Is bitcoin an inflation hedge?

Harold Rodriguez∗1 and Jéfferson Colombo†1

1 São Paulo School of Economics, Getulio Vargas Foundation (FGV-EESP)

2 São Paulo School of Economics, Getulio Vargas Foundation (FGV-EESP)

Abstract
Spot bitcoin ETFs have been recently approved in the U.S., increasing retail and insti-
tutional investors’ attention to the crypto space. Still, empirical evidence on whether
Bitcoin is an asset that protects investors against inflation is still inconclusive. To
contribute to this debate, we analyze the effect of inflation shocks on bitcoin returns
through the estimation and inference of Vector Autoregressive Models (VARs). Unlike
previous research on the topic, we identify inflation shocks as surprises in the US’s
CPI and Core PCE announcements: the difference between the announced inflation
and the analysts’ consensus. The results, based on monthly data between August 2010
and January 2023, indicate that bitcoin returns increase significantly after a positive
inflationary shock, corroborating empirical evidence that Bitcoin can act as an inflation
hedge. However, we observe that bitcoin’s inflationary hedging property is sensitive
to the price index – it only holds for CPI shocks – and to the period of analysis —
the hedging property stems primarily from sample periods before the increasing in-
stitutional adoption of BTC (“early days”). Thus, the inflation-hedging property of
Bitcoin is context-specific and is likely to be diminishing as adoption increases. This
research contributes to the still under-explored strand of literature that analyzes the
hedging and safe-haven properties of Bitcoin and benefits asset managers, investors,
and monetary authorities.

Keywords: Bitcoin, Hedge against inflation, Unexpected inflation, surprises in CPI,

surprises in PCE. JEL classifications: E44, E31, G11.

∗
E-mail address: haroldrodriguez152@hotmail.com.
†
Corresponding author. E-mail address: jefferson.colombo@fgv.br. ORCiD: https://orcid.org/
0000-0001-7221-8074. We appreciate comments from Thorsten Beck (discussant), Rodrigo de Oliveira
Leite, João Marco Braga da Cunha, Marcel Ribeiro, Alexandre Ludolf, Fernando Chague, and participants
at the 1st Elsevier Finance Conference (Rio de Janeiro, 2023). An earlier version of this paper has been
granted 1st place in the CFA Society Brazil award for Best Monographs in Finance (2023). Jéfferson Colombo
gratefully acknowledges financial support from the National Council for Scientific and Technological Develop-
ment (CNPq, Grant #313033/2022-6) and the Silicon Valley Community Foundation through the University
Blockchain Research Initiative (UBRI, Grant #2022-199610).

Electronic copy available at: https://ssrn.com/abstract=4763347

1 Introduction
Inflation is a critical investment risk and one of the foremost challenges to conducting business
in the upcoming years (World Economic Forum, 2023). Consequently, investors are urged
to formulate strategies to mitigate inflation risk. A viable safeguard involves incorporating
in the portfolio assets that appreciate alongside inflation (Tarbert, 1996). Such a need has
gained even more importance given the recent generalized price increase in various parts of
the world, including developed economies, where low interest rates and controlled inflation
have coexisted for decades. But not anymore.
The recent consideration of Bitcoin as an investment instrument (Conlon et al., 2021),
along with the growing interest among academics and policymakers (Phochanachan et al.,
2022), has contributed to positioning the cryptocurrency as a potential inflation hedge. Com-
monly cited justifications for this role include its limited supply and decentralized network,
conferring scarcity and resilience (Bouri et al., 2017a,b). However, existing theoretical and
empirical studies have not reached a consensus on Bitcoin’s ability to hedge for inflation –
while Blau et al. (2021) and Choi and Shin (2021) support Bitcoin as a robust inflation hedge,
most studies indicate that Bitcoin inflation hedging properties are context-specific (Conlon
et al., 2021, Matkovskyy and Jalan, 2020, Phochanachan et al., 2022, Smales, 2021). This
lack of consensus may be attributed to methodological differences, varying time horizons,
and the specific characteristics of the economies analyzed in these studies.
One key aspect of answering this research question is whether a given inflation announce-
ment carries unexpected information or not. To demonstrate this point, let us consider the
following two situations. If the observed CPI matches analysts’ consensus (i.e., market ex-
pectations), asset prices should not react, regardless of the level of inflation. However, if at
least part of the information was a surprise, the announcement generates new information
for financial markets, and asset prices should move accordingly. Thus, disentangling the
expected and unexpected parts of inflation announcements is critical to understanding the
role of inflation shocks in shaping the fluctuations in Bitcoin’s price.
That is precisely how we depart from previous research on the topic – by enhancing
the identification of inflation shocks in a Vector Autoregressive (VAR) system. Specifically,
instead of identifying inflation shocks as the residual from a typical regression of observed (or
expected) inflation on the lags of itself and of several variables (S&P 500, VIX, interest rates,
etc.) (for example, as in Choi and Shin, 2021), we focus on the unexpected component of
inflation releases, measured here as the difference between actual monthly inflation and the
consensus forecast of market analysts. By doing that, our innovations (or shocks) consider
all observed and unobserved factors that market agents consider in their projections of
inflation, and not only a few covariates included in the VAR system. As Nakamura and
Steinsson (2017) notes, any factors left out as covariates in the regression will be part of
the “shock”, even if they are anticipated by market agents. Thus, we believe that our
identification strategy provides a cleaner determination of the unexpected portion of inflation
announcements (that is, surprise inflation) and helps us understand the role of inflation in

