Understanding the Connection

between Real-World and Risk-

Neutral Scenario Generators
AUGUST | 2022
 2

Understanding the Connection between

Real-World and Risk-Neutral Scenario
Generators

AUTHOR Stephen J. Strommen, FSA, CERA, MAAA SPONSOR Quantitative Finance and Investment
Curriculum Committee

Caveat and Disclaimer

The opinions expressed and conclusions reached by the authors are their own and do not represent any official position or opinion of the Society of
Actuaries Research Institute, the Society of Actuaries or its members. The Society of Actuaries Research Institute makes no representation or warranty to
the accuracy of the information.

Copyright © 2022 by the Society of Actuaries Research Institute. All rights reserved.

Copyright © 2022 Society of Actuaries Research Institute

 3

CONTENTS
Introduction.............................................................................................................................................................. 4

Section 1: Why are there two kinds of generators? (Fitness for purpose) .................................................................. 5
Section 2: What it means to be market-consistent .................................................................................................... 8
2.1 Reflecting expectations ......................................................................................................................................... 8
2.2 Common mathematical form of scenario generators .......................................................................................... 9
2.3 The common connection between real-world and risk-neutral calibrations ...................................................... 9
Section 3: Interest rate models ............................................................................................................................... 10
3.1 Market behavior in interest rate models ............................................................................................................11
3.2 The market price of risk in interest rate models ................................................................................................13
Section 4: Equity index models ............................................................................................................................... 16
4.1 Market behavior in equity index models ............................................................................................................16
Alternatives for the expected drift term ......................................................................................................16
Alternatives for the random shock term ......................................................................................................17
4.2 Market price of risk in equity index models .......................................................................................................17
Section 5: Calibration of market-consistent and market-coherent models .............................................................. 18
5.1 General approach ................................................................................................................................................18
5.2 Risk-neutral calibration .......................................................................................................................................18
5.3 Real-world calibration .........................................................................................................................................19
5.4 Combined calibration: a market-coherent real-world model ............................................................................20

Section 6: Acknowledgments .................................................................................................................................. 23

Appendix A: The P-measure and the Q-measure .................................................................................................... 24
Interest rates .............................................................................................................................................................24
Equity returns ............................................................................................................................................................27

Appendix B: Some more complex interest rate models .......................................................................................... 28

Double mean reverting models.................................................................................................................................28
The three-factor Cox-Ingersoll-Ross model ..............................................................................................................30

Appendix C: The meaning of arbitrage-free ............................................................................................................ 32

Appendix D: Term premiums and the market price of risk ...................................................................................... 33
Notes on calibration of term premiums ...................................................................................................................35
Appendix E: Fair value, Principle-based Reserves, and the Valuation Manual ......................................................... 37
Obtaining scenario-specific values ............................................................................................................................37
Blending scenario-specific values into a single reported value ................................................................................38
Selecting a scenario generator that is fit for purpose ..............................................................................................39

References .............................................................................................................................................................. 40
About The Society of Actuaries Research Institute .................................................................................................. 41

Copyright © 2022 Society of Actuaries Research Institute

 4

Understanding the Connection between Real-

World and Risk-Neutral Scenario Generators

Introduction
Scenario generators are sometimes described as being as either “real-world” or “risk-neutral”. The difference
between real-world and risk-neutral generators is in the treatment of the market price of risk. This document
focuses on the difference in treatment of the market price of risk, and helps the reader understand the relationship
between real-world and risk-neutral scenarios and when it is appropriate to use each type.

Risk-neutral generators are referred to as “market-consistent” because they are calibrated to reproduce market
prices of traded financial instruments. It is often assumed that real-world generators cannot reproduce market
prices, but in fact the ability to reproduce market prices is a matter of calibration and a real-world generator can be
calibrated to reproduce market prices. We introduce the term “market-coherent” to describe a real-world
generator that is calibrated to reproduce market prices, explain how such calibration can be done, and discuss why
it can be useful. The term “market-coherent” was chosen because such a generator not only reproduces current
market prices but also reflects the real-world evolution of those prices into the future by using the P-measure (real-
world probabilities) rather than the Q-measure (risk-neutral probabilities).

To help illustrate how real-world and risk-neutral methods can be calibrated to produce the same market price, an
Excel workbook has been prepared to accompany this paper. The workbook illustrates the valuation of a simple put
option on equity stock using real-world and risk-neutral methods. The probability distributions associated with real-
world and risk-neutral probabilities are shown side-by-side, and the means by which the methods use different
probabilities to arrive at the same value are shown in detail.

Copyright © 2022 Society of Actuaries Research Institute

 5

Section 1: Why are there two kinds of generators? (Fitness for purpose)
Before addressing the connection between real-world and risk-neutral scenario generators, it is important to
understand why there are two different kinds of generators in the first place. The two kinds of generators were
created to address two fundamentally different problems. That means they are fit for different purposes, and it is
important when choosing a generator to choose one that is fit for the purpose at hand.

Real-world generators were created to provide realistic stochastic simulation of the path of future economic
conditions through time. Each path is a scenario, and such realistic but stochastically generated scenarios can be
used in financial models to quantify risk. By running simulations using such scenarios, one can try to answer the
questions “How bad could things get” and “How likely is that?” in the context of risk management for a financial
institution.

Therefore, real-world scenario generators are fit for the purpose of simulation across time.

Risk-neutral generators were created for market-consistent valuation of options and other derivatives with an
uncertain payoff. A market-consistent valuation method is one which reproduces the market price of market-traded
instruments including options and derivatives. Not all such options and derivatives are market-traded and have
observable market prices. When buying or selling a derivative that is not traded widely, it is very useful to
determine its value by comparison with similar derivatives that are widely traded. It is useful to determine a
“market-consistent” price or value. Risk-neutral methods, and generators based upon them, allow determination of
a market-consistent value without making any assumption about the market price of risk. That is useful because the
price of risk is not directly observable, but the market prices used to calibrate a risk-neutral generator are
observable.

Therefore, risk-neutral scenario generators are fit for the purpose of valuing an item with an uncertain payoff in a
market-consistent way.

The world would be much simpler if this distinction in purpose were always clear-cut. It is more complex when the
problem at hand is to value an item with an uncertain payoff by means of stochastic simulation across time. In the
realm of accounting for insurance contracts, that kind of valuation approach has been widely adopted. In that
context the choice as to which kind of scenario generator is fit for the purpose requires understanding the
difference between them at a deeper level.

One key to understanding this difference is to consider the universe of all possible scenarios. In abstract terms, any
scenario generator is a means of picking scenarios from that universe. The parameters of the generator essentially
assign a probability of selection to each possible scenario. The probability of selection differs in a fundamental and
systematic way between real-world and risk-neutral scenarios. Real-world scenarios use the P-measure, or real-
world probabilities. Risk-neutral scenarios use the Q-measure, or risk-neutral probabilities. These different
measures assign different probabilities to the same set of events in the future.

The chart below illustrates this concept in a very simple case. Suppose the problem at hand involves projecting the
future value of a share of equity stock one year from now. The universe of possible scenarios ranges from very low
value to very high value. At the starting date of the projection, the real-world probabilities are for a mean expected
value near the middle of that range. Real-world scenarios are probability weighted around that mean expected
value.

Copyright © 2022 Society of Actuaries Research Institute

 6

Figure 1
DISTRIBUTION OF REAL-WORLD SCENARIOS

What about risk-neutral scenarios? The next chart adds the probability distribution for a set of risk-neutral
scenarios. Note that there is a lot of overlap between the risk-neutral and real-world distributions, indicating that
many of the same scenarios may appear in both sets. Note also that the risk-neutral distribution is not centered on
the same central value: it is centered on a lower value. This means that the average projected value of a risky
investment in risk-neutral scenarios will be lower than in real-world scenarios. The probability weighting in risk-
neutral scenarios (Q-measure) gives more weight to adverse results (lower projected value in this case) than the P-
measure. For example, the central value in the risk-neutral probability weighting is based on the price increasing at
the risk-free rate. That rate is typically lower than a realistic market expectation that would include a risk premium.

Figure 2
DISTRIBUTION OF REAL-WORLD AND RISK-NEUTRAL SCENARIOS

Copyright © 2022 Society of Actuaries Research Institute

 7

In these charts the horizontal axis is a projected value. Keep in mind that the starting value is the same in both real-
world and risk-neutral scenarios. And, importantly, the expected present value of the stochastic future price is the
same in both the real-world and risk-neutral scenario sets. That’s because the discount rate differs between real-
world and risk-neutral methods. The risk-neutral discount rate is always the risk-free short-term rate. The real-
world discount rate is higher because it includes a risk premium for the equity price risk. That risk premium exists
because real-world markets are risk-averse; an uncertain future price is discounted at a rate that includes a risk
premium related to the degree of uncertainty in the future price. Therefore, projected values and prices are higher
in real-world market-coherent scenarios than in risk-neutral scenarios, but the present values are the same. This
means that projections of value based on risk-neutral scenarios diverge more and more from real-world projections
the further into the future you go. Risk-neutral methods are clearly meant for market-consistent valuation on the
scenario start date, where they match current real-world prices, but not for projections or simulations that involve
decision-making based on simulated conditions on dates in the future. In risk-neutral scenarios, simulated future
conditions do not have a realistic probability distribution.

This has implications when using stochastic scenario methods for valuation of insurance contracts or pension
obligations. When the valuation basis is market value, risk-neutral and real-world methods can (in theory) produce
the same present value in cases where there are no future decisions to be made based on future market conditions.
That would be the case for a block of insurance contracts or pensions backed by a true replicating portfolio. But in
many cases simulation of obligations involves decisions to be made by either the administrator or the consumer
based on future conditions1. These decisions have financial effects that are not considered in the calibration of risk-
neutral scenarios because that calibration is based on other financial instruments. Therefore, use of the risk-neutral
probabilities (Q-measure) to probability-weight the financial effects of such decisions is less than ideal because
those are not the real probabilities. It would be ideal if the real-world probabilities (P-measure) could be used.

Using real-world scenarios and the P-measure for this purpose requires that they be calibrated to reproduce market
prices. Such a calibration requires an assumption concerning the market price of risk, which is not directly
observable. Because of this, and since risk-neutral methods and the Q-measure do not require an assumption for
the market price of risk, risk-neutral methods are commonly used. Nevertheless, the market price of risk can be
inferred indirectly using some generally accepted models. With a reasonable assumption for the market price of
risk, the P-measure can be calibrated to reproduce market prices. We will call real-world scenarios based on such a
calibration “market-coherent” because they not only reproduce current market prices but also employ real-world
probabilities (the P-measure) concerning the evolution of financial conditions in the future. Market-coherent
scenarios can be used in valuation of the kind of obligations described above.

