further research, as they alone Figure 41: Performance comparison
are not enough to justify changing the format of our timing algorithm. – ADH alone versus ADH + timing
-8 -7 -6 -5 -4 -3 -2 -1 0
6.4 Timing or Alternative Delta Hedging? Average of top 5 drawdowns (ADH)

0
-1
Equities FX -2
So far we have shown two VRP-enhancing approaches: alternative delta
C'dys Bd Fut -3
hedging (ADH), introduced in Sections 5.4 and 6.2, and timing, introduced in
-4
Section 6.3. The former acts as an "income" provider, whereas the latter
-5
reduces drawdowns. A final question is whether we should combine both.
-6
-7
Average of top 5 drawdowns
Figure 41 compares the returns of the ADH strategy on its own versus the ADH (ADH + GSI) -8
strategy combined with timing. The drawdowns improved in most cases - as 1.5 ADH + GSI Sharpe
expected - but the resulting Sharpe ratios and modified Calmar ratios 89 are 1.3
better in only half of the markets. Timing mechanisms often curtail income, as 1.1
they can lead to under-leveraging during recovery periods, and timing the VRP 0.9
seems no different. 0.7
Equities FX
0.5
Figure 42 delves deeper into this topic. We zoom into the exact difference 0.3
C'dys Bd Fut
between the ADH strategy and the benchmark - the P&L difference between 0.1 ADH Sharpe
both delta hedging legs - and evaluate whether there is value in timing that
0
-0.1 0 0 1 1 1 1 1 2
spread.90 As the Sharpe ratios show, there is no strong evidence that timing the 2.5 ADH + GSI modified Calmar
spread adds value, just as we find no evidence that our aggregate delta one
2.0
portfolios from Section 4.1 can be timed. The results look particularly worse in
Equities, as the orange boxes in Figure 42 show. 1.5
Equities FX
In summary, these results show there exists a trade-off between income 1.0 C'dys Bd Fut
boosting and drawdown control when trying new VRP enhancement methods. 0.5
While we favour alternative delta hedging in all instances, we believe the ADH modified Calmar
decision to further apply strategy timing should be dependent on the investor’s 0.0
0 1 1 2 2 3
risk constraints.
6.5 An Equity VRP Strategy
We finalise this section by showing our results in 3 equity indices: S&P 500,
Eurostoxx 50 and Nikkei 225. In all 3 cases, we systematically sell USD 100
worth of a basket of 1-month options, rolled every month. The basket is
comprised of 10-delta, 25-delta and ATMF calls and puts, where the capital
allocated is inversely related to the strike level - as per variance swap formula.
As elsewhere in Section 6, costs are included.
88
The reader will likely see another, smaller pattern in the form of an inverted u-shape in the chart
pertaining to future 3-month returns against 1-month changes in implied volatility. It suggests that
significant drops in volatility are also detrimental to future VRP returns. It suggests, in fact, that the most
favourable environment is one where implied volatilities are not moving (the 50-67th percentile bucket).
89
Annualised strategy returns divided by the average of its 5 worst drawdowns.
90
We effectively divided the strategy P&L into 3 legs: the options P&L, the benchmark delta hedge P&L
and the additional delta hedge P&L from the ADH approach. We focus only on the third leg because we
have already shown earlier that the GSI performs well in timing the other 2.

20 April 2017
Figure 43 and Figure 44 show our backtested results. The benchmark strategy Figure 42: ADH delta hedge spread
applies daily delta hedging, and is plotted in gray. The strategy that applies the with and without timing
GSI indicator to adjust position sizes in the benchmark strategy, as described in
Equities Sharpe ADH Delta Hedge spread + GSI
Section 6.3, is plotted in blue. The strategy that delta hedges according to our 0.9
FX
expected spot returns, as described in Section 5.4, is shown in blue. Finally, the 0.7
C'dys
strategy that uses both alternative delta hedging and timing is plotted in Bd Fut 0.5
orange.
0.3
0.1
Alternative delta hedging boosts the returns in all 3 cases; it bumps the slope
-0.5 -0.3 -0.1
-0.1 0.1 0.3 0.5 0.7 0.9
of our 3 time series. The GSI timer reduces the average of the 5 worst
drawdowns in 2 out of 3 strategies. When combined, the results are also mixed -0.3 Sharpe ADH Delta Hedge
spread
– on average, we see lower risk-adjusted returns but lower drawdowns. -0.5
Figure 43: Backtested cumulative returns – VRP strategy applied to equity index options
30 ADH + GSI Timer Alt. Delta Hedging 30 ADH + GSI Timer 25 ADH + GSI Timer Alt. Delta Hedging
S&P 500_Benchmark Bmk with GSI Alt. Delta Hedging Nikkei_Benchmark Bmk with GSI
25 SXE50_Benchmark 20
25
Bmk with GSI
20
20 15
15
10
15
10
5
10
5
0
0 5
-5
-5 0
-10
-10 -5 Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan
Jan 02 Jan 04 Jan 06 Jan 08 Jan 10 Jan 12 Jan 14 Jan 16 Jan 02 Jan 04 Jan 06 Jan 08 Jan 10 Jan 12 Jan 14 Jan 16 07 08 09 10 11 12 13 14 15 16
Figure 44: Backtest performance characteristics

Benchmark Alternative Delta Hedging GSI timer ADH + GSI timer
S&P 500 SX5E Nikkei S&P 500 SX5E Nikkei S&P 500 SX5E Nikkei S&P 500 SX5E Nikkei
Ann. Ret 0.8 1.1 1.0 1.7 1.8 1.9 1.05 1.38 1.11 1.6 1.6 1.3
Ann. Vol 2.7 3.4 3.7 3.0 3.7 4.2 2.51 3.31 3.34 2.9 3.7 3.8
Sharpe Ratio 0.30 0.33 0.28 0.59 0.49 0.45 0.42 0.42 0.33 0.57 0.44 0.33
Sortino Ratio 0.34 0.42 0.32 0.68 0.64 0.53 0.50 0.54 0.39 0.67 0.55 0.37
Avg Top 5 DDs -4.0 -4.2 -4.1 -3.4 -3.9 -4.5 -3.20 -4.65 -3.34 -3.0 -4.1 -3.7
Length Top 5 DDs
1.6 1.7 1.2 0.9 1.1 0.8 1.7 1.9 1.1 1.3 1.5 1.1
(yrs)
Mod. Calmar 0.2 0.3 0.2 0.5 0.5 0.4 0.3 0.3 0.3 0.5 0.4 0.3

20 April 2017
7. Conclusion
This report introduces a framework for extracting value in volatility surfaces
across asset classes. It is based on comparing the market-implied distribution
of future returns - the Q-distribution - with our subjective expectations of the
same - in other words, our P-distribution. In building the latter, we use a
parametric approach to combine the information from our delta-one systematic
strategies with a regime-switching, event-calibrated multivariate risk factor
model.
Today's Derivatives Spotlight focuses on the volatility parameter of the

distributions above. More specifically, on the volatility risk premium, a primary
source of value in options trading. We describe the merits and shortcomings of
popular VRP enhancement methods, and introduce a different approach based
on volatility forecasting and alternative delta hedging.
We compare our methods to multiple benchmarks focusing on both statistical

accuracy and strategy returns. As is often the case with rigorous testing, the
results were not always as encouraging as we had hoped. That said, we
believe that the final recommendations - based on alternative delta hedging
and timing - should improve the way investors harness this rich source of risk-
adjusted returns.
Our conclusions are as follows:
 Having some predictive power of future asset returns – such as when

using our real world P-distribution – can be valuable in improving VRP
strategy returns.
 In our context, the first moment of this distribution has been key in
improving the way we delta hedge. That improvement came from higher
risk-adjusted returns, although not from lower drawdowns.
 For the latter, timing has shown some encouraging results. We proposed a
method based on our Global Sentiment Indicator – which adapts fast to
changing market conditions.
 Future research will focus on other uses of the P-distribution and how we
can use it to identify and extract more sources of value in the volatility
surface.