Electronic copy available at: https://ssrn.com/abstract=4763347

bitcoin returns.
Given this crucial conceptual issue, we contribute to this inconclusive but critical debate
by estimating and analyzing, through VARs, the short-term effect of inflation shocks on
Bitcoin price fluctuations. Despite the unexpected component of inflation releases mentioned
above, we follow previous studies and include in the system of equations the joint effects of
gold prices, the S&P500 and VIX indexes, the one-year US Treasury bill yield, and bitcoin
prices. Our data is monthly and ranges from August 2010 to January 2023.
The results of the impulse response functions indicate that Bitcoin returns increase fol-
lowing a positive inflation shock, ceteris paribus. Compared to other assets, such a pattern
is similar to the responses observed on Gold and contrasts with the one observed on the
S&P 500. This finding suggests that Bitcoin acts as a hedge against unexpected fluctuations
in inflation, which corroborates some of the empirical evidence. Furthermore, the variance
decomposition analysis indicates that the return-inflation relationship may exhibit greater
strength for gold than for Bitcoin, consistent with the findings of Smales (2021).
However, further analyses suggest that the efficacy of Bitcoin as an inflation hedge ap-
pears contingent on two aspects: the selected price index and the sample period. Regarding
the former, when considering the Core PCE index instead of the CPI, the cryptocurrency
return reveals a negative response to a positive inflation shock. This discrepancy can be
attributed to different criteria, such as composition, weighting, and release dates. Regarding
the former, we find a notable decrease in the inflation-hedging properties of Bitcoin when we
exclude the initial sampling period (“early days” of Bitcoin), suggesting that such hedging
property may be diminishing as adoption – and, consequently, market fluctuations – become
mainstream. Thus, the inflation-hedging property of Bitcoin is likely to be diminishing as
adoption increases.
This research contributes to a growing but still underexplored debate on Bitcoin hedging
properties in the face of inflationary shocks (Blau et al., 2021, Choi and Shin, 2021, Con-
lon et al., 2021, Matkovskyy and Jalan, 2020, Phochanachan et al., 2022, Smales, 2021).
Notably, our contributions are twofold. First, we identify inflation shocks through their sur-
prising or unexpected components. This approach fundamentally differs from most previous
studies that have examined Bitcoin’s hedging ability based on its expected (or actual) com-
ponent. Second, in contrast to studies that considered unexpected inflation in their analyses
(Matkovskyy and Jalan, 2020, Smales, 2021), our empirical methodology takes into account
the dynamic and temporal relationships among the variables in the model, which is especially
relevant in time series analyses. At the same time, we seek to mitigate possible endogeneity
problems that often arise in this type of investigation.
Furthermore, as this is a current discussion with significant practical implications for asset
managers, investors, and monetary authorities, the empirical evidence presented in this study
directly affects resource allocation decisions and allows for a recent analysis of the behavior
of Bitcoin, gold, and other assets following inflation shocks. In particular, recent research has
shown that sophisticated retail investors are attracted to cryptocurrencies for macroeconomic
and portfolio reasons (Colombo and Yarovaya, 2024). Our study is also timely since the SEC

Electronic copy available at: https://ssrn.com/abstract=4763347

approved spot Bitcoin ETFs only in 2024, bringing tremendous attention and demand for
Bitcoin exposure in retail investors’ portfolios.
The remainder of this paper is organized as follows: Section 2 presents the theoretical
framework and methodology used in this study. Section 3 presents the data used in this
work. Section 4 reports the empirical results, and Section 5 presents the conclusions.

2 Methodology
2.1 VAR model
To capture the dynamics of the variables and deal with the potential endogeneity in their
relationships, we use a VAR framework (Sims, 1980), which is reliable in describing, sum-
marizing, and forecasting macroeconomic data (Stock and Watson, 2001).
Specifically, let y1 , y2 , . . . , yT be multivariate time series, with yt = (y1t , y2t , . . . , ykt )′ . A
VAR model of order p, in its basic form, can be expressed as follows:

yt = v + A1 yt−1 + A2 yt−2 + . . . + Ap yt−p + ut ; t = 1, 2, 3, . . . , N (1)

All symbols used in this representation have usual meanings, that is, yt = (y1t , y2t , . . . , ykt )′
is a random vector (K × 1) of endogenous variables; v = (v1 , v2 , . . . , vk )′ is a fixed vector
of intercepts (K × 1), which allows the possibility of a non-zero mean E(yt ); Ai are fixed
matrices of coefficients (K ×K) where i = 1, . . . , p, which are interpreted as the sensitivity of
a model variable about a lag of another variable; ut = (u1t , u2t , . . . , ukt )′ is a K-dimensional
vector of white noise, such that E(u) = 0, E(ut u′t ) = Σu , and E(ut u′s ) = 0 for s ̸= t, that is,
it represents random shocks that do not correlate and are time-invariant.
Under the stability condition, the process Yt has the MA representation defined as:
∞
X
Yt = µ + Ai Ut−i (4)
i=0