What about real-world scenarios that are not market-coherent? Such generators can be useful for risk
management or in a regulatory context where black swan events are an important consideration. The next chart
illustrates one possible distribution for such a set of scenarios. In this case the distribution is wider with longer tails
and would not reproduce market prices of many derivatives. For this kind of purpose, the tail scenarios are the
focus of attention, and the exact center of the distribution is less important.

1 Such future decisions on the part of the insured may include whether to keep a contract in force because guarantees have become valuable or surrender
it because guarantees no longer have much value. Decisions on the part of the insurer may include how to invest future renewal premiums. The
distribution of future market prices and interest rates that would drive future investment decisions are not realistic in risk-neutral scenarios.

Copyright © 2022 Society of Actuaries Research Institute

 8

Figure 3
DISTRIBUTION OF REAL-WORLD SCENARIOS WITH EXTRA TAIL RISK

Section 2: What it means to be market-consistent

2.1 REFLECTING EXPECTATIONS

The term “market-consistent” means a valuation or a set of scenarios must be consistent with current market
prices.

Market prices for financial instruments reflect expectations of the future. For example, the market price of fixed
income investments will be lower if it is expected that interest rates will rise and higher if it is expected that interest
rates will fall in the future. The market price of equity investments like corporate stock depends on expectations
regarding future earnings. The market price of options depends not only on the strike price but also on the
expected volatility of market prices over the term of the contract because the probability of an out-of-the money
option having value depends on the volatility.

A calibration that reproduces current market prices focuses mainly on two aspects of market expectations:

• the central path or direction of future prices

• the volatility around that path; how far actual returns might vary from the expected path

Calibration of a scenario generator to reproduce current market prices involves using observed market prices to
quantify the central path and volatility around that path. Real-world and risk-neutral calibration methods infer
different central paths and different volatilities. If real-world and risk-neutral scenarios were used in the same way
to calculate a present value, they would produce different values. But real-world valuation methods and risk-neutral
valuation methods use the scenarios differently, and that difference allows them to both reproduce the same
market price. An Excel workbook accompanying this paper illustrates this idea using an example valuation of a put
option on equity stock.

Copyright © 2022 Society of Actuaries Research Institute

 9

Since both real-world and risk-neutral methods can be calibrated to reproduce market prices, under the definition
above they could both be categorized as market-consistent when calibrated that way. Common usage has come to
associate the term “market-consistent” only with the risk-neutral approach. To avoid confusion, the term “market-
coherent” can be used to refer to a real-world generator calibrated to reproduce market prices.

Scenario generators also need to reflect understandings about how markets behave. These include the idea that
markets are arbitrage-free and that market participants are risk-averse. Scenario generators can be designed and
calibrated to embody those expectations as well.

There are some popular forms of generators that are not arbitrage-free. These are commonly used for real-world
simulation. Since they are not arbitrage-free they do not reproduce market prices. There is generally no risk-
neutral version of such models, so one cannot explain the connection between risk-neutral and real-world
calibrations and they will not be discussed further in this document.

2.2 COMMON MATHEMATICAL FORM OF SCENARIO GENERATORS

The common mathematical form of a scenario generator is simple:

(next period value) = (current value) + (expected drift) + (random shock)

The expected drift is calibrated to the central expectation of the change in value. The random shock is calibrated to
volatility in the value.

2.3 THE COMMON CONNECTION BETWEEN REAL-WORLD AND RISK-NEUTRAL CALIBRATIONS

Consider the common mathematical form shown above. Taken together, the terms for the expected drift and the
random shock define a probability distribution for next period’s value. A key to understanding the connection
between real-world and risk-neutral models is that the distribution in a real-world calibration is different from the
distribution in a risk-neutral calibration. In particular, the drift term is always different between real-world and risk-
neutral calibrations.

The risk-neutral approach to calibration is based on the idea that the market price of risk is zero and the expected
drift in the price of all investments is the short-term risk-free rate. But since the real-world drift in price is higher for
riskier investments, the lower expected drift in a risk-neutral calibration amounts to giving more probability weight
to adverse events that result in lower future market prices. That is how a risk-neutral calibration reflects risk
aversion; it effectively gives more probability weight to adverse future events.

This is the key to understanding why the intended usage of risk-neutral scenarios is limited to valuation. Risk-
neutral scenarios are unrealistic in that they embed an implicit margin for risk by giving more weight to adverse
events – that is, events that lead to lower market prices. But that is exactly why they are useful for valuation
purposes. Valuations using risk-neutral techniques include risk margins implicitly and avoid the need to add an
explicit risk margin. That is why, when using risk-neutral scenarios, the estimated value is an average across all
scenarios with no add-on margin, but when using real-world scenarios some sort of risk margin needs to be
included. One common means of adding a risk margin in real-world valuations is to set the value to the contingent

Copyright © 2022 Society of Actuaries Research Institute

 10

tail expectation (CTE), the average of the subset of scenarios containing the most adverse results rather than the
average across all scenarios.2

There are countless different implementations of the common mathematical form shown above. The
implementation depends on the kind of value for which scenarios are to be generated, and on the kind of market
behavior over time that one wants to simulate. The following sections illustrate some generator formulations for
interest rates and for a stock price index that includes dividend re-investment. Differences between various
formulations are explained in terms of the market behavior they simulate and the way the market price of risk is
embedded.

Section 3: Interest rate models

Arbitrage-free interest rate models focus on the path of the short-term risk-free interest rate. The remainder of the
risk-free yield curve at any time is defined as an expectation based upon the path of the short-term risk-free rate.
The expectation operator is applied to the discounted value of a future payment using path-wise discounting. The
expected discounted values are treated as spot prices, and those spot prices define spot rates that represent the
yield curve.

If the short-term risk-free rate at time t is 𝑟𝑡 then the spot price P for maturity t=T is the risk-neutral expectation3:

𝑃𝑇 = 𝐸 [𝑒𝑥𝑝 (− ∫ 𝑟𝑡 𝑑𝑡)]
0

The expression for the spot price given above applies to risk-neutral calibrations. Real-world calibrations simulate
movement of the short-term rate differently and therefore are not consistent with this formula for the spot price.

Why is the formula defining movement of the short-term rate different between real-world and risk-neutral
models? The reason is the difference in purpose. The purpose of a real-world model is for simulation of the actual
path of the short-term risk-free rate. The purpose of a risk-neutral model is valuation, so the model must produce
forward paths of short-term rates that reproduce the prices of pure discount bonds based on the formula shown
above. The forward path in a risk-neutral model is different from the path in a real-world model because of the risk
associated with locking in the fixed interest rate in a long-term pure discount bond. That risk is reflected in the

2 Another common means of adding a risk margin in real-world valuation is to start with the average value across all real-world scenarios and then add an
amount that represents the present value of the cost of capital required to manage the risk of the item being valued. The theory in that approach is that
the cost of capital represents the market price of risk.
3 The spot price is not the same as the discounted value using the expected path of the short-term rate, which would be expressed as follows:

𝑃𝑇 ≠ [𝑒𝑥𝑝 (− ∫ 𝐸(𝑟𝑡 )𝑑𝑡)]

0

This expression may have a very similar value, but it is not the same. Since both expressions involve an expectation, they are often confused. The situation
here is somewhat analogous to measures of central tendency in a probability distribution, where the mean and the median are both measures of central
tendency, but they are not the same.

Copyright © 2022 Society of Actuaries Research Institute

 11

volatility of the market price of a pure discount bond – volatility that does not exist in the price of a short-term
money market fund earning the short-term risk-free rate. Investors in pure discount bonds demand a higher return
as a reward for taking that risk, so the forward rate paths consistent with the price of a pure discount bond are
higher than the forward paths of the truly risk-free short-term rate.

Previously it was said that risk-neutral scenarios give more weight to adverse events, meaning declines in prices. Yet
now we say that forward rate paths in a risk-neutral model are higher than in a real-world projection. How is this
consistent? It is consistent because an upward movement in the forward rate leads to a decline in the price. Higher
interest rates mean lower prices. Giving more weight to adverse events that decrease the market price means
giving more weight to increases in interest rates that decrease the market price.

Risk-neutral scenarios are not intended for use in simulation where those future higher interest rates might lead to
higher returns on future new investments. Risk-neutral scenarios are intended for use in valuation of investments
purchased in the past. Any use of risk-neutral scenarios to simulate investment purchases in the future violates the
intended purpose of such scenarios for these reasons4:

• It results in purchase of future investments at higher yields (i.e., more favorable future conditions)
than the market actually expects.

• It is inconsistent with the financial theory underlying risk-neutral models which was developed for the
purpose of valuation of existing investments, not simulation of future investments.

Given a formulation for 𝑟𝑡 it is often possible to derive a closed-form expression for the spot price consistent with
the formula above involving the central expectation along possible paths. A closed-form expression is desirable
because it allows the shape of the full yield curve to be calculated directly. When such a closed-form expression
does not exist, spot prices and the yield curve are approximated using Monte Carlo simulation of many scenarios for
𝑟𝑡 to estimate the central expectation in the formula above.

3.1 MARKET BEHAVIOR IN INTEREST RATE MODELS

There are two important and separate characteristics to any arbitrage-free model for interest rates, and both center
on the formula defining movements of the short-term risk-free rate. One characteristic is the kind of general
market behavior that is simulated. The other characteristic is whether the calibration is real-world or risk-neutral in
nature.

In what follows we first discuss a few different models of general market behavior. Then we discuss the difference
between real-world and risk-neutral versions of each model. This is done to emphasize that when choosing an ESG,
one needs to be aware of both characteristics of the model. All market-consistent models are not alike and do not
produce identical results, and the differences are not limited to whether they are real-world or risk-neutral; the
differences depend also on the underlying model of market behavior.

Perhaps the best-known and simplest interest rate model is the Vasicek model. The Vasicek model treats the short-
term risk-free interest rate as a mean-reverting random walk. Using the common mathematical form from above,
the Vasicek model in discrete time is expressed as follows:

4 The underlying theory does not support this usage in an asset-liability simulation where there is re-investment risk or where there are options in the
contracts that create decision points where the decision will be based on future economic conditions. In those situations, use of real-world market-
coherent scenarios is more consistent with the underlying theory.

Copyright © 2022 Society of Actuaries Research Institute

 12

(next period value) = (current value) + (expected drift) + (random shock)

𝑟𝑡+1 = 𝑟𝑡 + 𝜅 (𝜃 − 𝑟𝑡 ) + 𝜎𝑊𝑡

𝜅 = mean reversion strength or speed

𝜃 = mean reversion point

𝜎 = volatility

𝑊𝑡 = a random number from a Gaussian distribution with mean zero and variance one

To develop a closed-form expression for the spot price and the remainder of the yield curve, the discrete model
above is expressed as a stochastic differential equation in continuous time 5:

𝑑𝑟 = 𝜅(𝜃 − 𝑟)𝑑𝑡 + 𝜎𝑑𝑊

The Vasicek model assumes a certain kind of market behavior. Aspects of that behavior are that the volatility is
constant and that interest rates can become negative if enough negative random shocks accumulate.