20 April 2017
Appendix I: Our Multi-

Factor Risk Model
This section introduces our multi-factor risk model for cross-asset returns.
While its output can be used for multiple purposes in the future, here we use it
as input to the asset volatility forecasts introduced in Section 4. Our model is
based on the Arbitrage Pricing Theory (APT) – an extension of the Capital Asset
Pricing Model (CAPM) which evaluates the sensitivity of asset returns to a set
of risk factors. 91
The risk factors we use encompass macro drivers (inflation and growth),
market drivers (sector and asset class returns) and dynamic drivers
(Momentum and Carry). We calibrate the model to 80 assets across currencies,
equity indices, commodities and international Treasuries.92 Asset sensitivity to
our global drivers is estimated through stepwise robust regressions.
Section I.I provides a background to risk factor models. Section I.II provides the
algorithmic framework for setting up the factors and volatility forecasts. Finally,
Section I.III focuses on model estimation – in other words, how our betas are
derived.
I.I Background and Choice of Framework
Multi-factor models of asset returns, a.k.a. multi-factor risk models, are

commonplace in quantitative investing. They serve 3 primary purposes:
dimensionality reduction, better attribution of returns and better estimation of
how factors and assets interact. The first is key to large institutional portfolios;
instead of evaluating the covariances between thousands of securities, one can
model less than 100 factors instead – a much more palatable exercise. Risk
attribution is equally important; factor models decompose portfolio risk into a
smaller, manageable list of sources, making it easier to manage and protect
positions. Finally, risk models account for the interaction between factors and
assets, thus addressing spill-overs and other effects that are missed out by
univariate models. Combined together, these 3 features provide risk models
with an edge when predicting future asset volatility – our ultimate goal.93
Risk models are normally divided into 3 types: macroeconomic, fundamental

and statistical:94
 Macro factor models use observable economic anchors – tradable and

non-tradable – as risk drivers. The data pre-specifies the return of each
91
See Roll and Ross [1976] and Burmeister et al [1994].
92
21 equity index futures: ASX, Bolsa, Bovespa, CAC, DAX, Eurostoxx, HSI, IBX, ISE, JSE, Kospi, Nasdaq,
Nikkei, OMX, RDX, SMI, S&P 500, TSE, TWE, FTSE and WIG. 30 USD/FX: USD/G10 + USD vs BRL, CLP,
COP, CZK, HUF, IDR, ILS, INR, KRW, MXN, MYR, PEN, PHP, PLN, RON, RUB, SGD, THB, TRY, TWD, ZAR.
20 commodity futures: brent, cocoa, coffee, corn, cotton, gasoil, heating oil, natural gas, soybeans, sugar,
wheat, WTI, silver, aluminium, gold, copper, led, nickel, platinum, zinc. 9 bond futures: 10Y government
bonds in Australia, Canada, Switzerland, Germany, UK, Japan, Mexico, New Zealand, US.
93
See Ward [2010], Ward et al [2016] and Grinold & Kahn [1999] for an in-depth overview of risk factor
models and risk modelling.
94
See Zivot [2011] and Connor [1996] for more details on these 3 types.

20 April 2017
factor, and its loading to an asset is estimated through time series Figure 45: Correlation between
regression techniques. Macroeconomic factor models are the simplest and observed and estimated factor
most intuitive of the 3, though suffer from 2 potential drawbacks: mis-
returns
specification (when we use the wrong variable to represent the factor) and
factor omission (when we ignore a key driver).
 Fundamental factor models are distinct in that we assume that the factor
loadings are pre-specified, and factor returns are estimated through cross-
sectional regressions. They are the standard followed by risk factor
providers, and use asset-specific characteristics such as industry sector,
corporate accounting metrics and other style classification measures.
While popular in the equity investment community, this type of factor
model is less applicable to markets with few observable asset-specific
characteristics. It also suffers from the same drawbacks as macro factor
models, though have the advantage of being memory-less – it allows for
point-in-time risk estimation.
 Statistical factors take a completely different approach; neither the factor
returns nor their loadings are observable. Both are estimated instead
through statistical techniques – most commonly, factor or principal
component analysis. Mis-specification and factor omission are not a
concern in this approach, but statistical factors suffer from a lack of direct
economic interpretation.
Choosing from the 3 alternatives is not straight-forward; the researcher must
take a view on whether she needs the factors to be interpretable, whether
there are enough asset-specific characteristics that serve as drivers, whether
there are enough assets for cross-sectional regressions to be run efficiently,
and – ultimately – what she needs the model for.
In our case, we use it to forecast the volatility of asset returns. With that in
mind, the macro factor model approach suits us best. Our pool of assets is too
small for cross-sectional regressions to be reliable, and our assets lack enough
fundamental characteristics. At the same time, we want to be able to interpret
what each risk factor represents for future purposes. 95
Figure 45 illustrates part of our argument – that the fundamental factor model
approach does not fit our goals as well. It plots the correlation between factor
returns observed by market data – the “true” returns – and factor returns
estimated by the cross-sectional regressions defined using the fundamental
factor approach.96 The correlations are generally strong but not always; in the
case of rates markets, our estimated returns correlate little to the true factor
returns due to a lack of breadth of constituents.
I.II Selecting our Risk Factors
As the reader may suspect, selecting the qualifying factors was the biggest
challenge we faced when creating our risk model. We first had to define how
much of the variance in our basket of assets needed to be explained by
95
Examples of future use include portfolio risk hedging and risk factor investing.
96
We classified our pool of 80 assets into 6 sub-asset class buckets: DM and EM equities, DM and EM
FX, commodities and international Treasury markets. We assigned a value of 1 or 0 to an asset depending
on whether it belonged to that bucket, in the same manner as stocks are classified into industry or sector
buckets. We also added a global market bucket which applied to all assets. We then applied constrained
OLS regressions to estimate factor returns.

20 April 2017
common factors, and then we had to choose which factors best fit the task.
Both steps are challenging. The target explanatory power is a random variable
in itself, as the interaction between factor and idiosyncratic risk changes over
time. Choosing the actual factors is not straight forward either. We must
acknowledge our hidden bias – we know which factors explained the past –
and seek to minimize its adverse effect on future forecasts.
I.II.I Defining the explanatory power target

The first task is to define how much of the variance in our asset pool should be
covered by the multi-factor model. In other words, we must define a threshold
between factor risk and idiosyncratic risk. To do so, we resorted to Principal
Component Analysis (PCA). This method can differentiate between factor and
idiosyncratic variance as it outlines the marginal explanatory power of each
factor and how it loads into each asset.
There are 2 aspects to this task: defining the maximum number of statistical
factors, which is defined once using a long historical window, and choosing
the optimal number of factors applicable to a given rebalancing date. We
applied the Cattell’s Scree test to define the former, choosing a maximum
number of 8 (out of 80) statistical factors. 97 As for the latter, we applied the
information criteria proposed by Bai and Ng [2002].
Next, we want to evaluate how much of the variance in each asset can be
explained by the statistical factors chosen above. We run standard time series
regressions of asset returns on principal component returns and aggregate the
R-squared output from the 80 individual regressions.98 This number, which
changes at every rebalancing date, becomes our target explanatory power. It is
the amount of variance we will try to explain using the macro factors built
below.
I.II.II Choosing our factors

The premise of a macroeconomic risk factor model is, unsurprisingly, that Figure 46: Explanatory power using
asset returns are chiefly driven by macroeconomic developments. As such, it is nowcasting-based macro factors and
a sensible model choice for our markets given that they reflect the typical
interest rates versus the target
investment scope of a macro investor.
0.7 3-factor model Benchmark
0.6
This choice, however, creates 2 challenges: how to capture the desired factors,
0.5
and whether we are capturing enough of them.
0.4
0.3
We started with Chen et al [1986]; as per authors, we assumed that our asset
0.2
returns should be explained by surprises in GDP, inflation and interest rates.
0.1
We tried capturing the first 2 through the nowcasting indices introduced in
0
Natividade et al [2015], and the latter through 10Y US Treasury returns. Dec-01 Dec-04 Dec-07 Dec-10 Dec-13 Dec-16

Figure 46 compares the average explanatory power achieved through these
variables versus the target defined in Section I.II.I, making it clear that we have
not captured enough. It points us back to the 2 challenges highlighted above;
97
The Cattell’s Scree test plots the components as the X-axis and the corresponding eigenvalues as the Y-
axis. The contribution of each eigenvalue in explaining the total variance is presented in descending order
and linked with a line. Once the drop ceases and the additional eigenvalues explain little extra, we define
the cut-off. It suffers from the drawback that there is no deterministic solution as to the optimal number of
factors; the choice is subjective to the researcher. See Cattell [1978].
98
The OLS regressions were run on 2 years of daily data, rolled daily.