Where Yt is modeled as a function of the mean term µ and in terms of the past and
present of the innovation vector Ut .
Furthermore, the MA representation of yt can be found by pre-multiplying Yt by a matrix
J = [IK : 0 : . . . : 0] of dimension (K × Kp). Here, µ = Jµ, Φi = JAi J ′ , and ut = JUt .
∞
X ∞
X
yt = JYt = Jµ + Φi ut−i = µ + Φi ut−i (5)
i=0 i=0

The challenge of estimating these parameters incurs the same difficulties as obtaining
the parameters of the primitive model from the reduced model. This methodology does not
permit estimation in cases of underidentification, where the number of equations is fewer
than the number of unknowns. To identify the system written in MA form, we estimate each

Electronic copy available at: https://ssrn.com/abstract=4763347

equation by OLS and then compute the Cholesky factorization of the reduced form VAR
covariance matrix (Lütkepohl, 2005).

2.2 Model checking

Stationarity. The formal approach to ascertain the stationarity of time series involves em-
ploying the unit root test, with the Augmented Dickey-Fuller (ADF) test serving this pur-
pose. The null hypothesis indicates the existence of a unit root, indicative of non-stationarity.
Its rejection implies that the underlying stochastic process governing the series is time-
invariant. In that case, the VAR model can be estimated with the series at level. Throughout
the implementation of the ADF test, the analysis considered various model specifications,
including those without a constant or trend, with a constant, and with both a constant and
a trend.

Serial autocorrelation. To address the threat of serial autocorrelation, which has the
potential to compromise the integrity of impulse response functions, the authors employ the
Ljung and Box (1978) test. The number of lags, denoted as h, is determined by the Schwert
(1989) criterion. In short, the null hypothesis posits that residuals up to the h-th lag exhibit
white noise characteristics. When rejected, it implies that at least one autocorrelation is
statistically different from zero. In such a case, the model must be rejected, prompting a
re-specification and re-estimation of the VAR (Lütkepohl, 2011).

Model stability. To appraise the temporal consistency of a VAR model, it is imperative

to scrutinize its stability, denoting the absence of explosions in the time series and the
predictability of behavior over time. The stability examination aims to validate whether the
roots of the eigenvalues in the coefficient matrix fall within the unit circle. If the eigenvalues
modulus resides within the unit circle, the model attains stability, rendering the dependent
variables weakly stationary. Conversely, if the eigenvalues modulus lies outside the unit
circle, the model is deemed unstable, and the dependent variables are integrated.
According to Lütkepohl (2005), the stability of the VAR model assumes paramount sig-
nificance in ensuring the reliability of results, as an unstable model can engender inaccurate
predictions, particularly sensitive to minor fluctuations in input data. Furthermore, as high-
lighted by Baum (2013), adherence to stability conditions is imperative for the accurate
estimation of Impulse Response Functions and Variance Decomposition.

2.3 Model evaluation

Impulse Response Function (IRF). The IRF illustrates how a shock to a specific variable prop-
agates to others over time, enabling the measurement of the magnitude and time horizon of
the impact. In linear models with uncorrelated error terms, this process is straightforward.
However, in the presence of correlated errors, identifying shocks with specific variables be-
comes complex and lacks clarity. This complexity arises from common components in errors

Electronic copy available at: https://ssrn.com/abstract=4763347

affecting multiple variables. According to Farias (2008), a common approach in such cases is
to arbitrarily assign the effects of these shared components to the variable that appears first
in the system. The drawback of this approach is its dependency on the particular equation
order within the model.
In a VAR(p) model, the derivation of IRFs involves inducing a one-period shock to an
endogenous variable. Specifically, an increase in u1 by one standard deviation at time t = 0
is considered, maintained for only one period, constituting a ”boost”. The impact of this
shock permeates through the model, influencing all endogenous variables. Subsequently, a
one-period shock is introduced to the next endogenous variable, for instance, an increase in
u2 by one standard deviation. This iterative process extends to all endogenous variables,
allowing for a comprehensive tracking of effects across the entire model.

Forecast error variance decomposition (FEVD). A complementary approach to analyzing

the results of the VAR model is through variance decomposition. This technique enables the
breakdown of forecast error variances for each variable into components attributable to the
variable itself and others. The outcome is a percentage-wise representation of the impact
that a shock to a specific variable exerts on itself and other system variables.
According to Enders (2008), if the shocks observed in a variable, y2 t, cannot account
for the variance in the forecast error of variable y1 t, the sequence y1 t is deemed exogenous;
otherwise, it is considered endogenous. Lütkepohl (2005) further notes that this technique
is sensitive to changes in the considered system. The forecast error variance components
may undergo alterations with system expansions, including additional variables or exclu-
sions of series from the model. Furthermore, factors such as measurement errors, seasonal
adjustments, and the use of aggregated data can impact the outcomes of this approach.