Other models have been developed assuming different kinds of market behavior. For example, one can observe
that volatility is not constant but tends to be higher when interest rates are higher. One can also observe that there
is a strong tendency for interest rates not to go below zero. The Cox-Ingersoll-Ross model and the Black-Karasinski
model each embed those behaviors in different ways.

Here is the Cox-Ingersoll-Ross model:

𝑑𝑟 = 𝜅 (𝜃 − 𝑟)𝑑𝑡 + 𝜎√𝑟𝑑𝑊

The only difference from the Vasicek model is the random shock term which is multiplied by the square root of the
current rate. This makes the random shocks (i.e. the volatility) depend on the current level of the interest rate.
When interest rates approach zero, the random shocks get very small and mean reversion forces the interest rate
back up toward the mean reversion point. This prevents interest rates from ever getting to or below zero.

Here is the Black-Karasinski model:

𝑑 𝑙𝑛(𝑟) = 𝜅(ln (𝜃) − ln (𝑟))𝑑𝑡 + 𝜎𝑑𝑊

This model is also a mean-reverting random walk, but the simulation is done with the logarithm of the interest rate
rather than the interest rate itself. The volatility term is constant in absolute terms, but in proportion to the
logarithm of the interest rate the volatility is smaller when interest rates are lower and vice versa, as in the Cox-
Ingersoll-Ross model. But when the interest rate approaches zero, the magnitude of its logarithm approaches
infinity and that makes the mean reversion term very sensitive to the level of interest rates. In this model the mean
reversion term explodes when interest approach zero, providing an even stronger reversion to the mean than in the
Cox-Ingersoll-Ross model. And when interest rates are high, the mean reversion term gets substantially smaller and
has less effect.

5 Note that the values of the parameters to be used in the discrete time version of the model are not the same as those in the continuous time version.
The values depend on the size of the discrete time step in use.

Copyright © 2022 Society of Actuaries Research Institute

 13

These differences in assumed market behavior are illustrated in the chart below. The chart shows the distribution
of the short-term interest rate 15 years in the future based on each of the models, using comparable calibrations.
The distributions were simulated using 10,000 scenarios and counting the number of scenarios that fell into each
0.25% bucket. The distribution for the Vasicek model is a bell-shaped curve that includes some negative interest
rates. The distribution of the Cox-Ingersoll-Ross model is skewed with a longer tail on the high end and does not
include any negative interest rates. The distribution for the Black-Karasinski model is even more skewed.

Figure 4
SIMULATED INTEREST RATE DISTRIBUTIONS FROM THREE MODELS

All three of these distributions are based on calibrations with the same mean reversion point, the same value for the
mean reversion strength parameter, and the same volatility at the mean. The point of showing the difference
between them is to motivate the idea that the choice of model affects the valuation because it reflects the model’s
assumptions about market behavior and the likelihood of different events in the future. This dependence of any
market-consistent valuation on the choice of underlying model should not be overlooked.

3.2 THE MARKET PRICE OF RISK IN INTEREST RATE MODELS

The equations above do not explicitly include the market price of risk, which is normally denoted as 𝜆. Let’s briefly
review what the market price of risk is and then see how it has been incorporated into these models.

To define the market price of risk we start with the idea that risk-averse investors require a higher expected or
average return on riskier investments. We measure risk by the volatility of the market price. Greater volatility and
greater risk increase the probability that the price may decline causing a loss, but they are associated with higher
returns on the average. With that in mind we define the market price of risk as a ratio:

(𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑟𝑒𝑡𝑢𝑟𝑛)−𝑟

Market price of risk = 𝜆 = 𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑝𝑟𝑖𝑐𝑒

Copyright © 2022 Society of Actuaries Research Institute

 14

The excess average return over the risk-free rate is called a risk premium. The market price of risk is the ratio of the
risk premium to the volatility of the price.

One might ask why the market price of risk should be incorporated into a formula defining movements of the short-
term risk-free rate. After all, since the rate is risk-free, then the market price of risk does not apply. The explanation
is the formula given earlier for the spot price of a pure discount bond as an expectation based on the future
movement of the short-term rate. Pure discount bonds are not risk-free; their price is volatile because the market
interest rates used to discount their fixed future payoff are volatile. Since the price of a pure discount bond is
volatile, the market expects a return higher than the risk-free rate and uses that higher return to calculate the
market price. The path of the short-term rate used to price pure discount bonds (the risk-neutral path) must be
higher than the real-world expected path to a degree that depends on the market price of risk.

Therefore, it is common to introduce the market price of risk into these formulas as an upward drift in the risk-free
rate, an upward drift equal to the volatility times 𝜆. For example, the Vasicek formula including 𝜆 is this6:

𝑑𝑟 = 𝜅 (𝜃 − 𝑟)𝑑𝑡 + 𝜎(𝑑𝑊 + 𝜆)

Note that this is equivalent to simply increasing the mean reversion point parameter:

𝑑𝑟 = 𝜅 (𝜃 − 𝑟)𝑑𝑡 + 𝜎𝜆 + 𝜎𝑑𝑊

𝜎𝜆
𝑑𝑟 = 𝜅 ((𝜃 + ) − 𝑟) 𝑑𝑡 + 𝜎𝑑𝑊
𝜅

The last expression above is identical to the original expression for the Vasicek model, but with the mean reversion
𝜎𝜆
parameter 𝜃 replaced by (𝜃 + ). Because of that, sometimes the market price of risk in a model like this is re-
𝜅
defined to be the value added to the mean reversion parameter rather than the more abstract concept of risk
premium divided by the volatility.

The simple relationships shown above help one understand some of the language commonly used with reference to
these models.

The P-measure and the Q-measure. A measure refers to the probability distribution of future movements of the
variable under consideration, in this case 𝑑𝑟. As shown above, in an arbitrage-free model the spot price is defined
(measured) in terms of an expectation based on the path of 𝑟. Both the drift term and the random shock term are
part of the measure. As was shown above, the change in measure from real-world (the P-measure) to risk-neutral
(the Q-measure) in the Vasicek model is accomplished by modifying the random shock term, replacing 𝑑𝑊 with
𝑑𝑊̂ = 𝑑𝑊 + 𝜆. Equivalently, the change in measure can be accomplished by changing the drift term, replacing 𝜃
𝜎𝜆
with 𝜃̂ = 𝜃 + . This equivalence between changing the random term and the drift term is what makes this and
𝜅
many similar models mathematically tractable so that a fixed-form equation for the spot price at any maturity can
be derived. This equivalence is also what makes the P-measure and the Q-measure “equivalent Martingale
measures”. An appendix is included to provide further illustrations and explanations of the relationship between
the P-measure and the Q-measure.

6 In many references, lambda is introduced with a negative sign rather than a positive sign. When that is done, calibration shows that lambda has a
negative value. Here we introduce it with a positive sign and a positive value because the market price of risk is defined in a way that anticipates a positive
value.

Copyright © 2022 Society of Actuaries Research Institute

 15

The real-world parameters vs. the risk-neutral parameters. This expression refers to the values of the parameters,
not their definitions. The mathematical definition of the parameters is the same in real-world and risk-neutral
versions of the model. The parameters of the Vasicek model are 𝜅, 𝜃, and 𝜎. The difference between real-world
and risk-neutral parameter values is in the value of 𝜃. If the real-world value of that parameter is 𝜃 = 𝑥 then the
𝜎𝜆
risk-neutral value of that parameter is 𝜃̂ = 𝑥 + as shown in the formulas above7.
𝜅

Let’s turn now to the Cox-Ingersoll-Ross model. The market price of risk can be introduced into the Cox-Ingersoll-
Ross model in at least two different ways. The difference illustrates the complexity of reflecting the market price of
risk in other models.

̃𝑡 = 𝑑𝑊𝑡 − 𝜆√𝑟 𝑑𝑡. This

The first method is to replace the random term 𝑑𝑊 in the Cox-Ingersoll-Ross model with 𝑑𝑊 𝜎
is the approach used in the original Cox-Ingersoll-Ross paper8.

̃𝑡
𝑑𝑟𝑡 = 𝜅(θ − 𝑟𝑡 )𝑑𝑡 + 𝜎√𝑟𝑡 𝑑𝑊

𝜆 √𝑟
𝑑𝑟𝑡 = 𝜅(θ − 𝑟𝑡 )𝑑𝑡 + 𝜎√𝑟𝑡 (𝑑𝑊𝑡 − 𝑑𝑡)
𝜎

𝑑𝑟𝑡 = 𝜅 (θ − 𝑟𝑡 )𝑑𝑡 − 𝜆𝑟𝑡 𝑑𝑡 + 𝜎 √𝑟𝑡 𝑑𝑊𝑡

κθ
𝑑𝑟𝑡 = (𝜅 + 𝜆) ( − 𝑟𝑡 ) 𝑑𝑡 + 𝜎 √𝑟𝑡 𝑑𝑊𝑡
(𝜅 + 𝜆)

In the last line above we see that the terms containing lambda have been removed from the random term and put
equivalently into the drift term. The parameter 𝜅 has been replaced by 𝜅̂ = 𝜅 + 𝜆 and the parameter 𝜃 has been
𝜅𝜃
replaced by 𝜃̂ = . One can use this form of the model to derive a formulaic expression for the spot price. The
𝜅+𝜆
formulaic expression is very complex and is shown in Economic Scenario Generators – A Practical Guide (Pedersen
et. al, 2016).

Another method for introducing the market price of risk is to notice the following:

̃𝑡 = 𝜅̂(𝜃̂ − 𝑟𝑡 )𝑑𝑡 + 𝜎 √𝑟𝑡 𝑑𝑊𝑡

𝑑𝑟𝑡 = 𝜅 (θ − 𝑟𝑡 )𝑑𝑡 + 𝜎 √𝑟𝑡 𝑑𝑊

The second form of the model at the end of the line above can be used to develop a formulaic expression for the
spot price that depends on adjusted values of 𝜅 and 𝜃 without any reference to lambda.

If the adjustments are 𝛾1 and 𝛾2 then define 𝜅̂ = 𝜅 + 𝛾1 and 𝜃̂ = 𝜃 + 𝛾2 . The effect of the two adjustments
together is assumed to reflect the market price of risk. The value of this approach is that if one is only interested in
risk-neutral valuation, then one can use the formula for the spot price based on 𝜅̂ and 𝜃̂ to calibrate their values

7 One might be confused by the fact that lambda appears in this formula for the risk-neutral parameter value 𝜃̂. Isn’t the market price of risk zero in a risk-
neutral calibration? The explanation is that calibration can be done in a way that estimates the value of 𝜃̂ directly without the need to adjust for the
market price of risk. The market price of risk is not zero, rather it is reflected implicitly so no adjustment for the market price of risk is required. A risk-
neutral calibration does not really assume the market price of risk is zero; it could not reproduce market prices if it did. Rather, a risk-neutral calibration
provides parameter values that reflect the market price of investment risk implicitly, so no explicit adjustment for investment risk is required in a risk-
neutral valuation.
8 This is one of the references in which lambda has a negative value.