20 April 2017
nowcasting indices may be of higher frequency but perhaps still too slow, and
we may be missing other relevant drivers.
Next we tackled the factor representation issue. Market returns are often cited Figure 47: Explanatory power using
as representative of macroeconomic developments99, and as such we replaced market-based macro factors versus
our growth indices with asset class returns100 in G10 equities, G10 USD/FX,
the target
commodities and global Treasuries. These should capture both the macro
0.7
picture and asset class specific innovations, and correspond to the “market” Modf. 3-factor model Benchmark
0.6
variable alluded to in the CAPM. We also replaced our inflation series with a
0.5
market-derived metric: long nominal 10Y USTs while short 10Y TIPS, as per
0.4
Podkaminer [2013]. Figure 47 shows the effect of the new set on explanatory
0.3
power; we are capturing more, but still not enough.
0.2
0.1
Our next step was to either add market-representative categories or to sub-
0
categorise our factors in search for additional entropy. This is akin to industry Dec-03 Dec-05 Dec-07 Dec-09 Dec-11 Dec-13 Dec-15
and sector classification in fundamental equity risk models. In the case of Source: Deutsche Bank
equities and FX, we added emerging market return factors by orthogonalising
EM asset returns against G10 asset returns. In commodities, we split the
market factor into 3: energy, metals 101 and agricultural asset returns.
Separating between developed and emerging markets is not uncommon – the
latter often captures stronger country risk premium – see, for instance,
Podkaminer [2013] and Greenberg et al [2016]. Figure 48 shows the pickup in
explanatory power.
Finally, for completeness, we added two dynamic factors: time series Figure 48: Explanatory power using
Momentum and Carry. While these base factors are typically classified as sub-categorised, market-based
return drivers instead of risk drivers, there will be periods when they drive
factors versus the target
portfolio losses in ways not captured by the market factors outlined earlier. The
0.7
attractive long-term returns in Carry and time series Momentum investing exist Benchmark w/ Cdty sub sectors
so as to compensate for these specific short-term losses.102 The procedure is 0.6
0.5
consistent with Ang [2014].
0.4
0.3
Figure 50 outlines our final list of risk factors.103 We acknowledge a bias; we
0.2
assume these factors, which “worked” in the past, will also “work” in the
0.1
future, but our consistency with the literature is reassuring. Figure 49 shows
0
the explanatory power of our final multi-factor risk model; it looks very similar Sep-95 Sep-99 Sep-03 Sep-07 Sep-11 Sep-15
to the original, PCA-driven target set in Section I.II.I. in other words, we can Source: Deutsche Bank
now explain the desired amount of variance in our pool of 80 markets using
“tangible” – and even hedge-able – drivers.
99
See Ang [2014] for a detailed discussion.
100
We proxied asset class returns through (first) principal component baskets. The PCs are estimated
using a 1-year lookback window of daily returns, rebalanced monthly. The factors are extracted from a
correlation matrix – i.e. we assume unit variances.
101
We chose not to separate between base and precious metals because of the limited breadth in the
latter. Had we sub-categorised, the regressions run on their 1st principal components would have over-
estimated the explanatory power of the model.
102
The “momentum crash” effect is a good example.
103
The reader may also be interested in knowing that the only additional factor considered was “crash
risk”, introduced in David and Bhansali [2014] and proxied as the returns of a short 10-delta 1-month put on
the S&P 500. We ultimately opted against it because it failed to explain much extra variance.

20 April 2017
Figure 50: Our factors - description and how to attain exposure Figure 49: Explanatory power of our
Risk factors How to get exposure final risk factor model versus the
Inflation Return of long nominal US Treasuries, short TIPS (Treasury Inflation- target
Protected securities) portfolio
0.7 Final risk factor model Benchmark
Equity – Developed Broad-developed market equity index return. This is gained through the
Market returns of 1st principal component of our developed market indices. 0.6
Equity – Emerging Return of long EM equity index (1st principal component of emerging 0.5
Market market indices), short DM equity index portfolio. 0.4
Dollar – G10 countries Long 1st principal component of USD/G10 currencies. 0.3
Dollar – Emerging Return of long 1st principal component of USD/EM currencies, short 1st 0.2
Markets principal component of USD/G10 currencies portfolio 0.1
Commodities – Long 1st principal component of our energy basket. 0

Energy sector Sep 95 Sep 99 Sep 03 Sep 07 Sep 11 Sep 15
Commodities – Metals Long 1st principal component of our global metals (industrial and precious) Source: Deutsche Bank
sector basket.
Commodities – Long 1st principal component of our agriculture basket.
Agriculture sector
Rates – Global Long 1st principal component of our sovereign 10Y bond futures market.
Treasuries
Cross-asset Trend Return of our cross-asset trend strategy (Marta).
Cross-asset Carry Return of our cross-asset carry strategy (Carrie)
I.III Model Estimation
Having defined our ingredients in Section I.II, we now move into the final step:
how to estimate the model. Our final goal is to estimate the sensitivity of each
asset to our list of risk factors, as these factor loadings will serve as input to
how we forecast asset volatility in Section 4.
As is commonplace in quantitative research, we first need to address two

natural issues with market data: the interdependence between our factors, and
data outliers. More specifically:
 We addressed multicollinearity risk in 3 ways. First, we orthogonalised

some of the variables; that is, we regressed EM equity and FX returns on
G10 equity and FX returns, and used the respective residuals as our
emerging market factors. Second, we ran the Belsley Collinearity
Diagnosis 104 on a rolling basis to evaluate the degree of collinearity
between our explanatory variables. The results were largely satisfactory.105
Finally, as a means of protecting ourselves against the threat of collinearity
in the future, we applied stepwise regressions when estimating asset
sensitivity to each factor. Through this procedure, only relevant factors
were used at each estimation date; insignificant factors were dropped.106
 We sought to neutralise outlier influence in 2 ways.107 First, we removed
the edges of our factors returns (1st and 99th percentiles). Second, our
104
See Belsley and Kuh [1980] for details.
105
We used 17 years of data in total. The tests were run using 2 years of daily factor returns, rolled daily.
The test statistic did not breach the critical threshold value of 10 at any point of our sample history.
106
We also attempted more complex techniques, namely: partial least squares regressions, principal
component regressions, ridge regressions and LASSO. None showed enough improvement in our results
so as to justify the extra number of parameters required for calibration.
107
Outliers often lead to a violation of the assumption that our regression residuals are Gaussian,
prompting issues such as heteroskedastic and skewed residuals.

20 April 2017
stepwise regressions were conducted in robust form – in other words,

through robust stepwise regressions.108
At every rebalancing date, the steps of our algorithm are as follows:
1. We first take the last 2 years of daily observed returns in both asset and
factor data. We de-trend and remove the outliers109 in both cases.
2. We regress asset returns against each factor using a robust regression

approach:
We first define w as a Nx1 unit vector, where N is the number of data

 
i.
points, and W  uw   where u is a Nx1 unit vector,  is the
T
NxN identity matrix and  is the dot product. W is therefore a

diagonal matrix whose diagonal elements equal the elements of w .
 
ii. We estimate the sensitivity of asset returns to each factor 1
using a
weighted least squares regression, i.e.   X WX
T
X T WY ,
where Y is a Nx1 vector of asset returns, X is a Nx13 matrix of factor
returns (plus the intercept), and  is a 13x1 vector of coefficients.
iii. We redefine w  | s | 1  1  s  s   1  s  s . Further:

r
s , where r is a Nx1 vector of residuals from the
t 1  h
regression in ii,   MAD c , t  4.685 and c  0.6745 are
 
110
, MAD
 
constants 1
stands for mean absolute deviation and
h  diag X X T X X T representing a vector of leverage values
from the current regression fit.111
iv. Repeat steps i – iii until w is stabilized. It represents our robust

weights.
3. We run stepwise weighted regressions of asset returns on each of the risk

factors using the robust weights estimated in Step 2:
i. We run univariate (weighted) regressions on each risk factor and

choose the factor whose P-value is the smallest, assuming it is also
below a qualifying threshold.
ii. We re-run our regressions on every remaining factor, having already

included the term that qualified above, and keep the next relevant
factor – assuming the P-value qualifies.112
108
The method is called iteratively weighted least squares. It involves iteratively re-weighting each data
observation and re-running our regressions through weighted least squares until our betas converge to a
target tolerance level. We use a bi-square weighting function to re-weight our data observations. See
DuMouchel and O’Brien [1989] for details.
109
We remove observations below the 1st percentile and above the 99th percentile of daily returns over the
past 2 years.
110
Setting t = 4.685 gives us coefficients that are 95% as statistically significant as the OLS estimates.
Setting c = 0.6745 makes the estimate unbiased in a normal distribution.
111
See, for instance, https://en.wikipedia.org/wiki/Leverage_(statistics)

20 April 2017
iii. We repeat the step above until there are no qualifiers left. For every
asset, we record the qualifying factor coefficients and set the others to
zero.
The coefficients defined in Step 3 are used as input to our asset volatility
forecasts, as defined earlier in Section 4.3. This completes the estimation of
our cross-asset, macro factor risk model.
112
If the p-value of our regression F-statistic with the new explanatory variable is higher than that of
without it, we do not include the new variable. For reference, see:
https://uk.mathworks.com/help/stats/stepwiselm.html?searchHighlight=stepwiselm&s_tid=doc_srchtitle

20 April 2017
Appendix II: Our Global

Sentiment Indicator
Modeling investor risk appetite has always been a popular topic in both
academic and industry research. In March 2012, we – Chen and Natividade
[2012] – introduced Deutsche Bank’s Global Sentiment Indicator (GSI), a
variable that points to the current state of market sentiment. It is available on
Bloomberg as DBQSGSI Index. Figure 51 shows its historical performance prior
to and after publication.
Figure 51: Global Sentiment Indicator time series