3 Data
This study utilizes monthly data from August 2010 to January 2023 (150 observations).
The beginning of this time frame stems from the earliest available data on Bitcoin trading
prices. The variables used in this study (in stationary form) are the following: returns on
S&P500 ( Ret SP500), Gold (Ret Gold), and bitcoin ( Ret BTC); first differences of the
US 1y interest rates (Dif US 1Y) and the VIX (Dif VIX); and surprise inflation (Surp Inf).
Table I describes and details the variables considered in this study.
Our framework for variable selection, estimation, and inference is similar to Choi and
Shin (2021), but we introduce modifications to enhance the identification of inflation shocks.
These adjustments are inspired by the work of Kuttner (2001), Romer and Romer (2004), and
Nakamura and Steinsson (2017), and the specification proposed by Lowenkron and Garcia
(2007). Particularly, as described in Table I, we include in the VAR system inflation surprises
estimated as the difference between the actual (CP It ) and the expected (Et−1 (CP It )) CPI
MoM (also CPI YoY and Core PCE MoM in further analyses). With such refinements,

Electronic copy available at: https://ssrn.com/abstract=4763347

we seek more assertive estimates and robust outcomes by identifying inflation through its
surprising component.
One key operational concern is to homogenize each variable’s measurement period to
coincide with inflation announcements, in the same spirit as Kuttner (2001), Cochrane and
Piazzesi (2002), and Hanson and Stein (2015).1 Thus, we consistently use the closing quote
on the day of the inflation announcement relative to the closing price of the previous day for
all series. By doing that, our variables capture the impact of the inflation announcement,
regardless of different trading hours across various financial markets.
Particularly, the specification of variables, excluding Surp Inf , involves calculating changes
from t − 1 to t for each announcement of the U.S. economy price index, CPI or PCE (day
t). This approach aims to capture the complete market response to the announcement, op-
erating under the implicit assumption that the entire reaction to the economy price index
announcement may not be immediate. Investor uncertainty about the implications of news
may lead to a gradual reaction, where beliefs are updated as others’ interpretations unfold
through trading volume, the pricing process, and financial media. Consistent with this idea
and similar to our case, Hanson and Stein (2015) argues the full reaction to an FOMC
announcement might not be instantaneous.2
Regarding the scale transformation applied to specific variables, this is a practice fre-
quently adopted in this type of study. As Morettin and de Castro Toloi (2006) emphasize,
working with scale-free returns rather than asset prices is recommended since the former
have more advantageous statistical properties.

3.1 Descriptive analysis

As shown in Table II, the centrality measures across all series are similar and nearly zero.
Regarding asymmetry values, the return series exhibit negative asymmetry, with Ret Gold
approaching a distribution closest to symmetry. Kurtosis results depart from normal stan-
dards, with most demonstrating a leptokurtic distribution. Surp Inf is the closest to a
normal pattern, with a kurtosis of 2.59. Regarding volatility, it is noteworthy that the stan-
dard deviation of Bitcoin returns significantly exceeds that of gold, consistent with findings
by Choi and Shin (2021), Smales (2021) and Phochanachan et al. (2022).
Since the surprise measures of inflation are key to our analysis, we take a closer look on
them. Figure A1, available in Appendix A, illustrates the frequency distribution graphs for
the Surp Inf , Surp Inf Y oY , and Surp Inf P CE series. For the first two, the actual value
aligned with market expectations approximately 33% of the time. This proportion is notably
lower compared to the Surp Inf P CE series, where 55% of the time the inflationary surprise
1
The authors use a window of one or two days around FOMC meetings to identify the effects of changes
in monetary policy on financial markets.
2
More specifically, Fleming and Remolona (1999) explored price dynamics in the financial market following
the release of crucial economic announcements, such as monetary policy decisions. Their findings indicate a
gradual formation of prices in the market after such information is released, characterized by elevated levels
of volume and volatility persisting for at least 90 minutes.

Electronic copy available at: https://ssrn.com/abstract=4763347

 Table I. Data sources and description of model variables

Variable Calculation Definition and source

Ret SP500 Ln(SP 500t /SP 500t−1 ) The log-return is derived from a theoretical portfolio comprising the 500 largest
companies traded on the NYSE or NASDAQ exchanges. It is widely used to
capture general financial market conditions. Source: Yahoo Finance.
Ret Gold Ln(Goldt /Goldt−1 ) The log-return is obtained from gold futures contracts traded on the COMEX,
a division of NYMEX. It serves as a crucial benchmark for investors seeking to
track the price dynamics of gold in the financial market. Source: Yahoo Finance.
Ret BTC Ln(BT Ct /BT Ct−1 ) The log-return is calculated from the real-time market price of Bitcoin, deter-
mined by global transactions. Source: Investing.com.
Dif US 1Y U S1Yt − U S1Yt−1 Represents the change in the par yield of U.S. Treasury securities with a one-year
maturity. It is based on the closing bid prices of the most recently auctioned
securities in the over-the-counter market. The rate is published daily and reflects
market expectations concerning the economic trajectory and monetary policy.
Source: United States Department of the Treasury.
Dif VIX V IXt − V IXt−1 Represents the change in the reference indicator measuring the expected short-
term volatility of the stock market. It is based on the stock options prices
comprising the S&P500 index and is updated in real-time, during trading ses-
sions. Source: Yahoo Finance.
Surp Inf CP It − Et−1 (CP It ) Represents the unexpected component of inflation, calculated at the moment
of the official inflation rate announcement as the difference between the actual
(CP It ) and the expected (Et−1 (CP It )) CPI MoM. Alternative metrics such as
CPI YoY and Core PCE MoM will also be taken into account. MoM (YoY)
data are (are not) seasonally adjusted. Source: U.S. Bureau of Labor Statistics,
Bureau of Economic Analysis and Investing.com.