Copyright © 2022 Society of Actuaries Research Institute

 16

directly to market prices and never need to specify the implicit values for 𝛾1 and 𝛾2 or the real-world values 𝜅 and
𝜃.

The simple interest rate models discussed above are not commonly used in practice because their simplicity does
not reproduce actual market behavior very well. An appendix is included in this document to illustrate the
treatment of the market price of risk in some more complex interest rate models.

The market price of risk can seem obscure when presented as a factor in a mathematical formula. Understanding
can be improved by viewing the effect of the market price of risk on the yield curve. Recall that there is often a
closed-form formula for the yield curve based on a definition of the process for 𝑟𝑡 . The result of the formula
depends on the parameter values used. When there are both real-world and risk-neutral parameter values for the
same model, they will produce different yield curves starting from the same short-term rate. The difference
between those yield curves is caused by the market price of risk because the market price of risk is the difference in
the parameters. The difference between those yield curves is a set of “term premiums” that vary by term to
maturity.

Term premiums arise because of the risk associated with locking in a fixed rate for a long time in a world where
future interest rates are uncertain. Term premiums represent the market price of that risk. One can judge the
reasonableness of any formula for the market price of risk based on the term premiums that it implies, that is, based
on the difference between yield curves produced using real-world and risk-neutral parameter values.

Section 4: Equity index models

Arbitrage-free equity index models focus on the total rate of return of investments in the equity index. Here we
consider only models that include re-investment of dividends. Such models follow the general form outlined above:

(next period value) = (current value) + (expected drift) + (random shock)

Generally, these models are simpler than interest rate models. The expected drift for any period is typically the sum
of the short-term risk-free rate and a risk premium. The realized returns in any specific scenario are dominated by
the random shocks which tend to be much larger in magnitude than the expected drift.

In risk-neutral calibrations of these models, the risk premium is zero so the expected drift is always equal to the
short-term risk-free interest rate. Differences between risk-neutral versions of such models are limited to the
random shock term.

4.1 MARKET BEHAVIOR IN EQUITY INDEX MODELS

The design of an equity index model focuses on ways to model the expected drift and the random shock (volatility).
When multiple equity indices are included in one ESG, there is focus on the relationships of the drift and volatility of
various indices. We will not address relationships between separate equity indices in this document.

ALTERNATIVES FOR THE EXPECTED DRIFT TERM

There are no alternatives for behavior of the expected drift in a risk-neutral model. The expected drift in a risk-
neutral model is always the current short-term risk-free rate.

Alternatives for market behavior of the expected drift in a real-world model include:

Copyright © 2022 Society of Actuaries Research Institute

 17

• Whether the expected drift is the sum of the current short-term risk-free rate and a fixed risk premium or
is constant. If the expected drift is a constant, then the implied risk premium changes over time to offset
changes in the short-term risk-free rate.

• Whether the expected drift is regime-switching.

• Whether the expected drift is mean-reverting over time, so that the level of the equity index tends to
mean-revert to a level that grows at a more stable compounded rate over long periods.

ALTERNATIVES FOR THE RANDOM SHOCK TERM

Alternatives for market behavior of the random shock include:

• Whether the volatility of random shocks is constant or is stochastic or regime-switching in nature.

• Whether the volatility of random shocks is supplemented by a “jump process” that creates large jumps in
value at random intervals.

Many of these alternatives are discussed in Economic Scenario Generators – A Practical Guide (Pedersen et. al. 2016)
and are not elaborated here.

4.2 MARKET PRICE OF RISK IN EQUITY INDEX MODELS

The market price of risk in an equity index model is reflected in the excess of the drift term over the short-term risk-
free rate. That excess is the equity risk premium. In a risk-neutral model the drift term is equal to the short-term
risk-free rate so the market price of risk and the equity risk premium are zero.

Risk-neutral equity models are intended for valuation of options and other derivatives. One does not need a model
to perform valuation of a stock index – the value is the current level of the index. Since the drift term in a risk-
neutral equity model is always the short-term risk-free rate, the valuation of options depends only on the random
shock term in the model. Risk-neutral models can vary in the way the random shock term is constructed, and
calibration is important. The basic risk-neutral equity model is the Black-Scholes model in which volatility is Gaussian
and calibration amounts to determining the implied volatility level that corresponds to an observed market price.

Implied volatility in the Black-Scholes model is determined by calibration to the market price of an option. In the
Black-Scholes model the drift in price is the short-term risk-free rate and the only variable is the volatility. Implied
volatility is the volatility which, when used in the Black-Scholes formula, leads to an option value that equals the
market price of that option. When this is done, implied volatility tends to depend on the strike price and tenor of
the option used for calibration. The variation by strike price and tenor is called the “volatility surface”. The volatility
surface is very useful for valuation. But it presents a problem when the purpose of a simulation is not valuation of
an option at a point in time but focuses on the future path of the underlying index. In that context, a choice must be
made concerning the level of volatility for the underlying index in the scenarios. Reasonable choices come from the
range of volatilities in the volatility surface.9

9 The volatility surface corresponds to the Black-Scholes model, which is a constant volatility model. More complex models involve stochastic volatility,
and such models are often calibrated to value options using simulation. In that context, the parameters of a stochastic volatility model are calibrated to
market prices of some options, and the result may be several sets of parameters that differ by the strike and tenor of the option. An analogous issue arises
there because when the purpose of a simulation is not valuation of an option at a point in time but focuses on the future path of the underlying index, a
single set of parameters is needed to generate scenarios for the path of the index.

Copyright © 2022 Society of Actuaries Research Institute

 18

In a real-world model the equity risk premium tends to be positive. But some of the options for real-world models
involve the idea that equity returns are driven by forces other than a risk premium over the risk-free rate. Such
forces may include fiscal and monetary policy, GDP growth, unemployment, and others. Because of those other
influences, the equity risk premium may not be constant and the drift in equity prices may not be tied to the sum of
the risk-free rate and a fixed risk premium.

Section 5: Calibration of market-consistent and market-coherent models

5.1 GENERAL APPROACH

Economic Scenario Generators – A Practical Guide (Pedersen et. al. 2016) provides a good discussion of calibration
procedures in section 7. Importantly, it defines separate calibration approaches for risk-neutral and real-world
calibrations. Here we generalize so that the two calibration approaches can be viewed as two versions of the same
process. The generalized process involves iterative adjustment of parameter values to improve fit to historical
observations. The difference between risk-neutral and real-world calibration is not in the use of iterative fitting, it is
in selecting which historical observations are used as targets of the ESG’s fit, and how the goodness of fit is
measured.

No model yet invented can perfectly fit all historical observations. There will always be some historical data that a
model can fit better than other data. Historical data that is not used directly in calibrating the model may not be fit
well at all. As noted above, risk-neutral and real-world calibrations focus on fitting different aspects of historical
data. Because of this, it should not come as a surprise that sometimes risk-neutral and real-world calibrations are
very different from one another and appear inconsistent. The kind of conceptual connection between risk-neutral
and real-world calibrations that was presented earlier in this document can be reflected when both kinds of model
are calibrated together. But often they are calibrated independently using different samples of historical data. Odd
relationships between real-world and risk-neutral parameters can arise from independent calibration to separate
data sets.

In the sections below we briefly discuss risk-neutral and real-world calibration separately and then explain the kind
of combined calibration required for a real-world market-coherent model.

5.2 RISK-NEUTRAL CALIBRATION

The purpose of a risk-neutral ESG is valuation. Calibration of a risk-neutral ESG is the process of choosing
parameters that reproduce market prices, so that the ESG can be used to determine a market-consistent value for
items that are not widely traded and therefore have no observable market price.

Typically a risk-neutral ESG is formulated in a way that leads to closed-form expressions for the prices of various
kinds of items traded in the market, including both direct investments and derivatives. The process of calibration
amounts to iteratively adjusting the values of the parameters in the closed-form expressions for market prices to
achieve the best fit to the widest array of observed market prices.

These considerations involved in risk-neutral calibration deserve special attention:

• Market conditions are constantly changing. Often (but not always) a risk-neutral ESG is re-calibrated on
every valuation date to current prices in order to be most consistent with current market prices.
Sometimes recalibration is less frequent, and the calibration is intended to be consistent with market prices
over a period of time, but the period of time is selected to represent relatively stable market conditions.

Copyright © 2022 Society of Actuaries Research Institute

 19

The point here is that the time frame of the historical data sample used for risk-neutral calibration is
typically shorter than that used for real-world calibration.

• Once a formula for a market price based on a set of parameters is developed, the process of fitting
parameter values to reproduce market prices often does not associate any meaning to those parameters.
They are just numbers to be adjusted in a mathematical optimization. The resulting parameter values may
not make much sense in the real-world. This is not viewed as a problem in the context of valuation. But
care must be exercised when using such risk-neutral parameter values as a starting point when building a
real-world model.

• Sometimes the calibration process includes solving for different parameter values for different derivatives
of the same investment. This is most common in connection with the volatility parameter. For example, in
the Black-Scholes model the “volatility surface” is an array of different values for the volatility parameter
depending on the strike prices and tenors of options on some underlying index. By using an array of values
like this, market prices can be matched more exactly. But the user should understand that there is only
one underlying item with only one underlying volatility, and that the need for an array of different
volatilities for derivatives is evidence that the model does not fit the underlying item and its derivatives
with the same parameters. The use of a “volatility surface” is just a way to improve the fit of the model to
current prices. It is useful in the context of valuation, but not in the context of simulation because there
can be only one volatility for the underlying item. In a simulation the volatility of the underlying item
governs payoffs of any derivatives of that item.

5.3 REAL-WORLD CALIBRATION

Calibration of real-world models is covered well in Economic Scenario Generators – A Practical Guide (Pedersen et.
al. 2016).

There are some key differences from risk-neutral calibration.

• The “goodness of fit” is measured differently. Instead of measuring the fit of formulaic prices to actual
market prices, “goodness of fit” is measured with reference to expected movements over time in the
generated scenarios. The Practical Guide says:

o “Because real-world parameterizations are forward looking, they require explicit views as to how the
economy will develop in the future, so they require a significant amount of expert judgment to determine the
veracity of the scenarios that result from the parameterization process. In practice, real-world calibrations
often are parameterized to be consistent with historical dynamics of economic variables, although the long-
term steady-state levels associated with these parameterizations can differ from long-term historical
averages in favor of current consensus expectations or even individual viewpoints.