Sentiment indicator [14,15]
[1] [2] [3] [4] [5] [6] [7] [9] [11] [12] [13]
100 [8] [10]
90
80
70
60
50
40
30
20
10
0
1990 1993 1996 1999 2001 2004 2007 2010 2012 2015
Live
[1] Black Wednesday, [2] Mexican devaluation, [3] Asian crisis, [4] Russian crisis, [5] Dot-com bubble, [6] 9/11, [7] Market crash, [8] Credit crunch starts, [9] Lehman Brothers, [10] Greek downgrade, [11] European
debt woes, [12] Taper Tantrum, [13] Greek debt woes, [14] Oil Price Shock, [15] British EU Referendum. Source: Deutsche Bank
The GSI aggregates 11 barometers of risk across asset classes. These were
Figure 52: GSI – Kernel density and
collected through a survey of our colleagues in macroeconomic research and
focused on variables that were highly adaptive and historically persistent – in Gaussian mixture decomposition
other words, they did not predict turbulent periods but adapted quickly to Kernel Density Low Risk
Medium Risk High Risk
them. They are:
 Equity implied volatility: the VIX, a weighted average of implied vols which
helps gauge the market expectation of how volatile the S&P 500 will be
over the next 1 month.
 Financial sector risk (equity perspective): the ratio of the MSCI Financials
Local Index over the MSCI World Index.
 Financial sector risk (rates perspective): the difference between 30Y and 2Y
asset swap spreads, as the former represents possible stress in the
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
-0.1
pension fund and insurance sector while the former represents banking
sector stress. This measure of liquidity risk has been more efficient at Source: Deutsche Bank
market crises prior to the most recent.

 Interest rate implied volatility: the average of USD, EUR and JPY 3M5Y
swaption volatility.
 Short-term interest rate liquidity risk: TED spread, the difference between
interest rates on interbank loans and US T-bills.
 Investment grade credit spreads: the spread between Moody's BAA
corporate bond index and 10Y US Treasuries.
 Non-financial CDS spreads: the iBoxx US Non-Financials Index.

20 April 2017
 FX implied volatility: the CVIX, a turnover-weighted average of 3M implied

volatility which helps gauge the market expectation of how volatile
currency markets will be over the next 1 month.
 FX volatility slope: the average term spread between 1Y and 1M volatility
in EUR/USD, USD/JPY and EUR/JPY.
 FX volatility skew: the average level of 1M 25-delta risk reversals between
USD/JPY, EUR/USD and EUR/JPY.
 EM sovereign risk: JP Morgan's EMBI+ index.
We standardized the data by calculating the rolling 1-year percentile of each Figure 53: GSI levels across historical
series, and distilled the output through its first principal component. 113 The final risk events
variable is bounded between 0 and 100, where higher readings imply a higher Average
Event Start End
risk environment. GSI Level
UK withdraw
from EUR. Jul-92 Oct-92 41
A Gaussian mixture decomposition of the GSI’s empirical distribution pointed ERM
to the existence of 3 distinct regimes of risk appetite – low, medium and high
US rates
risk – as consistent with its tri-modal kernel density. Over the past 27 years the shock/Orange Feb-94 Jun-94 77
GSI has spent roughly 33% of the time in low risk, 40% in the intermediate county
regime and the remaining in high risk. Figure 52 plots this historical Asian
Financial May-97 Feb-98 69
distribution. Crisis
RUS default Jul-98 Sep-98 59
Figure 53 illustrates how efficient the GSI has been at capturing turbulent .COM bubble Feb-00 Jun-00 79
periods in both backtest and live history. In 13 out of the 17 stress periods (and .COM bubble Oct-00 Nov-00 65
4 out of 4 live stress periods) the indicator has been at its upper tercile, .COM bubble Feb-01 May-01 74
suggesting it adapted to high risk conditions when needed. It did not predict
2002
any particular stress period, as it is not designed to do so, but adapted quickly Jul-02 Oct-02 72
downturn
to each of the periods below. Start of US
Mar-04 Jul-04 65
hiking cycle
We have used the GSI for 2 purposes: timing risk-sensitive strategies and Start of
Jun-07 May-08 89
financial crisis
estimating regime-dependent covariance matrices:
Collapse of
Sep-08 Jan-09 88
Lehman
 Timing a strategy implies increasing or decreasing the capital allocated to it
EU Sov. Debt May-10 Jul-10 82
according to an algorithm that is often based on values of an exogenous
EU Sov. Debt Jun-11 Dec-11 80
variable. In this and prior reports, we have used GSI levels to increase or
decrease exposure according to the ratio: Lt ,t  h  2  I t  2 , where Taper Tantrum May-13 Sep-13 72
L  [0,2] . Figure 54 shows the rolling drawdowns on a standard FX Carry Greek Debt
Sep-14 Jul-15 79
Crisis
strategy (DBHVBUSI Index) before and after we applied the algorithm
above, over the long run and since we published the idea in Anand et al. Oil Price
Feb-16 Mar-16 79
Stress
[2014]. Risk indicators, when applied to risk-sensitive strategies, provide
British EU
rare instances when timing is promising. On average, timing is a very hard Referendum
Jun-16 Jul-16 68
task.114 Source: Deutsche Bank
113
We set negative sample covariance values to zero, as per Luo et al [2009]. All data is pre-processed
such that it has the same direction to risk appetite, therefore minimising the instances where we get
negative covariances in the first place.
114
See Asness [2016] for a longer discussion of this topic.

20 April 2017
Figure 54: Rolling FX Carry (DB Balanced Harvest) drawdowns – with and
without GSI timing, historical and since September 2014
0 0
-0.05 -0.05
-0.1 -0.1
-0.15 -0.15
-0.2 -0.2
-0.25 DB FX Carry Index -0.25 DB FX Carry Index
DB FX Carry Index (GSI) DB FX Carry Index (GSI)

-0.3 -0.3
2001 2004 2007 2010 2013 2016 Sep 14 Jun 15 Mar 16 Dec 16
 Regime-based co-movement estimation is another area in which the GSI Figure 55: Volatility of the average
can be applied. Quantitative researchers often calculate co-variances using
pairwise correlation using different
asset returns sequenced in time. Another approach, however, is to
estimate co-movements conditioned on data of the same market regime (rolling) correlation matrices
over different – potentially non-sequential – periods in history. The premise 0.02 Time domain
Low Risk
is that asset dependencies may be better modeled as a function of risk 0.018 Med. Risk
High Risk
states; they may, for instance, be more strongly correlated in risk aversion 0.016
0.014
as one factor – “risk” – becomes the primary driver of asset returns.
0.012
Conversely, they may be less correlated under lower risk conditions as 0.01
multiple drivers affect price action. Figure 55 plots the volatility of the 0.008
average pairwise correlation between 18 cross-asset underlyings, 0.006
bucketed according to 3 GSI states and according to time.115 It shows that 0.004
0.002
the regime-conditioned estimates are less volatile than the estimates 0
calculated in time domain; which suggests the former could potentially 2004 2006 2008 2010 2012 2014 2016
lead to more stable and adaptive asset weights when building portfolios.116 Source: Deutsche Bank
In summary, the GSI seeks to update the investor about current market
conditions in as adaptive way as possible. It can be a powerful tool for
modeling financial data – either for timing or estimating co-movements.
115
We used 4 flagship assets in FX, commodities and global Treasuries, and 5 equity indices. In this
exercise, we effectively applied a bucketing approach to estimate joint relationships. First, we estimated
the 3 risk-state dependent correlation matrices by separating asset returns into 3 buckets based on the GSI
level at the time. In this case, we used an anchored lookback window so as to maximise the number of
data points. In the time domain, however, our correlation matrices were estimated using a lookback
window equal to the average size of 3 risk-state buckets. As such, the number of data points in each
bucket rises in time equally in both domains. Thereafter, we calculated the volatility of average pair-wise
correlation over time as a measure of the stability.
116
Researchers then control for covariance estimation error using shrinkage or factorization. For shrinkage,
see James and Stein [1961], Jorion [1986] and Frost and Savarino [1986]. For factorization, see Sharpe
[1963], Chan et al [1999] and MacKinlay and Pastor [2000].