Note: The subscript t (t − 1) denotes, except for Surp Inf , the closing price of the asset/index
on the day (previous day) of the official inflation rate announcement.

equaled zero. This discrepancy may be attributed to differing criteria, such as composition
and disclosure dates (the PCE announcement occurs after the CPI release for the same period
of reference; thus, market agents update beliefs following any surprise in CPI). We can see
in the figures that the dispersion of surprises is larger for CPI MoM and CPI YoY than for
PCE MoM. Furthermore, surprises in CPI are positively skewed – i.e., extreme values are
more common in the right tail (positive ones).
Another important issue to discuss is the timing and correlation of these surprises. Fig-
ure A2 shows the actual, forecast, and surprise (actual - forecast) values for our baseline
measure of inflation (CPI MoM, Panel A) and for the further analyses (CPI YoY, Panel B,
and PCE MoM, Panel C). We can observe from the Figure that average surprises are close
to zero; however, positive surprises started to become more frequent in the recent period,
starting in the middle of 2021, following a supply-side disruption brought by COVID-19.3
Such a pattern is more pronounced in the CPI MoM and CPI YoY, suggesting that positive
surprises in CPI are incorporated in PCE forecasts, reducing the forecast error in the latter
3
On May 12, 2021 (reporting inflation from April 2021), the actual CPI MoM came on 0.8%, while the
expected inflation was 0.2%. Such a surprise of 0.6 p.p. is the largest value in our sample period. After that,
market agents witnessed a sequence of positive unexpected inflation values that made the FOMC start, on
March 16, 2022, eight successive increases in the policy rate.

Electronic copy available at: https://ssrn.com/abstract=4763347

 Table II. Summary statistics

Variable Mean p50 SD Asymmetry Kurtosis Min Max Obs.

Ret SP500 -0.000 0.001 0.012 -0.680 6.759 -0.050 0.054 150
Dif VIX -0.123 -0.175 1.899 1.794 9.214 -5.710 11.090 150
Surp Inf -0.000 0.000 0.136 0.769 2.588 -0.400 0.600 150
Dif US 1Y 0.003 0.000 0.045 2.607 13.714 -0.160 0.230 150
Ret Gold 0.001 0.001 0.009 -0.210 1.341 -0.029 0.028 150
Ret BTC -0.004 0.000 0.091 -5.094 49.918 -0.849 0.288 150

Note: This table provides statistics for all series covering the period from 2010:08 to 2023:01.
The Ret (Dif) prefix indicates that the series was derived from returns (differences) calculated
within one-day windows around monthly CPI MoM announcements.

inflation index.
Finally, Figure A3 plots the three surprise measures and shows the correlation coefficient
among them: the contemporaneous linear association is stronger for the pair CPI MoM-CPI
YoY (0.63), but almost uncorrelated on the pairs CPI MoM-PCE MoM (0.12) and CPI YoY-
PCE MoM (0.02). Such evidence reinforces that markets adapt expectations and reduce the
forecast error for the PCE, whose announcement date always occurs later than the CPI.
Finally, Figure A4 shows the cross-correlation among each surprise pair, and again shows a
clear pattern for CPI MoM-CPI YoY: the correlation coefficient is a lot stronger in t0 (i.e.,
their contemporaneous values). On the other hand, the relationship between CPI and PCE
is noisier, showing that the correlation is near zero for contemporaneous values and for other
lead-lag relationships.

3.2 Model identification

Since the literature does not provide clear theoretical or empirical evidence on the causal
relationships among all variables, we identify the structural shocks (i.e., the uncorrelated or
orthogonal errors, not the reduced-form errors that correlate across equations) based on an
economically reasonable temporal ordering criterion. This procedure represents a recursive
VAR, where the error term in each regression equation is constructed to be uncorrelated
with the error the the preceding equations (Stock and Watson, 2001).
Our ordering criteria for the Choleski triangular decomposition of the residuals are the
following. Initially, variables representing broad measures of economic activity and market
volatility are selected for their ability to assimilate new information promptly. Subsequently,
series reflecting the dynamics of prices in the economy and the stance of monetary policy are
incorporated. Lastly, variables theoretically presumed to be influenced by the aforementioned
factors are considered: the price of a reference financial asset and the price of an emerging
financial asset.
Given the relatively lower significance of Bitcoin in financial markets (Choi and Shin,
2021), as well as empirical findings suggesting its lesser efficiency compared to stock and

Electronic copy available at: https://ssrn.com/abstract=4763347

gold markets in assimilating available information (Al-Yahyaee et al., 2018), which also
indicate its responsiveness to changes in inflation, interest rates (Li and Wang, 2017), and
financial market volatility (Bouri et al., 2017a), it is reasonable to treat Bitcoin as the most
endogenous variable in the system.
Regarding the order of remaining variables, support for the identification assumption is
drawn from the literature on the monetary VAR model, particularly highlighted by Choi and
Shin (2021). Studies such as those by Christiano et al. (2005) and Coibion (2012) placed
inflation before short-term interest rates in the Cholesky identification method applied in
their analyses. Additionally, empirical research conducted by Beaudry and Portier (2006) and
Jordà et al. (2017) provides evidence indicating that movements in financial variables, such
as stock prices, may precede changes in macroeconomic variables like inflation and interest
rates. The analyses by Rigobon and Sack (2003) and Kurov et al. (2022) are particularly
relevant in this context, revealing the FED’s responsiveness to the stock market dynamics,
adjusting its monetary policy in response to changes in market conditions.
In light of the evidence provided, the baseline VAR model includes an intercept and six
variables arranged in the following sequence: the log-return of the S&P500 (Ret SP 500), the
variation of the VIX index (Dif V IX), the surprise inflation (Surp Inf ), the variation in
the one-year US Treasury bill yield (Dif U S 1Y ), the log-return of gold prices (Ret Gold),
and the log-return of Bitcoin prices (Ret BT C).