• Because goodness of fit is measured in terms of movements over time, parameters like the mean reversion
point and volatility have real meaning and the steady-state levels mentioned in the Practical Guide have
real meaning. This contrasts with risk-neutral calibration where, as noted above, sometimes the parameter
values diverge from their real meaning.

• The market price of risk is always non-zero in a real-world model and is always zero in a risk-neutral model.
That means that the expected return of any investment that involves risk will be different on average when
comparing real-world model and risk-neutral models. That makes real-world and risk-neutral calibrations
of the same model different from one another.

Copyright © 2022 Society of Actuaries Research Institute

 20

5.4 COMBINED CALIBRATION: A MARKET-COHERENT REAL-WORLD MODEL

A real-world model can be calibrated to reproduce market prices, but this is achieved differently in the real-world
context than in the risk-neutral context.

In both cases, reproducing market prices means

• Choosing a probability measure to use in calibration. The real-world measure is called the P-measure. The
risk-neutral measure is called the Q-measure.

Calibrating the path of future interest rates using observed spot prices of pure discount bonds.

• Reflect market volatility around that path through calibration to some observed data.

Real-world and risk-neutral calibrations accomplish these things differently.

• Choosing a probability measure

o Risk-neutral calibration uses the Q-measure

o Real-world calibration uses the P-measure

• Reflecting expectations for the path of interest rates

o Risk-neutral calibration centers the path of the short-term risk-free rate on the forward rate path
implied by observed spot prices on pure discount bonds.

o Real-world calibration centers the path of the short-term risk-free rate on the forward rate path
implied by observed spot prices on pure discount bonds, adjusted downward to remove the term
premiums included in observed long term bond rates for the risk of locking in a long-term rate.

• Reflecting market volatility

o Risk-neutral calibration solves for the “implied volatility”, the volatility parameter value required in
a closed-form formula for the price of a security to reproduce the market price. This often results
in a “volatility surface” of different implied volatilities for different options on the same underlying
security.

o Real-world calibration uses historical realized volatility in observed market prices for the
underlying security.

Before discussing further details of the real-world calibration process, it is important to emphasize that real-world
market-coherent scenarios are designed to be used differently than risk-neutral market-consistent scenarios in any
stochastic valuation exercise. When using risk-neutral scenarios, the valuation result is the average of the scenario-
specific valuations. When using real-world scenarios, the average of the scenario-specific values needs to be
adjusted by adding a margin for investment risk, either at the scenario level or in total.

There is some debate over whether the margin for risk can be determined in a reliable way. If not, then the real-
world valuation cannot reproduce market prices even if it is based on scenarios calibrated in the real-world market-

Copyright © 2022 Society of Actuaries Research Institute

 21

coherent manner described above. In this document we take the view that risk margins can be determined in a
reliable way10. Techniques for calculating risk margins are outside the scope of this document.

We will first discuss characteristics of a real-world model that allows these things to be done, and then provide
more detail on how they are done.

Characteristics of a real-world market-coherent model include:

• It must start with an arbitrage-free model framework because a model must be arbitrage-free to be
market-coherent.

• An arbitrage-free model framework can have both real-world and risk-neutral calibrations. A market-
coherent real-world model must have both real-world and risk-neutral calibrations because both are used
in scenario generation.

• The difference between the parameter values in real-world and risk-neutral calibrations represents a
specific assumption about the market price of risk.

• The model must have a clear division between “state variables” and “parameters”. There are two sets of
parameters (real-world and risk-neutral) but only one set of “state variables”.

The statement that such a model must have both real-world and risk-neutral calibrations deserves explanation.
Both calibrations are required because they are both used when generating interest rate scenarios.

Interest rate scenarios in a market-coherent real-world model are generated using the following general process:

1. The initial values of the state variables are fit to the observed starting yield curve (spot prices) using the
risk-neutral parameters.

2. Future scenario paths for the short-term rate are generated using the real-world parameters to create
paths for the state variables.

3. The yield curve at any time step in a scenario path is calculated using the state variables and the risk-
neutral parameters. This process assumes we have a model that has a closed-form formula or other fast
procedure to calculate the yield curve based on the state variables and the parameters.

This is the same as the process used by a risk-neutral model, except for step 2 where the real-world parameters are
used in place of the risk-neutral parameters for simulating the forward path of the state variables. Steps 1 and 3 are
the same as in a risk-neutral model because those steps deal with fitting the state variables to a full yield curve at a
point in time. Step 2 is different because it deals with generating a path through time based on the probabilities of
different paths. In step 2 a real-world generator uses the real probabilities (P-measure) while a risk-neutral
generator uses the risk-neutral probabilities (Q-measure) which implicitly give more weight to adverse events that
lead to lower market prices.

10 Valuation of insurance contracts for accounting purposes requires inclusion of margins for both investment risks and insurance risks. Risk-neutral
scenarios are often used to provide an implicit market-consistent margin for investment risks. But the margin for insurance risks must then be added
separately. If it is possible to determine a separate margin for insurance risks, one may argue that it is also possible to determine a separate margin for
investment risks, or for the sum of all risks in a block of insurance contracts.

Copyright © 2022 Society of Actuaries Research Institute

 22

To say this another way, the future real-world path of the short-term rate is governed by the real-world parameter
values, while the forward rates that define the yield curve (and spot prices) are governed by the risk-neutral
parameter values.

Equity return scenarios in a market-coherent real-world model are generated using the following general process:

1. Generate the risk-free rate for the next time step

2. Add a risk premium (expected equity return spread) based on the market price of risk

3. Add a random shock based on the volatility process

This differs from the process in a risk-neutral equity return generator because the risk premium in step 2 is zero in a
risk-neutral generator. Also, the volatility in step 3 may differ.

In the modeling of interest rates, flexibility in simulating market behavior generally means using a multi-factor
model for interest rates. Each factor corresponds to a “state variable” whose value represents the current state of
that factor. In a one-factor interest rate model, the state variable is typically the current short-term risk-free
interest rate. Multi-factor models are discussed here in the Appendix on “More complex interest rate models”.

In a multi-factor model the values of the state variables are shared between real-world and risk-neutral versions of
the model, while the values of the parameters differ. The state variables define conditions on the starting date and
any simulated date thereafter, while the parameters govern the stochastic paths of the state variables. The
parameters remain fixed over time, but the state variables move over time.

It is this division between “state variables” and parameters that enables the creation of a market-coherent real-
world model. The state variables are shared between real-world and risk-neutral versions of the model. Risk-
neutral parameters are calibrated for consistency with market prices, including the spot prices in the yield curve.
Real-world parameters are calibrated for consistency with movements of the state variables across points in time.
In a real-world market-coherent model, the prices (yield curve) at any point in time are based on the risk-neutral
parameters while the movements of the state variables are governed by the real-world parameters, as is most
consistent with the way each set of parameters is calibrated.

The use of both real-world and risk-neutral parameters in the same model imposes a calibration constraint that
does not exist when such parameters are used in separate models. The constraint arises because the difference
between the real-world and risk-neutral parameters is the market price of risk. For the two sets of parameters to
make sense together, the market price of risk that connects them must be reasonable. That means, at a minimum,
that it should generally be positive so that the generated scenarios tend to show higher expected average returns
for riskier investments.

Calibration of risk-neutral parameters is typically done by measuring “goodness of fit” to market prices. Calibration
of real-world parameters is typically done by measuring “goodness of fit” to movements in state variables. The
connection in theory is the market price of risk, but when the two calibrations are done independently, the market
price of risk implied by the difference between them may not make sense. That’s because, as noted earlier, models
do not fit perfectly, and the approach used in risk-neutral calibration can lead to parameter values that lose their
real meaning.

Therefore, the calibration of the risk-neutral parameters in a real-world market-coherent model may be done
differently than for other risk-neutral models. The real-world parameters are calibrated first and taken as given.
The risk-neutral parameters are treated as the sum of the real-world parameters and adjustments that reflect the
market price of risk, and those adjustments are the subject of the risk-neutral calibration.

Copyright © 2022 Society of Actuaries Research Institute

 23

One way to calibrate the market price of risk in an interest rate model is to set it so that it produces reasonable term
premiums. As noted earlier, term premiums implied in the yield curve are defined by the difference in two yield
curves. One curve is generated using the real-world parameters and the other is generated using the risk-neutral
parameters which are the real-world parameters adjusted for the market price of risk. The difference between
those curves is a set of term premiums by term to maturity.

The two yield curves that are compared to get the term premiums are both based on the same values of the initial
state variables. Since state variables move over time, one should determine the term premiums implied starting
from a range of different values for the state variables.

The process just described for calibrating the market price of risk involves a fair amount of judgment along with
knowledge of what a reasonable set of term premiums should be. Generally they should be positive if one believes
that markets are risk-averse, but historical studies have provided more refined estimates. An appendix to this
document discusses term premiums and the market price of risk in more detail.

Section 6: Acknowledgments
The researchers’ deepest gratitude goes to those without whose efforts this project could not have come to fruition:
the Project Oversight Group for their diligent work overseeing, reviewing and editing this report for accuracy and
relevance.

Project Oversight Group members:

Charles V Ford, FSA, MAAA, CFA

Gary Hatfield, FSA, CERA, CFA, PhD

Jason Kehrberg, FSA, MAAA, PhD

James Kosinski, FSA, MAAA, CFA, PhD

Steve Marco, ASA, CERA

Hal Pedersen, ASA, MAAA, PhD

At the Society of Actuaries Research Institute:

Doug Chandler, FSA, FCIA, Canadian retirement research actuary

Dale Hall, FSA, CERA, CFA, MAAA, SOA managing director of research

Copyright © 2022 Society of Actuaries Research Institute

 24

Appendix A: The P-measure and the Q-measure

This appendix provides graphic depictions of the future distributions of returns created by scenario generators,
focusing on the difference between real-world and risk-neutral versions. The real-world distributions are called the
P-measure and the risk-neutral distributions are called the Q-measure.

INTEREST RATES
The chart below illustrates the drift of the future short-term risk-free rate in one calibration of a Cox-Ingersoll-Ross
model. Both the real-world (RW) and risk-neutral (RN) drifts are shown. Since the drift is stochastic in nature, three
percentile points in the probability distribution are shown: the 10th percentile, the 50th percentile and the 90th
percentile. In the scenarios charted here, the starting level of the short-term risk-free rate is 4.00% and the
calibrated real-world mean reversion point is also 4.00%.

The point of this chart is to show that the risk-neutral drift is higher than the real-world drift.

The next chart shows the probability density of the short-term risk-free rate based on the same scenarios as the
previous chart. The density is charted based on the number of scenarios in each 0.25% interval out of a set of
10,000 scenarios.

The point of this chart is that the risk-neutral probabilities (Q-measure) define a distribution that has a higher mean
and is slightly wider than the real-world probabilities (P-measure).