20 April 2017
8. Bibliography
Ahmad, R. and Wilmott, P. [2005], Which Free Lunch Would You Like Today,
Sir?: Delta Hedging, Volatility Arbitrage and Optimal Portfolios, Wilmott
Magazine (November 2005): 64-79.
Aït-Sahalia, Y., & Mancini, L. [2008]. Out of sample forecasts of quadratic

variation. Journal of Econometrics, 147(1), 17-33.
Alexander, C., & Lazar, E. [2006], Normal mixture GARCH (1, 1): Applications to
exchange rate modeling, Journal of Applied Econometrics, 21(3), 307-336.
Alexander, C., & Lazar, E. [2009], Modelling Regime‐Specific Stock Price

Volatility, Oxford Bulletin of Economics and Statistics, 71(6), 761-797.
Alexander, C., Lazar, E., & Stanescu, S. [2011], Analytic moments for GARCH
processes, ICMA Centre Discussion papers.
Alvarez, M., Luo, Y., Cahan, R., Jussa, J., Chen, Z., Wang, S. [2012]. Portfolios
Under Construction: Risk and Alpha Alignment. Deutsche Bank Quantitative
Strategy.
Anand, V., Natividade, C., Mesomeries, S., Davies, C., Jiang, S. and Capra, J
[2014], Riding Carry, Deutsche Bank Quantcraft, 3 September 2014.
Ang, A. [2014], Asset Management: A Systematic Approach to Factor Investing,

Oxford University Press, 1 Edition.
st
Asness, Clifford S., [2016], The Siren Song of Factor Timing, Journal of
Portfolio Management, Vol. Special Issue, No. 1, 2016.
Bai, J., & Ng, S. [2002], Determining the number of factors in approximate
factor models, Econometrica Vol. 70, No. 1, 191-221
Ball, C. A., & Torous, W. N. [1983], A simplified jump process for common
stock returns, Journal of Financial and Quantitative analysis, 18(01), 53-65.
Bauwens, L., Omrane, W. B., & Giot, P. [2005], News announcements, market
activity and volatility in the euro/dollar foreign exchange market, Journal of
International Money and Finance, 24(7), 1108-1125.
Bauwens, L., Hafner, C. M., & Laurent, S. [2012], Handbook of volatility models
and their applications, Volume 3,. John Wiley & Sons.
Baz, J., Granger, N., Harvey, C., Le Roux, N., Rattray, S. [2015], Dissecting
Investment Strategies In The Cross Section And Time Series, available at
SSRN.
Belsley, D. A., E. Kuh, & R. E. Welsh., Regression Diagnostics. New York, NY:
John Wiley & Sons, Inc., 1980.

20 April 2017
Benzoni, L. Collin-Dufreane, P. and Goldstein, R. [2010], Explaining asset

pricing puzzles associated with the 1987 market crash, Working Paper,
available at SSRN.
Bollerslev, T. [1986], Generalized autoregressive conditional heteroskedasticity,

Journal of econometrics, 31(3), 307-327.
Bollerslev, T. [1990], Modelling the coherence in short-run nominal exchange

rates: a multivariate generalized ARCH model, The review of economics and
statistics, 498-505.
Broadie, M., Chernov, M. and Johannes, M. [2009], Understanding Index

Option Returns, Review of Financial Studies 22(11), 4493-4529.
Burmeister et al [1994] A Practitioner's Guide to Arbitrage Pricing Theory, The

Research Foundation of the Institute of Chartered Financial Analysts.
Carr, P., Jin, X., & Madan, D. B. [2001]. Optimal investment in derivative
securities, Finance and Stochastics, 5(1), 33-59.
Carr, P. and Wu, L. [2009], Variance Risk Premia, Review of Financial Studies
22, 1311-1341.
Cattell, R. B. [1966], The scree test for the number of factors, Multivariate
Behavioral Research, 1, 245-276.
Chan, L., J. Karceski, and J. Lakonishok, [1999], On Portfolio Optimization:

Forecasting Covariances and Choosing the Risk Model, Review of Financial
Studies 12, 937974.
Chen, N., Roll, R. and Ross, S. [1986], Economic Forces and the Stock Market,
Journal of Business, Vol. 59, No. 3 (1986), pp. 383-403.
Chen, L. [2010], Option Pricing And Hedging With Transaction Costs, PhD
Disstertation – Stanford University.
Chen, L., Lai, T. L. and Lim, T. W. [2011], A New Approach to Pricing and
Hedging Options with Transaction Costs, Available at SSRN.
Clenow, A. [2013], Following the Trend: Diversified Managed Futures Trading,

John Wiley & Sons, 1 Edition.
st
Connor, G. [1996], The Three Types of Factor Models: A Comparison of Their

Explanatory Power, Barra Research Insights.
Connor, G. [2000]. Robust confidence intervals for the bias test of risk
forecasts. Technical report, MSCI Barra.
Cont, R. and Tankov, P. [2004], Financial Modelling with jump processes, Rama
Cont, Peter Tankov, Boca Raton, Fla.: Chapman & Hall/CRC.
Corradi, V., Distaso, W., & Mele, A. [2006], Macroeconomic determinants of

stock market volatility and volatility risk-premia, Working Paper, University of
Warwick, UK.

20 April 2017
Cosemans, M. [2011], The Pricing of Long and Short Run Variance and
Correlation Risk in Stock Returns, Working paper, Erasmus University.
Davis, J. and Bhansali, V. [2014], Insuring Against Hidden Financial

Catastrophe Risk, Tail Risk Hedging: Theory and Practice, Risk Books, 1 Edition. st
Derman, E. and Miller, M. [2016], The Volatility Smile, John Wiley & Sons, 1 st
Edition.
DuMouchel, W. H., and F. L. O'Brien [1989], Integrating a Robust Option into a

Multiple Regression Computing Environment, Computer Science and Statistics:
Proceedings of the 21st Symposium on the Interface. Alexandria, VA: American
Statistical Association, 1989.
Engle, R. F. [1982], Autoregressive conditional heteroscedasticity with

estimates of the variance of United Kingdom inflation, Econometrica: Journal
of the Econometric Society, 987-1007.
Engle, R. [2002], Dynamic conditional correlation: A simple class of multivariate

generalized autoregressive conditional heteroskedasticity models, Journal of
Business & Economic Statistics, 20(3), 339-350.
Engle, R. F., & Rangel, J. G. [2005], The spline garch model for unconditional
volatility and its global macroeconomic causes, Working Paper.
Engle, R. F., Ng, V. K., & Rothschild, M. [1990], Asset pricing with a factor-
ARCH covariance structure: Empirical estimates for treasury bills, Journal of
Econometrics, 45(1-2), 213-237.
Franq, C. and Zakoian, J. [2010], GARCH Models: Structure, Statistical Inference

and Financial Applications, John Wiley & Sons, 1 Edition.st
Frost, P. A., and J. E. Savarino, [1986], An empirical Bayes approach to efficient

portfolio selection, Journal of Financial and Quantitative Analysis 21, 293-305.
Gatheral, J. and Jacquier, A. [2013], Arbitrage-free SVI volatility surfaces,

Working Paper, available at SSRN.
Glosten, L. R., Jagannathan, R., and Runkle, D. [1993], On the relation between
the expected value and the volatility of the nominal excess return on stocks,
The journal of finance 48.5 (1993): 1779-1801.
Greenberg et al [2016], Factors to Assets: Mapping Factor Exposures to Asset

Allocations, Special Issue 2016.
Grinold, R. and Kahn, R. [1999], Active Portfolio Management, McGraw Hill, 2 nd
Edition.
Guegan, D., Ielpo, F., & Lalaharison, H. [2013], Option pricing with discrete
time jump processes, Journal of Economic Dynamics and Control, 37(12),
2417-2445.
Haas, M., Mittnik, S., & Paolella, M. S. [2004a], A new approach to Markov-
switching GARCH models, Journal of Financial Econometrics, 2(4), 493-530.

20 April 2017
Haas, M., Mittnik, S., & Paolella, M. S. [2004b]. Mixed normal conditional
heteroskedasticity. Journal of Financial Econometrics, 2(2), 211-250.
Hamilton, J. [2010], Regime switching models, In Macroeconometrics and

Time Series Analysis (pp. 202-209), Palgrave Macmillan UK.
Hodges, S. D. and Neuberger, A. [1989], Optimal Replication of Contingent

Claims under Transaction Costs, Review of Futures Markets, 8, 222–239.
Honore, P. [1998], Pitfalls in estimating jump-diffusion models, Centre for

Analytical Finance, University of Aarhus, Aarhus School of Business, Working
Paper Series No.18.
Ilmanen, A. [2011], Expected Returns: An Investor’s Guide to Harvesting Market

Rewards, John Wiley & Sons, 1 Edition.
st
James, W., and C. Stein, [1961], Estimation with quadratic loss, Proceedings of
the Fourth Berkeley Symposium on Mathematics and Statistics 361-379.
Johnson, N. L. [1949], Systems of frequency curves generated by methods of

translation. Biometrika, 36(1/2), 149-176.
Jorion, P., [1986], Bayes-Stein estimation for portfolio analysis, Journal of

Financial and Quantitative Analysis 21, 279-292.
Kaminski, K. and Lo, A. [2007], When Do Stop-Loss Rules Stop Losses?,

Working Paper, available at SSRN.
Leland, H. [1985], Option Pricing and Replication with Transaction Costs,

Journal of Finance, 5, 1283–1301.
Luo, J., Saks, P. and Satchell, S. [2009], Implementing Risk Appetite in the
Management of Currency Portfolio, Journal of Asset Management, 9, 380-397.
MacKinlay, A. C., and L. Pastor, [2000], Asset pricing models: implications for
expected returns and portfolio selection, Review of Financial Studies 13, 883-
916.
Mandelbrot [1963], The Variation of Certain Speculative Prices, The Journal of

Business, 36(4), 394 – 419.
Marcucci, J. [2005], Forecasting stock market volatility with regime-switching

GARCH models, Studies in Nonlinear dynamics and Econometrics, 9(4), 1-53.
Mccormick, S. and Natividade, C. [2008], “Volatility Drivers Part 1: Does

Liquidity Affect Volatility?”, DB FX Derivatives Focus, 26 February 2008.
Merton, R. C. [1976], Option pricing when underlying stock returns are

discontinuous, Journal of financial economics, 3(1-2), 125-144.
Natividade, C. [2008], Volatility Drivers Part 2: The Cyclical Factors, DB FX

Derivatives Focus, 15 April 2008.
Natividade, C. [2012a], Meet Vanessa, Deutsche Bank Quantcraft, 11 May

2012.