3.3 Model checking

The model-checking outcomes, featured in Appendix B, demonstrate favorable statistical
properties, enhancing the reliability of the empirical results. Particularly, Table B1 exhibits
the ADF test findings, indicating that all series have no unit root and are stationary at a
0.01 significance level. Such a result is robust to several model specifications – without drift
or trend, with drift, with drift and trend, and with drift and linear and quadratic trends.
Furthermore, Table B2 displays the outcomes of the portmanteau test, revealing that, at
a 0.05 significance level, only the VAR(6) model residuals exhibit white noise characteristics
up to the maximum lag.4 Finally, Table B3 showcases the findings of the stability diagnosis
test applied to the VAR(6) model, indicating that all eigenvalues reside inside the unit circle,
affirming the dynamic stability of the model.

4 Empirical Results
4.1 Baseline model
Figure 1 depicts the main empirical results of this study, considering the CPI MoM as the
underlying index to calculate the inflation surprises. Specifically, the Figure shows the cumu-
4
Defined by the Schwert (1989) criterion.

10

Electronic copy available at: https://ssrn.com/abstract=4763347

lative responses of each variable in the system following an impulse in the surprise component
of the CPI MoM announcements.5 Overall, we can infer that an unexpected, positive shock
in CPI MoM increases Gold, BTC, and US1y interest rates, while it decreases the S&P
500 and has no effect on the VIX. While the results are statistically non-significant in most
cases (the exception is the impact on interest rates in some horizons), such a pattern reveals
that responses in BTC are similar to the ones of Gold and interest rates and contrast those
of the S&P500.6 Put differently, interest rates, Gold, and BTC returns increase following
inflation shocks, while S&P returns diminish. Thus, the results suggest that Bitcoin could
be a helpful hedge against unexpected fluctuations in inflation, an overall finding that aligns
with empirical evidence supporting Bitcoin’s role as an inflation hedge (Blau et al., 2021,
Choi and Shin, 2021).

Figure 1. COIRF - Baseline model (CPI MoM)

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of the six
variables to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2010:08 and 2023:01. The Ret (Dif) prefix indicates that the series
was constructed from returns (differences) calculated in one-day windows around monthly CPI
MoM announcements. The horizontal axes show the number of months after the impulse.

The baseline, entire period results highlight further similarities in the behavior of BTC
and GOLD. Firstly, both assets exhibit an initial negative effect, with Bitcoin displaying a
5
Figure C1 in Appendix C plots the entire set of impulse response functions for a comprehensive picture.
6
Relative to BTC, The COIRFs suggest a similar, albeit less pronounced, response in gold returns to
inflation shocks, indicating that the returns of both assets react in the same direction but with different
intensities.

11

Electronic copy available at: https://ssrn.com/abstract=4763347

more significant but less enduring impact. For both assets, there is a convergence to the final
cumulative effect after the tenth month. Second, the forecast error variance decomposition
of Bitcoin and gold returns (as shown in Figure 2) indicates that Bitcoin (gold) return
fluctuations are predominantly explained by a shock to itself, decreasing over time from 94%
(92%) in the initial period to 69% (64%) in the final period. These results partially contrast
with the findings of Choi and Shin (2021), who found this property only in gold. Regarding
the contribution of a shock in inflation to the variance in the forecast error of the two series,
gold is superior (8%) compared to Bitcoin (6%). This finding suggests that the relationship
between return and inflation may be slightly more substantial for the commodity than for
the cryptocurrency, aligning with the results of Smales (2021) and consistent with studies
by Corbet et al. (2020) and Pyo and Lee (2020), which found that CPI announcements have
a negligible impact on Bitcoin prices.
Figure 2. FEVD - Baseline model

Note: This figure shows forecast error variance decomposition (FEVD) of Bitcoin and gold for
the sample period between 2010:08 and 2023:01. The Ret (Dif) prefix indicates that the series
was constructed from returns (differences) calculated in one-day windows around monthly CPI
MoM announcements. The horizontal axes show the number of months after the impulse.

4.1.1 CPI MoM vs. CPI YoY

So far, we have considered inflation measured on a month-over-month (MoM) basis. One
drawback of this approach is that short-term measures of inflation may be noisy. Thus,
in this section, we replace the MoM CPI with the 12-month, year-over-year (YoY) CPI.
This further analysis brings forth several potential benefits. Firstly, employing CPI YoY
enables the capture of a more comprehensive perspective on price variations over time,
providing a more representative view across an extended period. Additionally, it facilitates
the identification of long-term inflation trends, as monthly fluctuations may be influenced
by temporary factors that could distort the analysis. Hence, it is possible to obtain a more
consistent view of inflationary pressures by opting for an annual inflation measurement.
Figure 3 illustrates that the results from cumulative orthogonalized impulse-response
functions (COIRFs) are preserved even with the alternative specification of inflation shocks.