Copyright © 2022 Society of Actuaries Research Institute

 25

Risk-neutral scenarios have been described as weighting adverse events more heavily and thereby leading to lower
future values of current investments. Consider how that is reflected in a simulation of a pure discount bond that
matures for $1,000. We know the starting value, and we know the maturity value, and those must be the same in
both real-world and risk-neutral scenarios. But what about the value between those two points?

Consider a situation where the current risk-free short-term rate is at its mean reversion point at the start of a
simulation. In the real world the short-term rate is equally likely to move up or down, so let’s look at a scenario
where it remains constant, and the rest of the yield curve remains constant. The chart below shows the yield curve
we’ll use for illustrative purposes. The one-year spot rate is 4% and the 10-year spot rate is 6%. The 10-year rate
includes term premiums that compensate for the risk of locking in the rate for 10 years.

Copyright © 2022 Society of Actuaries Research Institute

 26

If we generate both real-world and risk-neutral scenarios with this starting yield curve, we know that the scenarios
will differ, with the risk-neutral scenario drifting higher. While the short-term rate in the real-world scenario will be
unchanging, the short-term rate in the risk-neutral scenario will follow the forward rate path in the initial yield
curve. This will lead to higher interest rates in the risk-neutral scenario. That means lower discounted present
values for the bond in the risk-neutral scenario. The chart below illustrates the average path of the value of the
bond. At every point between the start date and the maturity date, the value is lower in the risk-neutral scenario.

One can calculate the return on the bond for each projection year in each of these scenarios. In the real-world
scenario the return “walks down the curve” in the usual way, showing a return equal to long-term forward rates
when the maturity date is far in the future, and declining over time to the short-term rate just before maturity. In
the risk-neutral scenario, all investments are assumed to earn the same short-term risk-free rate (on average), but
the short-term risk-free rate follows the path of the forward rate curve. The forward rate curve starts with the
short-term rate and ends with the long-term rate. Basically, the forward rate curve is earned in reverse order.

Copyright © 2022 Society of Actuaries Research Institute

 27

EQUITY RETURNS
For equities, the real-world probabilities define a distribution of future values that is higher than the distribution
based on risk-neutral probabilities. The higher values include the market reward for taking risk. The chart below is
based on simulation of 10,000 scenarios, counting the number of accumulated values in each $0.05 interval.

When percentile points on these distributions are tracked over time one can see how the difference in probabilities
affects the drift of accumulated value over time.

Copyright © 2022 Society of Actuaries Research Institute

 28

Appendix B: Some more complex interest rate models

The interest rate models discussed in the body of this document are not used widely in practice for simulation. They
were used to explain concepts in a simple context.

The reason those simple models are not used in practice is that the market behavior that they capture is
unrealistically simple. They are one-factor models where the only factor describing the current state of the
economy is the short-term risk-free interest rate. Since those models are mean reverting with a fixed mean
reversion point, the market behavior they imply is:

• if the short-term rate is above the mean reversion point then then forward rates will trend down

• if the short-term rate is below the mean reversion point then forward rates will trend up

The only shape of yield curve such models can produce is monotonic, either upward sloping or downward sloping,
depending on where the current short-term rate is relative to the mean reversion point.

Interest rate models for simulation in practice are multi-factor in nature. Such models can simulate behavior where
interest rates will trend in one direction in the short run and in another direction in the long run. That makes it
possible to simulate behavior consistent with the shape of market yield curves that are not monotonic; hump-
shaped and valley-shaped yield curves are possible.

The role of “state variables” is a key to such models. Each “factor” in a one-factor or multi-factor model has a state
variable that defines the current state of that factor. In a one-factor model, the state variable is typically the current
level of the short-term risk-free interest rate. In multi-factor models, the additional factors have state variables that
correspond to the current level of other variables that are included in the model.

An important aspect of multi-factor models is that each factor has both a state variable that defines its current value
and parameters whose values define future stochastic paths of that state variable. The general idea is that the
parameter values are anticipated to be stable over time and the movements of the state variables define the
behavior of the yield curve over time. Fitting a current yield curve to such a model means selecting values for the
state variables that best fit the model yield curve to the observed yield curve.

Here we describe two very different approaches to multi-factor models. In a “double mean reverting” model, each
factor governs a different aspect of market behavior. In a “three-factor Cox-Ingersoll-Ross” model, each factor is a
separate CIR model of the short-term risk-free rate that has its own parameter values, and the simulated short-term
rate is the sum of the three factors.

DOUBLE MEAN REVERTING MODELS

A double mean reverting model has two state variables: the short-term rate and a short-term mean. The short-term
rate is attracted to the short-term mean in a random walk, and the short-term mean itself follows a random walk
and is attracted to a long-term mean which remains fixed.

Definitions of the state variables:

𝑟𝑡 = current short-term risk-free rate at time t

𝑚𝑡 = current short-term mean reversion point for the risk-free rate at time t

Copyright © 2022 Society of Actuaries Research Institute

 29

Definitions of the parameters:

𝜇 = long-term mean reversion point

𝜎1 = volatility of the short-term rate

𝜎2 = volatility of the short-term mean reversion point

𝜅1 = mean reversion strength of the short-term rate to the short-term mean

𝜅2 = mean reversion strength of the short-term mean to the long-term mean

Definition of the stochastic process in discrete time:

𝑟𝑡+1 = 𝑟𝑡 + 𝜅1 (𝑚𝑡 − 𝑟𝑡 ) + 𝜎1 𝑊2

𝑚𝑡+1 = 𝑚𝑡 + 𝜅2 (𝜇 − 𝑚𝑡 ) + 𝜎2 𝑊2

The random shocks 𝑊1 and 𝑊2 are random draws from a Gaussian (Normal) distribution with mean 0 and variance
1.

There are several variations on this model.

• The two random shocks may or may not be correlated.

• The two stochastic processes (two factors) shown above are in a form analogous to the Vasicek model.
Each could be changed to be analogous to a Cox-Ingersoll-Ross, Black-Karasinski, or other formulation, with
suitable changes to values of the parameters.

Characteristics of this model type that deserve note:

• A variety of yield curve shapes can be generated using a fixed set of parameter values while changing the
values of the state variables. For example, a humped yield curve is generated when the short-term mean is
greater than either the short-term rate or the long-term mean.

• In some scenarios the short-term mean will remain far from the long-term mean for an extended period of
time. Since the short-term rate reverts to the short-term mean, this tends to create scenarios that may be
characterized as “low-for-long” or “high-for-long” in greater proportion than with a model with a single
fixed mean reversion point.

Reflecting the market price of risk in a double-mean-reverting model

The market price of risk is reflected by the difference in drift due to the difference between real-world and risk-
neutral parameter values. The simplest approach is to simply increase the long-term mean reversion point so that
the risk-neutral value of the parameter 𝜇 is 𝜇̂ = 𝜇 + 𝜆𝜇 . Alternately, one can use the approach discussed earlier
where one starts with the formulaic yield curve based on real-world parameters and add a vector of term premiums
to reflect the market price of risk. There are, of course, other approaches as well.

Copyright © 2022 Society of Actuaries Research Institute

 30

THE THREE-FACTOR COX-INGERSOLL-ROSS MODEL

A three-factor Cox-Ingersoll-Ross model is simply the sum of three independent one-factor Cox-Ingersoll-Ross
models. The three factors each have their own parameters for mean, volatility, and mean reversion strength. The
short-term rate is the sum of the state variables of the three factors.

Definitions of the state variables:

𝑟1,𝑡 = current short-term risk-free rate in factor 1

𝑟2,𝑡 = current short-term risk-free rate in factor 2

𝑟3,𝑡 = current short-term risk-free rate in factor 3

Definitions of the parameters:

Each of the three factors has its own set of the three parameters in a Cox-Ingersoll-Ross model:

𝜅 = mean reversion strength or speed

𝜃 = mean reversion point

𝜎 = volatility

The values of these parameters are different between the factors, so there are nine parameters in total before
reflecting the market price of risk.

Definition of the stochastic process in discrete time:

Each of the factors follows the basic Cox-Ingersoll-Ross process using its own state variable and its own set of
parameter values. Here is the process used for each factor expressed in continuous time:

𝑑𝑟 = 𝜅 (𝜃 − 𝑟)𝑑𝑡 + 𝜎√𝑟𝑑𝑊

Since the short-term rate is the sum of the state variables, the price of a pure discount bond is

𝑃𝑇 = 𝐸 [𝑒𝑥𝑝 (− ∫(𝑟1,𝑡 + 𝑟2,𝑡 + 𝑟3,𝑡 )𝑑𝑡)]

0

The mathematical form of the Cox-Ingersoll-Ross model allows derivation of a closed-form expression for this price.
That is a significant advantage of this type of model.

Characteristics of this model type that deserve note:

• The three factors in this model are loosely analogous to the three components of the yield curve that arise
from principal component analysis. A calibration of this model generally results in the three factors having
very different parameter values for the mean reversion strength or speed. The parameter values are low,
medium, and high. These correspond to the principal components as follows:

o Low mean reversion speed corresponds to the “level” component

o Medium mean reversion speed corresponds to the “slope” component

Copyright © 2022 Society of Actuaries Research Institute

 31

o High mean reversion speed corresponds to the “curvature” component

• The three factors each have a fixed mean reversion point, so in sum they have a fixed mean reversion
point. This makes this kind of model less likely to produce scenarios that remain far from the mean for an
extended period of time as in “low for long” or “high for long”.

• The movements of the state variables in this model are assumed to be independent, and scenarios are
generated using independent random shocks to each factor. However, when the state variables are fit to
historical yield curves, the historical movements of the state variables are significantly correlated. This
indicates that while this model can generate reasonable distributions of future interest rates, the path-wise
behavior of scenarios may be less realistic than some other models.

Reflecting the market price of risk in a three-factor Cox-Ingersoll-Ross model

As was discussed earlier for the one-factor Cox-Ingersoll-Ross model, it is common to reflect the market price of risk
by simply defining adjustments to the parameters 𝜃 and 𝜅 in each of the factors. That makes six different
adjustments (2 per factor x 3 factors). Combined with the 9 parameters already in the model, that makes 15 total
parameters in a combined real-world / risk-neutral model11.

As with any mean-reverting model, adding a market price of risk that produces positive term premiums in the
generated yield curves is most directly done by an adjustment (increase) in the mean reversion point. In this model
each factor has its own mean reversion point and its own mean reversion speed. One can manipulate the shape of
the curve of term premiums by deciding how much of the adjustment to the total mean reversion point is allocated
to each of the three factors. Putting most of the adjustment in the faster-reverting factors will lead to a curve of
term premiums that is steeper on the short end and levels off at the long end, which some actuaries believe is most
realistic.