20 April 2017
Natividade, C. [2012b], Catching Trends, Deutsche Bank Quantcraft, 25

October 2012.
Natividade, C. [2013a], Machine (un)learning?, Deutsche Bank Quantcraft, 6

March 2013.
Natividade, C. [2013b], Clock-watchers, Deutsche Bank Quantcraft, 10 April

2013.
Natividade, C. [2013c], Pairing Singles, Deutsche Bank Quantcraft, 13 May

2013.
Natividade, C. and Chen, K. [2011a], Breaking Out?, Deutsche Bank Quantcraft,

01 November 2011.
Natividade, C. and Chen, K. [2011c], Pairing Up, Deutsche Bank Quantcraft, 28

November 2011.
Natividade, C. and Chen, K. [2012a], Well Seasoned Vol Forecasting: SET Goes
Global, Deutsche Bank Quantcraft, 04 January 2012.
Natividade, C. and Chen, K. [2012b], Switching Styles, Deutsche Bank

Quantcraft, March 2012.
Natividade, C., Chen, K. and Wang, Y. [2011], Well Seasoned Vol Forecasting,
Deutsche Bank Quantcraft, 14 November 2011.
Natividade, C., Mesomeris, S., Alvarez, M., Beceren, M. and Davis, C [2013],
Colours of Trend, Deutsche Bank Quantcraft, 17 September 2013.
Natividade, C., Mesomeris, S., Davies, C., Jiang, C., Capra, J., Anand, V. and
Qu, S. [2014], Value Investing: Tricky But Worthy, Deutsche Bank Quantcraft, 3
December 2014.
Natividade, C., Anand, V., Mesomeris, S., Davies, C., Ward, P., Capra, J., Qu, S.
and Osiol, J. [2015], Delving Into New Territories, Deutsche Bank Quantcraft,
10 November 2015.
Natividade, C. and Anand, V. [2016], Trend Following: Asymmetries and

Thresholds, Deutsche Bank Quant Snaps, 4 May 2016.
Natividade, C., Anand, V. and Mesomeris, C. [2016a], Carry: Think Different,

Deutsche Bank Quant Snaps, 16 May 2016.
Oomen, R. and Sheppard, K [2014], Trading on-exchange versus over-the-

counter: some observations from the equity and currency markets,
Unpublished manuscript.
Prasad, R., Stanescu, S., Carter, S., and Jain, S. [2016], Designer Dispersion,
Deutsche Bank Special Report, 5 February 2016.
Podkaminer E., [2013], Risk Factors as Building Blocks for Portfolio

Diversification, Investment Risk and Performance, CFA Institute.

20 April 2017
Ross, Stephen A., [1976], The Arbitrage Theory of Capital Asset Pricing, Journal
of Economic Theory (December): 341-60.
Saravelos, G. and Grover, R. [2017], Why FX volatility is not cheap:

understanding the drivers, DB FX Special Report, 14 March 2017.
Sepp, A. [2013], When You Hedge Discretely: Optimization of Sharpe Ratio for
Delta-Hedging Strategy under Discrete Hedging and Transaction Costs, Journal
of Investment Strategies, 2013, Vol. 3, No.1, pp.19-59.
Sharpe, W., [1963], A simplified model for portfolio analysis, Management

Science 9, 277-293.
Simonatto, J. G. [2013], Approximating the Multivariate Distribution of Time-

Aggregated Stock Returns Under GARCH, Journal of Risk, 16(2), 25-49.
Sinclair, E. [2008], Volatility Trading, John Wiley Y Sons, 1 Edition.

st
Tse, Y. K., & Tsui, A. K. C. [2002], A multivariate generalized autoregressive

conditional heteroscedasticity model with time-varying correlations, Journal of
Business & Economic Statistics, 20(3), 351-362.
Tuenter, H. [2001], An algorithm to determine the parameters of SU-curves in

the johnson system of probabillity distributions by moment matching. Journal
of Statistical Computation and Simulation, 70(4), 325-347.
Valenzuela, M. [2014], Implied Correlation and Expected Returns, Working

Paper, available at SSRN.
Ward, P., Mesomeris, S., Davies, C., Natividade, C., Anand, V., Capra, J., Qu, S.
and Osiol, J. [2016], Exploring The Risk Horizon, DB Quantitative Strategy:
Portfolios Under Construction, 4 February 2016.
Whalley, A. E. and Wilmott, P. [1997]. An Asymptotic Analysis of an Optimal

Hedging Model for Option Pricing with Transaction Costs, Mathematical
Finance, 7 (3), 307–324.
Zakamouline, V. [2006a], Optimal Hedging of Option Portfolios with

Transaction Costs, Working paper, Faculty of Economics, Agder University
College, Norway.
Zakamouline, V. [2006b], Efficient Analytic Approximation of the Optimal

Hedging Strategy for a European Call Option with Transaction Costs,
Quantitative Finance 6:435–445.
Zakamouline, V. [2007], The Best Hedging Strategy in the Presence of

Transaction Costs, Working paper, Faculty of Economics, Agder University
College, Norway.
Zivot, E., [2011], Factor Models for Asset Returns, Lecture Handouts, University
of Washington, 14 March 2011.

20 April 2017
Appendix 1
Important Disclosures
*Other information available upon request
Prices are current as of the end of the previous trading session unless otherwise indicated and are sourced from local
exchanges via Reuters, Bloomberg and other vendors . Other information is sourced from Deutsche Bank, subject
companies, and other sources. For disclosures pertaining to recommendations or estimates made on securities other
than the primary subject of this research, please see the most recently published company report or visit our global
disclosure look-up page on our website at http://gm.db.com/ger/disclosure/DisclosureDirectory.eqsr. Aside from
within this report, important conflict disclosures can also be found at https://gm.db.com/equities under the
"Disclosures Lookup" and "Legal" tabs. Investors are strongly encouraged to review this information before investing.
Analyst Certification
The views expressed in this report accurately reflect the personal views of the undersigned lead analyst(s). In addition,
the undersigned lead analyst(s) has not and will not receive any compensation for providing a specific recommendation
or view in this report. Caio Natividade/Silvia Stanescu/Vivek Anand/Paul Ward/Simon Carter/Pam Finelli/Spyros
Mesomeris
Hypothetical Disclaimer
Backtested, hypothetical or simulated performance results have inherent limitations. Unlike an actual performance
record based on trading actual client portfolios, simulated results are achieved by means of the retroactive application of
a backtested model itself designed with the benefit of hindsight. Taking into account historical events the backtesting of
performance also differs from actual account performance because an actual investment strategy may be adjusted any
time, for any reason, including a response to material, economic or market factors. The backtested performance
includes hypothetical results that do not reflect the reinvestment of dividends and other earnings or the deduction of
advisory fees, brokerage or other commissions, and any other expenses that a client would have paid or actually paid.
No representation is made that any trading strategy or account will or is likely to achieve profits or losses similar to
those shown. Alternative modeling techniques or assumptions might produce significantly different results and prove to
be more appropriate. Past hypothetical backtest results are neither an indicator nor guarantee of future returns. Actual
results will vary, perhaps materially, from the analysis.