12

Electronic copy available at: https://ssrn.com/abstract=4763347

Relative to the previous set of results (Figure 1), the cumulative responses are very similar
in terms of magnitude for both BTC and GOLD, reinforcing that they behave similarly.

Figure 3. COIRF - CPI YoY

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of Bitcoin
and gold to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2010:08 and 2023:01. The Ret prefix indicates that the series was
constructed from returns calculated in one-day windows around monthly CPI YoY announce-
ments. The horizontal axes show the number of months after the impulse.

4.1.2 CPI MoM vs. Core PCE MoM

Complementing the CPI, another relevant consumer price index in the U.S. is the Personal
Consumer Expenditures (PCE), released by the Bureau of Economic Analysis (BEA). Al-
though similar in spirit, the CPI and the PCE carry distinctions in four dimensions (Bureau
of Labor Statistics (BLS), 2011): formula effect (Laspeyres vs. Fisher-Ideal), weight effect
(Consumer Expenditure Survey vs. business surveys), scope effects (CPI considers all urban
households, while the PCE considers goods and services consumed by all households, and
nonprofit institutions serving households), and other effects (e.g., seasonal-adjustment dif-
ferences, price differences, etc.). Furthermore, the PCE has been the official inflation target
of the FOMC since January 2012, and looking to the core inflation may eliminate temporary
distortions as it excludes the volatile components of food and energy prices.
Figure 4 indicates that responses to inflation shocks measured by the PCE MoM generate
partially contrasting results. While there is a shift in the response sign, BTC and Gold
continue to react in the same direction. On the left graph, Bitcoin returns are observed to
react with the same intensity, but negatively after a positive inflation shock. In contrast,
the right graph shows that gold returns also respond negatively, although insignificantly, to
the inflation shock. Furthermore, it is noteworthy that convergence to the final cumulative
effect continues to occur after the tenth period.

13

Electronic copy available at: https://ssrn.com/abstract=4763347

 The discrepancy in the response sign may be attributed to differing criteria, such as
composition and disclosure dates, that can impact inflation surprises. Since PCE announce-
ments occur after CPI releases, every surprise in the CPI – positive or negative – is plausibly
incorporated in PCE expectations. This could explain why, as previously discussed, there is
a very low contemporaneous correlation between surprise inflation in the pairs CPI MoM-
PCE MoM and CPI YoY-PCE MoM.7 Although such adjustments may explain the negative
responses in BTC and Gold returns, it is imperative to note that the empirical validation
of this mechanism is beyond the scope of the present study. However, this observation is
suggested as a potential avenue for future research to deepen its understanding.

Figure 4. COIRF - CPI MoM vs. Core PCE MoM

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of Bitcoin
and gold to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2010:08 and 2023:01. The Ret prefix indicates that the series
was constructed from returns calculated in one-day windows around monthly Core PCE MoM
announcements. The horizontal axes show the number of months after the impulse.

4.2 Robustness checks

4.2.1 Alternative VAR lag order
Acknowledging that estimated impulse responses may exhibit bias due to an insufficient
number of lags (Enders, 2008, Lütkepohl, 2005), we perform a sensitivity analysis by rees-
timating the model using nine and twelve lags. Figure D1 in Appendix D illustrates the
findings, highlighting that the primary results remain consistent regardless of the lag length.
However, we note that with extended lag, the impact of impulses on the results intensifies.
7
As Figure A3 shows, correlations between CPI MoM (CPI YoY) and PCE MoM surprises are 0.12 (0.12)
in our sample. They are not just low, on average, but they also differ in time trends: while the CPI has shown
a sequence of positive surprises in 2021 onwards (following supply-chain imbalances brought by COVID-19
and, in 2022, the Russian invasion of Ukraine), the PCE has not shown such a trend.

14

Electronic copy available at: https://ssrn.com/abstract=4763347

4.2.2 Alternative identification #1: changing order of variables in the Choleski
decomposition
The model is identified based on the premise that movements in the rest of the economy are
less endogenous concerning the Bitcoin market. While this assumption is reasonable, it is
crucial to note that any recursive assumption can pose challenges, especially in the presence
of financial variables at a low frequency (Furlanetto et al., 2019). Furthermore, unless there
are special reasons for a recursive structure, the order of the variables used in the Choleski
decomposition of the white noise covariance matrix is arbitrary (Lütkepohl, 2005). Because
there is no trivial solution to this issue, we re-estimate the model with modifications to the
identification scheme.
Firstly, recognizing the inherent complexity of the relationship between inflation and
interest rates and the possibility that this connection is bidirectional and variable in different
economic contexts and market conditions (Alvarez et al., 2001), the time sequence of both
series was inverted. Then, it was decided to position these variables at the beginning of the
system, drawing on studies such as Roley and Sellon (1998), Bomfim (2003), Rigobon and
Sack (2004) and Farka (2009) that illustrate the endogenous nature of interactions between
the stock market, monetary policy, and inflation. Finally, Bitcoin was chosen to remain
the most endogenous variable in the system, aligning with research by Bouri et al. (2017a),
Li and Wang (2017), Al-Yahyaee et al. (2018), and Choi and Shin (2021). Therefore, the
Cholesky ordering of the VAR system is as follows: Dif U S 1Y , Surp Inf , Ret SP 500,
Dif V IX, Ret Gold, and Ret BT C.
Figure D2, available in Appendix D, confirms that the change in model identification
does not affect any of the baseline findings.