11 That’s before any shift that may be added to allow the model to simulate negative interest rates. The shift would be a 16 th parameter.

Copyright © 2022 Society of Actuaries Research Institute

 32

Appendix C: The meaning of arbitrage-free

Arbitrage is the generation of profit with no risk. It occurs when the same item can be purchased somewhere at one
price and sold elsewhere at a higher price. Markets are assumed to be arbitrage-free, so any market consistent or
market-coherent ESG must be arbitrage-free.

In a world where countless kinds of derivative securities exist, the detection of arbitrage in any scenario generator is
a complex task. While arbitrage is defined as buying and selling the same “item” at different prices, the “item”
could be composed of two or more other items, often including a derivative of the item itself. A test for arbitrage
must evaluate all equivalent combinations of items. Complicated mathematical tests have been devised for that
purpose. Section 9 of Economic Scenario Generators – A Practical Guide (Pedersen et. al. 2016) describes such tests.

A simple concept underlies those complicated tests. The concept is that the generated paths of future returns (i.e.,
interest rates, stock prices) must be consistent with market expectations on the starting date of the scenarios. One
of those expectations is the path of the short-term risk-free rate. To be arbitrage-free, an ESG for investments that
involve risk must produce some scenario returns higher and some lower than the short-term risk-free rate. That
way there are some scenarios that result in gains and some that result in losses, relative to keeping money risk-free
in a money market fund. If all scenarios result in gains, or all result in losses, then arbitrage exists, and the generator
is not arbitrage-free. To be arbitrage-free, any possibility of gain (earnings higher than the risk-free rate) must be
accompanied by a possibility of loss.

This does not mean that the average return must equal the risk-free rate. The average scenario return for a class of
investments can be much different from the risk-free rate while still allowing the possibility of both gains and losses
and thereby remaining arbitrage-free. This is one of the keys to understanding the connection between real-world
and risk-neutral ESGs. In real-world ESGs the average scenario return for an investment depends on the level of risk
in the investment. In risk-neutral ESGs the average scenario return for all investments is the same short-term risk-
free rate, but that risk-free rate follows an unrealistic average path. Both kinds of ESGs can be arbitrage-free; the
difference is that the market price of risk is nonzero in a real-world generator and that leads to average returns on
risky investments that differ from the risk-free rate.

Copyright © 2022 Society of Actuaries Research Institute

 33

Appendix D: Term premiums and the market price of risk

Term premiums are the reward investors demand for taking the risk of locking in a fixed interest rate in a world
where interest rates change over time. The risk is that interest rates may rise, so a locked-in rate has both an
element of opportunity cost and an element of risk that the market value of a fixed-rate investment may decline.
Those two risks are flip sides of the same thing – they are 100% correlated.

Since term premiums are a reward for taking risk, they are a function of the market price of risk. If you can quantify
the market price of risk, you should be able to calculate the corresponding term premiums. This appendix explains
how to do that.

The market price of risk has been defined as this ratio, where r is the short-term risk-free rate:

(𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑟𝑒𝑡𝑢𝑟𝑛)−𝑟

Market price of risk = 𝜆 = 𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑝𝑟𝑖𝑐𝑒

For default-free fixed-income investments, the numerator of the ratio is the term premium. The denominator is the
volatility of the price. Recall that in an arbitrage-free model, if the short-term risk-free rate at time t is 𝑟𝑡 then the
spot price P for maturity t=T is the expectation:

𝑃𝑇 = 𝐸 [𝑒𝑥𝑝 (− ∫ 𝑟𝑡 𝑑𝑡)]
0

The price is a function of the risk-free rate at time 0, which is 𝑟0 . One should be able to estimate the sensitivity of
𝜕𝑃𝑇
the price to a change in 𝑟0 , which would be noted as . We also can estimate the volatility of 𝑟0 which can be
𝜕𝑟0
notated as 𝜎𝑟 . We can then quantify term premiums using the following algebra.

(𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑟𝑒𝑡𝑢𝑟𝑛) − 𝑟

𝜆=
𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑝𝑟𝑖𝑐𝑒

(𝑇𝑒𝑟𝑚 𝑝𝑟𝑒𝑚𝑖𝑢𝑚)
𝜆=
𝜕𝑃
𝜎𝑟 ( 𝑇 )
𝜕𝑟0

𝜕𝑃𝑇
𝜆𝜎𝑟 ( ) = (𝑇𝑒𝑟𝑚 𝑝𝑟𝑒𝑚𝑖𝑢𝑚)
𝜕𝑟0

This gives an expression for the term premium given three things:

• An estimate of the market price of risk 𝜆

• The volatility of the short-term interest rate 𝜎𝑟

𝜕𝑃𝑇
• The sensitivity of the price to the short-term rate 𝜕𝑟0

Copyright © 2022 Society of Actuaries Research Institute

 34

Consider the situation where the values of 𝜆 and 𝜎𝑟 are constants12. The sensitivity of the price is a function of the
term to maturity and is greater for longer maturities. This means that term premiums can be expected to increase
𝜕𝑃𝑇
with term to maturity and follow a curve that mirrors the shape of which varies by term to maturity.
𝜕𝑟0

The Cox-Ingersoll-Ross model is one such model and it has a formulaic expression for the spot prices 𝑃𝑇 . That makes
𝜕𝑃𝑇
it possible to calculate the value of and quantify the term premiums resulting from a set of chosen parameter
𝜕𝑟0
values. The chart below illustrates term premiums from the Cox-Ingersoll-Ross model based on one reasonable
calibration.

Illustrative term premiums

from a Cox-Ingersoll-Ross model
𝜆𝜎 (𝜕𝑃/𝜕𝑟)
2.00%

1.50%

1.00%

0.50%

0.00%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Years to maturity

Similar curves arise from other mean-reverting models.

When a curve of term premiums is quantified in this way, one needs to be clear about the measure of the observed
yield curve to which term premiums correspond. When term premiums are quantified in this way, they are part of
the forward rate curve. If they were quantified as part of the spot curve or the par coupon curve, they would
typically be lower and less steep on the short end 13.

The connection between real-world and risk-neutral interest rate models can be seen clearly in this context.
Consider the formula for the price of a pure discount bond:

𝑃𝑇 = 𝐸 [𝑒𝑥𝑝 (− ∫ 𝑟𝑡 𝑑𝑡)]
0

12 There are of course models where 𝜆 and 𝜎𝑟 are not constants. These include multi-factor models or models with stochastic volatility. The mathematics
associated with such models gets very complex and is beyond the scope of this document. Nevertheless, the same conceptual framework applies. One
can gain much intuitive insight from the simple analysis presented here.
13 The reader is assumed to know how to transform a yield curve into its three equivalent forms – the par coupon curve, the spot curve, and the forward
rate curve. One source is Fabozzi 2021, pages 692-695, 709-713.

Copyright © 2022 Society of Actuaries Research Institute

 35

The price is defined based on the path of 𝑟𝑡 , the instantaneous forward discount rate. We can break the path of 𝑟𝑡
into two parts: the term premium 𝑝𝑡 and the remainder 𝑠𝑡 . The remainder conceptually corresponds to the path of
the risk-free short-term rate, which does not include term premiums. We then have:

𝑃𝑇 = 𝐸 [𝑒𝑥𝑝 (− ∫(𝑠𝑡 + 𝑝𝑡 )𝑑𝑡)]

0

The risk-neutral parameters for an interest rate scenario generator define the stochastic path for 𝑟𝑡 . The real-world
parameters for an interest rate scenario generator define the stochastic path for 𝑠𝑡 , which is the path of 𝑟𝑡 with
term premiums removed. The difference between those paths is the path of term premiums 𝑝𝑡 .

While term premiums are not directly observable at a point in time, cross-sectional studies over time can provide
estimates of term premiums. Such studies have been done, and they suggest that term premiums follow a pattern
by maturity that is much like that shown in the chart above, rising quickly for short maturities and leveling off for
longer maturities.

NOTES ON CALIBRATION OF TERM PREMIUMS

The chart shown above is based on one reasonable calibration of a model. Term premiums are sensitive to the
calibration. The chart below illustrates term premiums based on some alternate parameter values for volatility and
mean reversion speed, all using the same market price of risk within the same model.

In more complex interest rate models, term premiums can also be sensitive to the level and shape of the starting
yield curve.

One may wish to adapt an ESG so that it can reflect a desired set of term premiums. The exhibit below outlines the
steps used to generate scenarios in a market-coherent real-world ESG. The basic process is shown on the left, and
the changes needed to reflect pre-specified term premiums are shown on the right.

Copyright © 2022 Society of Actuaries Research Institute

 36

Step Basic process Revised process

(specified term premiums)

Initialize state variables Fit state variables to the spot prices in Subtract the specified term premiums
the observed starting curve using the from the observed starting curve and fit
risk-neutral parameters. the state variables to the resulting curve
using real-world parameters.

Generate paths of state variables Use the real-world parameters to Same.

generate the stochastic paths.

Calculate yield curves at each time step Use the risk-neutral parameters to Use the real-world parameters to
based on the state variables at that calculate the yield curve based on the calculate a yield curve based on the
time state variables. state variables, and then add the
specified term premiums to that curve.

Note that there is no use of risk-neutral parameters in the revised process. The use of pre-defined term premiums
replaces the role of the risk-neutral parameters. This emphasizes the fact that the role of risk-neutral parameters in
a real-world market-coherent ESG is simply to define the term premiums. If term premiums are provided explicitly
in another way, then there is no role for risk-neutral parameters in generation of real-world market-coherent
scenarios.

The model under discussion here assumes that the yield curve is governed by two forces – the future expected path
of the short-term rate and the term premiums based on the market price of risk. The observed yield curve can also
be influenced by other things such as supply and demand factors and inflation expectations. These issues need to
be considered when evaluating the fit of a model to historical data.

Copyright © 2022 Society of Actuaries Research Institute

 37

Appendix E: Fair value, Principle-based Reserves, and the Valuation Manual

The valuation of insurance reserve liabilities (“reserves”) for accounting purposes has evolved in the 21st century so
that it is often based on stochastic simulation and discounting of future cash flows. In the U.S. such valuation
methods have been called “principle-based” (Center for Insurance Policy Research 2021) and are encoded in the
National Association of Insurance Commissioner’s (NAIC) Valuation Manual for use in required regulatory financial
reporting. Internationally, a different and new standard of accounting for insurance contracts has also employed
such methods (IFRS 2021), but with important differences.

There is a common principle on which all such methods are based. The principle is that the value of a block of
insurance contract reserves is equal to the value of the assets required on the valuation date to pay them off, with
some degree of uncertainty or margin for risk. This principle ties the valuation of the insurance contract reserves to
the valuation of the assets. The largest difference between reserve valuations in the international accounting
standard and the U.S. statutory Valuation Manual is that the valuation of assets is at market in the international
standard while it is largely based on amortized cost in U.S. insurance regulatory (“statutory”) reporting.