20 April 2017
Additional Information
The information and opinions in this report were prepared by Deutsche Bank AG or one of its affiliates (collectively
"Deutsche Bank"). Though the information herein is believed to be reliable and has been obtained from public sources
believed to be reliable, Deutsche Bank makes no representation as to its accuracy or completeness.
If you use the services of Deutsche Bank in connection with a purchase or sale of a security that is discussed in this
report, or is included or discussed in another communication (oral or written) from a Deutsche Bank analyst, Deutsche
Bank may act as principal for its own account or as agent for another person.
Deutsche Bank may consider this report in deciding to trade as principal. It may also engage in transactions, for its own
account or with customers, in a manner inconsistent with the views taken in this research report. Others within
Deutsche Bank, including strategists, sales staff and other analysts, may take views that are inconsistent with those
taken in this research report. Deutsche Bank issues a variety of research products, including fundamental analysis,
equity-linked analysis, quantitative analysis and trade ideas. Recommendations contained in one type of communication
may differ from recommendations contained in others, whether as a result of differing time horizons, methodologies or
otherwise. Deutsche Bank and/or its affiliates may also be holding debt or equity securities of the issuers it writes on.
Analysts are paid in part based on the profitability of Deutsche Bank AG and its affiliates, which includes investment
banking, trading and principal trading revenues.
Opinions, estimates and projections constitute the current judgment of the author as of the date of this report. They do
not necessarily reflect the opinions of Deutsche Bank and are subject to change without notice. Deutsche Bank provides
liquidity for buyers and sellers of securities issued by the companies it covers. Deutsche Bank research analysts
sometimes have shorter-term trade ideas that are consistent or inconsistent with Deutsche Bank's existing longer term
ratings. Trade ideas for equities can be found at the SOLAR link at http://gm.db.com. A SOLAR idea represents a high
conviction belief by an analyst that a stock will outperform or underperform the market and/or sector delineated over a
time frame of no less than two weeks. In addition to SOLAR ideas, the analysts named in this report may from time to
time discuss with our clients, Deutsche Bank salespersons and Deutsche Bank traders, trading strategies or ideas that
reference catalysts or events that may have a near-term or medium-term impact on the market price of the securities
discussed in this report, which impact may be directionally counter to the analysts' current 12-month view of total return
or investment return as described herein. Deutsche Bank has no obligation to update, modify or amend this report or to
otherwise notify a recipient thereof if any opinion, forecast or estimate contained herein changes or subsequently
becomes inaccurate. Coverage and the frequency of changes in market conditions and in both general and company
specific economic prospects make it difficult to update research at defined intervals. Updates are at the sole discretion
of the coverage analyst concerned or of the Research Department Management and as such the majority of reports are
published at irregular intervals. This report is provided for informational purposes only and does not take into account
the particular investment objectives, financial situations, or needs of individual clients. It is not an offer or a solicitation
of an offer to buy or sell any financial instruments or to participate in any particular trading strategy. Target prices are
inherently imprecise and a product of the analyst’s judgment. The financial instruments discussed in this report may not
be suitable for all investors and investors must make their own informed investment decisions. Prices and availability of
financial instruments are subject to change without notice and investment transactions can lead to losses as a result of
price fluctuations and other factors. If a financial instrument is denominated in a currency other than an investor's
currency, a change in exchange rates may adversely affect the investment. Past performance is not necessarily
indicative of future results. Unless otherwise indicated, prices are current as of the end of the previous trading session,
and are sourced from local exchanges via Reuters, Bloomberg and other vendors. Data is sourced from Deutsche Bank,
subject companies, and in some cases, other parties.
The Deutsche Bank Research Department is independent of other business areas divisions of the Bank. Details regarding
our organizational arrangements and information barriers we have to prevent and avoid conflicts of interest with respect
to our research is available on our website under Disclaimer found on the Legal tab.
Macroeconomic fluctuations often account for most of the risks associated with exposures to instruments that promise
to pay fixed or variable interest rates. For an investor who is long fixed rate instruments (thus receiving these cash

20 April 2017
flows), increases in interest rates naturally lift the discount factors applied to the expected cash flows and thus cause a
loss. The longer the maturity of a certain cash flow and the higher the move in the discount factor, the higher will be the
loss. Upside surprises in inflation, fiscal funding needs, and FX depreciation rates are among the most common adverse
macroeconomic shocks to receivers. But counterparty exposure, issuer creditworthiness, client segmentation, regulation
(including changes in assets holding limits for different types of investors), changes in tax policies, currency
convertibility (which may constrain currency conversion, repatriation of profits and/or the liquidation of positions), and
settlement issues related to local clearing houses are also important risk factors to be considered. The sensitivity of fixed
income instruments to macroeconomic shocks may be mitigated by indexing the contracted cash flows to inflation, to
FX depreciation, or to specified interest rates – these are common in emerging markets. It is important to note that the
index fixings may -- by construction -- lag or mis-measure the actual move in the underlying variables they are intended
to track. The choice of the proper fixing (or metric) is particularly important in swaps markets, where floating coupon
rates (i.e., coupons indexed to a typically short-dated interest rate reference index) are exchanged for fixed coupons. It is
also important to acknowledge that funding in a currency that differs from the currency in which coupons are
denominated carries FX risk. Naturally, options on swaps (swaptions) also bear the risks typical to options in addition to
the risks related to rates movements.
Derivative transactions involve numerous risks including, among others, market, counterparty default and illiquidity risk.
The appropriateness or otherwise of these products for use by investors is dependent on the investors' own
circumstances including their tax position, their regulatory environment and the nature of their other assets and
liabilities, and as such, investors should take expert legal and financial advice before entering into any transaction similar
to or inspired by the contents of this publication. The risk of loss in futures trading and options, foreign or domestic, can
be substantial. As a result of the high degree of leverage obtainable in futures and options trading, losses may be
incurred that are greater than the amount of funds initially deposited. Trading in options involves risk and is not suitable
for all investors. Prior to buying or selling an option investors must review the "Characteristics and Risks of Standardized
Options”, at http://www.optionsclearing.com/about/publications/character-risks.jsp. If you are unable to access the
website please contact your Deutsche Bank representative for a copy of this important document.
Participants in foreign exchange transactions may incur risks arising from several factors, including the following: ( i)
exchange rates can be volatile and are subject to large fluctuations; ( ii) the value of currencies may be affected by
numerous market factors, including world and national economic, political and regulatory events, events in equity and
debt markets and changes in interest rates; and (iii) currencies may be subject to devaluation or government imposed
exchange controls which could affect the value of the currency. Investors in securities such as ADRs, whose values are
affected by the currency of an underlying security, effectively assume currency risk.
Unless governing law provides otherwise, all transactions should be executed through the Deutsche Bank entity in the
investor's home jurisdiction. Aside from within this report, important conflict disclosures can also be found at
https://gm.db.com/equities under the "Disclosures Lookup" and "Legal" tabs. Investors are strongly encouraged to
review this information before investing.
United States: Approved and/or distributed by Deutsche Bank Securities Incorporated, a member of FINRA, NFA and
SIPC. Analysts located outside of the United States are employed by non-US affiliates that are not subject to FINRA
regulations.
Germany: Approved and/or distributed by Deutsche Bank AG, a joint stock corporation with limited liability incorporated
in the Federal Republic of Germany with its principal office in Frankfurt am Main. Deutsche Bank AG is authorized under
German Banking Law and is subject to supervision by the European Central Bank and by BaFin, Germany’s Federal
Financial Supervisory Authority.
United Kingdom: Approved and/or distributed by Deutsche Bank AG acting through its London Branch at Winchester
House, 1 Great Winchester Street, London EC2N 2DB. Deutsche Bank AG in the United Kingdom is authorised by the
Prudential Regulation Authority and is subject to limited regulation by the Prudential Regulation Authority and Financial
Conduct Authority. Details about the extent of our authorisation and regulation are available on request.
Hong Kong: Distributed by Deutsche Bank AG, Hong Kong Branch.
India: Prepared by Deutsche Equities India Pvt Ltd, which is registered by the Securities and Exchange Board of India

20 April 2017
(SEBI) as a stock broker. Research Analyst SEBI Registration Number is INH000001741. DEIPL may have received
administrative warnings from the SEBI for breaches of Indian regulations.
Japan: Approved and/or distributed by Deutsche Securities Inc.(DSI). Registration number - Registered as a financial
instruments dealer by the Head of the Kanto Local Finance Bureau (Kinsho) No. 117. Member of associations: JSDA,
Type II Financial Instruments Firms Association and The Financial Futures Association of Japan. Commissions and risks
involved in stock transactions - for stock transactions, we charge stock commissions and consumption tax by
multiplying the transaction amount by the commission rate agreed with each customer. Stock transactions can lead to
losses as a result of share price fluctuations and other factors. Transactions in foreign stocks can lead to additional
losses stemming from foreign exchange fluctuations. We may also charge commissions and fees for certain categories
of investment advice, products and services. Recommended investment strategies, products and services carry the risk
of losses to principal and other losses as a result of changes in market and/or economic trends, and/or fluctuations in
market value. Before deciding on the purchase of financial products and/or services, customers should carefully read the
relevant disclosures, prospectuses and other documentation. "Moody's", "Standard & Poor's", and "Fitch" mentioned in
this report are not registered credit rating agencies in Japan unless Japan or "Nippon" is specifically designated in the
name of the entity. Reports on Japanese listed companies not written by analysts of DSI are written by Deutsche Bank
Group's analysts with the coverage companies specified by DSI. Some of the foreign securities stated on this report are
not disclosed according to the Financial Instruments and Exchange Law of Japan. Target prices set by Deutsche Bank's
equity analysts are based on a 12-month forecast period.
Korea: Distributed by Deutsche Securities Korea Co.
South Africa: Deutsche Bank AG Johannesburg is incorporated in the Federal Republic of Germany (Branch Register
Number in South Africa: 1998/003298/10).
Singapore: by Deutsche Bank AG, Singapore Branch or Deutsche Securities Asia Limited, Singapore Branch (One Raffles
Quay #18-00 South Tower Singapore 048583, +65 6423 8001), which may be contacted in respect of any matters
arising from, or in connection with, this report. Where this report is issued or promulgated in Singapore to a person who
is not an accredited investor, expert investor or institutional investor (as defined in the applicable Singapore laws and
regulations), they accept legal responsibility to such person for its contents.
Taiwan: Information on securities/investments that trade in Taiwan is for your reference only. Readers should
independently evaluate investment risks and are solely responsible for their investment decisions. Deutsche Bank
research may not be distributed to the Taiwan public media or quoted or used by the Taiwan public media without
written consent. Information on securities/instruments that do not trade in Taiwan is for informational purposes only and
is not to be construed as a recommendation to trade in such securities/instruments. Deutsche Securities Asia Limited,
Taipei Branch may not execute transactions for clients in these securities/instruments.
Qatar: Deutsche Bank AG in the Qatar Financial Centre (registered no. 00032) is regulated by the Qatar Financial Centre
Regulatory Authority. Deutsche Bank AG - QFC Branch may only undertake the financial services activities that fall
within the scope of its existing QFCRA license. Principal place of business in the QFC: Qatar Financial Centre, Tower,
West Bay, Level 5, PO Box 14928, Doha, Qatar. This information has been distributed by Deutsche Bank AG. Related
financial products or services are only available to Business Customers, as defined by the Qatar Financial Centre
Regulatory Authority.
Russia: This information, interpretation and opinions submitted herein are not in the context of, and do not constitute,
any appraisal or evaluation activity requiring a license in the Russian Federation.
Kingdom of Saudi Arabia: Deutsche Securities Saudi Arabia LLC Company, (registered no. 07073-37) is regulated by the
Capital Market Authority. Deutsche Securities Saudi Arabia may only undertake the financial services activities that fall
within the scope of its existing CMA license. Principal place of business in Saudi Arabia: King Fahad Road, Al Olaya
District, P.O. Box 301809, Faisaliah Tower - 17th Floor, 11372 Riyadh, Saudi Arabia.
United Arab Emirates: Deutsche Bank AG in the Dubai International Financial Centre (registered no. 00045) is regulated
by the Dubai Financial Services Authority. Deutsche Bank AG - DIFC Branch may only undertake the financial services

20 April 2017
activities that fall within the scope of its existing DFSA license. Principal place of business in the DIFC: Dubai
International Financial Centre, The Gate Village, Building 5, PO Box 504902, Dubai, U.A.E. This information has been
distributed by Deutsche Bank AG. Related financial products or services are only available to Professional Clients, as
defined by the Dubai Financial Services Authority.
Australia: Retail clients should obtain a copy of a Product Disclosure Statement (PDS) relating to any financial product
referred to in this report and consider the PDS before making any decision about whether to acquire the product. Please
refer to Australian specific research disclosures and related information at
https://australia.db.com/australia/content/research-information.html
Australia and New Zealand: This research is intended only for "wholesale clients" within the meaning of the Australian
Corporations Act and New Zealand Financial Advisors Act respectively.
Additional information relative to securities, other financial products or issuers discussed in this report is available upon
request. This report may not be reproduced, distributed or published without Deutsche Bank's prior written consent.
Copyright © 2017 Deutsche Bank AG

David Folkerts-Landau
Group Chief Economist and Global Head of Research
Raj Hindocha Michael Spencer Steve Pollard

Global Chief Operating Officer Head of APAC Research Head of Americas Research
Research Global Head of Economics Global Head of Equity Research
Anthony Klarman Paul Reynolds Dave Clark Pam Finelli

Global Head of Head of EMEA Head of APAC Global Head of
Debt Research Equity Research Equity Research Equity Derivatives Research
Andreas Neubauer Stuart Kirk

Head of Research - Germany Head of Thematic Research
International Locations
Deutsche Bank AG Deutsche Bank AG Deutsche Bank AG Deutsche Securities Inc.
Deutsche Bank Place Große Gallusstraße 10-14 Filiale Hongkong 2-11-1 Nagatacho
Level 16 60272 Frankfurt am Main International Commerce Centre, Sanno Park Tower
Corner of Hunter & Phillip Streets Germany 1 Austin Road West,Kowloon, Chiyoda-ku, Tokyo 100-6171
Sydney, NSW 2000 Tel: (49) 69 910 00 Hong Kong Japan
Australia Tel: (852) 2203 8888 Tel: (81) 3 5156 6770
Tel: (61) 2 8258 1234
Deutsche Bank AG London Deutsche Bank Securities Inc.
1 Great Winchester Street 60 Wall Street
London EC2N 2EQ New York, NY 10005
United Kingdom United States of America
Tel: (44) 20 7545 8000 Tel: (1) 212 250 2500

Deutsche Bank - Volatility Risk Premium

Uploaded by

Copyright:

Available Formats

Deutsche Bank - Volatility Risk Premium

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deutsche Bank - Volatility Risk Premium

Uploaded by

Copyright:

Available Formats

Provided for the exclusive use of harel@oportod.com on 2020-09-15T09:46+00:00.

Europe Derivatives Strategy Date

Volatility Risk Premium: New caio.natividade@db.com

Dimensions Silvia Stanescu

Source: Getty Images.

While we assume some technical knowledge, the report is aimed at both

 Section 2 looks at the primary drivers of options returns – our starting

Page 2 Deutsche Bank AG/London

As is often the case in quantitative research, some results are not as

Deutsche Bank AG/London Page 3

2. The Primary Factors

Page 4 Deutsche Bank AG/London

-0.08 Option delta

Today’s report focuses on the PC1 of delta-hedged returns, whose historical

dynamic; we should not just blindly sell options.

Deutsche Bank AG/London Page 5

3. Exploring the Volatility

Page 6 Deutsche Bank AG/London

Figure 6: Cumulative, volatility-adjusted pre-cost 1-month VRP basket returns

Deutsche Bank AG/London Page 7

cross-market replication, factor timing and buying tail options.

Cross-market replication improves the return profile – the expected shortfall

Page 8 Deutsche Bank AG/London

3.2 Market Timing

To test it, we created a linearly continuous leverage mechanism to the VRP

Deutsche Bank AG/London Page 9

Page 10 Deutsche Bank AG/London

In the remaining sections, we seek to introduce methods that are distinctly

Deutsche Bank AG/London Page 11

This abundance of market-implied views is valuable to the quantitative investor

4.1 P-distribution Basics

 We use our delta-one portfolios to estimate expected returns - the first

Page 12 Deutsche Bank AG/London

 We account for volatility jumps by segregating between event and non-

Vart 1 rt ,i   Bi  t| t 1 BiT  t t 1

 t ,i is determined exogenously by our delta-one signals. It is non-

  t is a time-varying factor covariance matrix, our systematic risk. It is

4.2 P-Distribution: the Mean

As the P-distribution should represent our "best guess" of future values of

For any horizon, our forecasted expected return,  , is computed as follows:

Deutsche Bank AG/London Page 13

 t ,i  f  rt ,i , rt 1,i ,..., r0,i   w

4.3 P-Distribution: the Variance

To model  t we apply a multivariate factor constant conditional correlation

 Common factor vs idiosyncratic risk: we discriminate between systematic

 Computational efficiency. In essence, factor models reduce dimensionality.

Page 14 Deutsche Bank AG/London

variations better. As has been argued in the literature 31 , the inter-

Vart 1 rt ,i   Bi  t| t 1 BiT  t t 1 Investable: Commodities:

is a KxK factor correlation matrix.   diag h  is a

: The estimation of our CCC-MGARCH includes the following steps:

Deutsche Bank AG/London Page 15

The variance of the k-th de-meaned factor Fk,, conditional on information

vart 1 Ft , k   ht , k  k   k Ft 21, k  k , st 1 Ft 21, k   k ht 1, k

vart 1  t ,i   ht , i  i   i t21,i  i ,st 1  t21,i   i ht 1, i

where st is an observable random variable. At any time t, we assume that st is

For parameter estimation via maximum likelihood, as well as forecasting

ht ,k   k   k Ft 21,k  1,k Ft 21,k I t11 ,k  2 Ft 21,k I t21,k   k ht 1,k

4.3.3 Event versus non-event volatility

k ,3  0 . The same treatment applies to i ,s t 1