There are two ways to tie the value of the reserves to the value of the assets in this kind of principle-based
approach. One way is to directly determine the amount of assets needed on the valuation date to pay off the
contracts and set the value of the reserve equal to the value of those assets. The other way is to determine the rate
of investment return in each time step of an asset-liability cash flow simulation and discount the liability cash flows
at a rate equal to the path of projected investment returns. If the amount of starting assets in the simulation is just
sufficient to run out when the last contract payment is made, then these two methods should produce identical
results14.

Different accounting bases prescribe different asset valuation methods, and different asset valuation leads to the
reporting of different investment yields. When the reserve is calculated by discounting cash flows at the projected
investment return on the assets, one must use the investment return as constrained by the applicable accounting
standard. The reported investment return may differ from the return on market value in cases where assets and
reserves are not held at a market value. For example, the investment return as measured under U.S. statutory
accounting can be substantially different from current returns based on market value.

Stochastic valuation of reserves with an asset-liability simulation model involves calculation of a separate liability
value for each stochastic scenario and then using the set of scenario-specific values to determine a single reported
value. The next two subsections discuss methods commonly used to get scenario-specific values and to determine
the single reported value.

OBTAINING SCENARIO-SPECIFIC VALUES

When using stochastic scenarios for valuation of reserves in an asset-liability model, the amount of starting assets
required to fulfill the contract differs for each scenario, unless the asset portfolio is a true replicating portfolio. True
replicating portfolios are rare, so one must deal with the question of how to set the amount of starting assets.

As a practical matter, one does not use a different amount of starting assets for each scenario. This creates two
issues. First, one must determine what amount of assets to start with, and second, one must have a method to

14 Significant differences in results from these two approaches can arise when the path of investment returns has not been determined correctly based on
the accounting standard in use.

Copyright © 2022 Society of Actuaries Research Institute

 38

calculate the value of the reserve when the assets do not exactly satisfy the contract cash flows in a particular
scenario.

A generally accepted approach is to start with an amount of assets very close in value to the ending single reported
value of the reserve. This approach is somewhat circular and may require a bit of trial and error to arrive at that
value. Nevertheless, this approach is required in the Valuation Manual, where the amount of starting assets is
required to be within a specified narrow range around the reported reserve amount. Generally, this means that the
starting assets will be more than sufficient to fund the reserve in most scenarios, but there many still be some
particularly adverse scenarios where the assets run out before all contract obligations are paid. In those scenarios,
the asset-liability model simulates borrowing to pay the obligations, and the investment return becomes the rate of
interest simulated as paid on the borrowed funds.

If the same amount of assets is used at the start of each scenario, then the scenario-specific value of the reserve
cannot be set equal to the value of the starting assets. Therefore, the approach of discounting the cash flows at the
path of investment returns (as modeled scenario by scenario and as reported under the applicable accounting
standard) can be used to determine the scenario-specific value of the liability.

In the Valuation Manual, another method is sometimes required; it is the “greatest present value of accumulated
deficit” (GPVAD). Under this method, one determines a deficit defined as the excess of the outstanding liability over
accumulated assets at each time step in the projection 15. Each of those amounts is discounted back to the valuation
date, and the GPVAD is the most negative of those discounted values. The scenario-specific reserve value is then
defined as the sum of the value of the starting assets and the GPVAD. In a scenario where the starting assets are
exactly sufficient to pay off the obligations, the GPVAD is zero and the scenario-specific reserve value is equal to the
value of the starting assets.

The discount rate used for the GPVAD theoretically reflects the expected return on the extra starting assets that
would be required to fully fund the obligation and eliminate any deficits in a particular scenario. For regulatory
purposes that could be the same as the projected return on the assets in the asset-liability projection, or it could be
set conservatively at a level close to the risk-free rate. In either case, it is important to understand that this discount
rate is not used to discount cash flows; it is only used to discount accumulated deficits. The biggest part of the
liability value is set equal to the value of the starting assets, with the GPVAD often being a relatively small
adjustment that varies by scenario.

BLENDING SCENARIO-SPECIFIC VALUES INTO A SINGLE REPORTED VALUE

When the accounting basis for asset valuation is market value, then risk-neutral methods can be used, and the
single liability value is the simple average of the scenario-specific values.

Any valuation of insurance contracts needs to include a margin for risk. Risk-neutral methods include a margin
implicitly because risk-neutral calibration gives added probability weight to adverse scenarios. Whenever real-world
scenarios are used, the simple average of the scenario-specific values does not include such an implicit margin and a
margin must be added in some other way.

The two most common methods for adding such a margin in the context of real-world scenarios are these:

15 Since it can be difficult or time-consuming to determine the outstanding liability value at each time step in a scenario, an approximation or “working
reserve” may be defined for this purpose. Sometimes the “working reserve” is simply zero and the deficit is simply the negative of the amount of
accumulated assets.

Copyright © 2022 Society of Actuaries Research Institute

 39

• Contingent tail expectation (CTE). Under this method the scenario-specific values are sorted from least to
greatest, and only the greatest are included when calculating an average value. The number of scenarios
included in the average is a specified percentage of the total. The Valuation Manual specifies the “70 CTE”
which means that the smallest 70% of the scenarios are excluded from the average and only the largest
30% are included.

• Cost of capital method. Under this method the scenario specific values are each increased by adding an
imputed cash flow equal to the cost of capital in each time step of the scenario. The margin is the present
value of those imputed cash flows. The cost of capital is intended to represent the market price of risk. In
theory this approach may be considered consistent with market pricing but estimating the cost of capital is
felt by some actuaries to be problematic.

SELECTING A SCENARIO GENERATOR THAT IS FIT FOR PURPOSE

If the accounting basis for assets is market value, then the scenarios used for liability valuation should be able to
reproduce market prices of traded assets. While risk-neutral scenarios are always calibrated to do that, real-world
scenarios may or may not be calibrated that way.

When using a real-world ESG it is important that the calibration approach be appropriate for the accounting basis.
Real-world scenarios that are not calibrated to reproduce market prices may be appropriate when the accounting
basis is not market value.

Recall that valuation using real-world scenarios must include an add-on margin for risk. When the scenarios used
are not market-coherent, then the margin may arise partly from conservatism in the scenario calibration and partly
from an explicit margin. That is the case with the Valuation Manual. U.S. regulators mandate the use of a real-
world scenario generator that is not market-coherent and then add a margin using the CTE approach.

A special issue arises when hedging is to be simulated within each scenario in the asset-liability model. Hedging
depends on market-consistent16 valuation, including “Greeks”. To simulate hedging activity in each time step one
needs market-consistent values for both the hedges and the contractual obligations being hedged. Such market-
consistent values can be obtained through separate stochastic valuations at each time step. Stochastic valuations at
each time step in a long-term simulation are often called “inner loop” valuations to indicate that the scenarios used
for valuation are separate and different from the “outer loop” scenario that defines the conditions at each time
step. The starting date for a set of “inner loop” scenarios is a valuation date anywhere within time span of the outer
loop scenario. Since inner loop valuations are often used to determine a market-consistent value for derivatives
used for hedging, they often employ risk-neutral scenarios calibrated on the fly to the yield curve at their starting
date. This is true even if the outer loop scenario is real-world and neither market-consistent nor market-coherent in
nature. In situations involving hedging, a risk-neutral ESG may be used for the inner loop valuations even though
the outer loop is based on scenarios from a real-world ESG. That can be appropriate, for example, when simulating
a clearly defined hedging strategy in a valuation that complies with the Valuation Manual where the outer loop
scenarios are real-world and neither market-consistent nor market-coherent.17

16 In this paragraph the term “market-consistent” means a valuation technique that reproduces observed market prices. Both “market-consistent” and
“market-coherent” valuations satisfy this definition.
17 See Question 12.1 (Slutsker 2019).

Copyright © 2022 Society of Actuaries Research Institute

 40

References

Center for Insurance Policy and Research. Principle-Based Reserving (PBR). National Association of Insurance
Commissioners (NAIC.org). Last updated May 28, 2021. https://content.naic.org/cipr-topics/principle-based-
reserving-pbr

Fabozzi, Frank. 2021. Fixed Income Securities. Ninth Edition. McGraw Hill.

IFRS. 2021. IFRS 17 Insurance Contracts Standard 2022 Issued. International Financial Reporting Standards
Foundation. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/

Pedersen, Hal, Mary Pat Campbell, Stephen L Christiansen, Samuel H. Cox, Daniel Finn, Ken Griffin, Nigel Hooker,
Matthew Lightwood, Stephen M Sonlin and Chris Suchar. 2016. Economic Scenarios – A Practical Guide. Chicago:
Society of Actuaries. https://www.soa.org/globalassets/assets/Files/Research/Projects/research-2016-economic-
scenario-generators.pdf

Slutsker, Benjamin and Dylan Strother, Cindy McGovern et al. 2019. Principle-Based Approach Projections Practice
Note. Washington. American Academy of Actuaries. https://www.actuary.org/sites/default/files/2019-
12/PBA_Projections_Practice_Note.pdf

Copyright © 2022 Society of Actuaries Research Institute

 41

About The Society of Actuaries Research Institute

Serving as the research arm of the Society of Actuaries (SOA), the SOA Research Institute provides objective, data-
driven research bringing together tried and true practices and future-focused approaches to address societal
challenges and your business needs. The Institute provides trusted knowledge, extensive experience and new
technologies to help effectively identify, predict and manage risks.

Representing the thousands of actuaries who help conduct critical research, the SOA Research Institute provides
clarity and solutions on risks and societal challenges. The Institute connects actuaries, academics, employers, the
insurance industry, regulators, research partners, foundations and research institutions, sponsors and non-
governmental organizations, building an effective network which provides support, knowledge and expertise
regarding the management of risk to benefit the industry and the public.

Managed by experienced actuaries and research experts from a broad range of industries, the SOA Research
Institute creates, funds, develops and distributes research to elevate actuaries as leaders in measuring and
managing risk. These efforts include studies, essay collections, webcasts, research papers, survey reports, and
original research on topics impacting society.

Harnessing its peer-reviewed research, leading-edge technologies, new data tools and innovative practices, the
Institute seeks to understand the underlying causes of risk and the possible outcomes. The Institute develops
objective research spanning a variety of topics with its strategic research programs: aging and retirement; actuarial
innovation and technology; mortality and longevity; diversity, equity and inclusion; health care cost trends; and
catastrophe and climate risk. The Institute has a large volume of topical research available, including an expanding
collection of international and market-specific research, experience studies, models and timely research.

Society of Actuaries Research Institute

475 N. Martingale Road, Suite 600

Schaumburg, Illinois 60173

www.SOA.org

Copyright © 2022 Society of Actuaries Research Institute