4.2.3 Alternative specification #2: adding a trend term

The baseline model comprises six variables plus the intercept. To incorporate more precise
information regarding the historical trajectory and potential patterns in the temporal evolu-
tion of the series, we include in this robustness check a trend term in the VAR specification.
Figure D3 in Appendix D illustrates that, although the intensity of impulse responses for
both assets is reduced, the results exhibit qualitative similarity overall when incorporating
a linear trend component into the model specification.

4.2.4 Changing sample period: before COVID-19

Although the COVID-19 pandemic is a unique opportunity to test Bitcoin’s inflation-hedging
properties, it is crucial to acknowledge potential influences from this atypical event.8 As
highlighted by Maneejuk et al. (2021), an asset’s ability to hedge against inflation may
be subject to the state of the economy, encompassing both stable and turbulent economic
8
Previous research has shown that the sensitivity of performance measures to the COVID-19 period was
higher for portfolios with cryptocurrency relative to otherwise identical portfolios (Colombo et al., 2021).

15

Electronic copy available at: https://ssrn.com/abstract=4763347

regimes. Taking into account this possibility, the model was re-estimated using data up to
December 2019.
Figure D4, available in Appendix D, underscores that the findings are partially preserved.
Despite an increased intensity in the impulse response, Bitcoin’s primary results are con-
firmed. However, excluding the pandemic episode revealed a slight negative reaction of gold
returns to inflation shocks. These findings align with empirical evidence from Phochanachan
et al. (2022), suggesting that Bitcoin’s reactions to inflation shocks are more significant in
stable scenarios, unlike gold, which demonstrates better performance in turbulent contexts.

4.2.5 Changing sample period: after the BTC structural break

Over the last decade, perceptions of cryptocurrencies, especially Bitcoin, have undergone sig-
nificant transformations, impacting the trading system and market dynamics. In the initial
phase of the examined period, Bitcoin trading volumes were minimal, and the cryptocur-
rency was not widely recognized as an investment option. Identifying the precise moment
when Bitcoin emerged as a viable investment alternative is challenging. However, for this
sensitivity test, the authors adopted Choi and Shin (2021) proposition of a structural break
in the Bitcoin market around 2013.
Figure D5 in Appendix D confirms that restricting the analysis to data from 2014 onward
has no discernible impact on gold prices. However, the positive cumulative effect on Bitcoin
prices, while still present, diminishes significantly. This outcome implies that the early years
of trading significantly influenced Bitcoin’s performance as a hedge against inflation.

4.2.6 Including observed inflation in the model

Including a variable representing realized inflation enhances our understanding of the for-
mation of expectations over time and their impact on economic behavior (Armantier et al.,
2011, Binder and Kamdar, 2022, Weber et al., 2022), offering a more thorough analysis of the
dynamic relationships between inflation and the price of assets like Bitcoin. Moreover, this
approach provides a more comprehensive assessment of market conditions by considering the
cumulative effects of inflation and structural changes in the economy.
Figure D6, available in Appendix D, illustrates the key empirical findings of this ex-
amination. Firstly, actual inflation increases over time in response to a positive inflation
shock. Secondly, Bitcoin returns experience a significant upswing following a positive shock
to realized inflation, aligning with empirical evidence that supports Bitcoin’s potential as
an effective hedge against inflation (Blau et al., 2021, Choi and Shin, 2021). Lastly, gold
returns display limited responsiveness to the realized inflation shock.

16

Electronic copy available at: https://ssrn.com/abstract=4763347

5 Concluding remarks
The recent consideration of Bitcoin as an investment option (Conlon et al., 2021), along with
the growing interest among academics and policymakers (Phochanachan et al., 2022), has
contributed to positioning the cryptocurrency as a potential inflation hedge. Commonly cited
justifications for this role include its limited supply and decentralized network, conferring
scarcity and resilience (Bouri et al., 2017a,b). However, like gold, existing theoretical and
empirical studies have not reached a consensus regarding its hedging ability (Choi and Shin,
2021). This lack of consensus may be attributed to methodological differences, varying time
horizons, and the specific characteristics of economies analyzed in these studies.
This article investigates the dynamic response of Bitcoin returns to inflation shocks,
considering different inflation measures, changes in the sampling period, and various model
specifications. The results suggest that Bitcoin can be a valuable hedge against unexpected
fluctuations in inflation. However, the efficacy of Bitcoin as an inflation hedge appears
contingent upon the selected price index (evidence is found on CPI MoM and CPI YoY,
but not for the Core PCE MoM) and the analyzed sampling period. Regarding the latter,
the inflation-hedging properties of Bitcoin diminish significantly when we exclude the initial
sampling period (“early days” of Bitcoin), suggesting that mainstream adoption may be
driving BTC returns to be closer to returns of other risky assets (like stocks). At the end
of the day, bitcoin does not seem to be as reliable as Gold in protecting portfolios against
unexpected, positive inflation shocks.
Our paper contributes to the still incipient literature on the role of Bitcoin as an inflation
hedge (see, e.g., Blau et al., 2021, Choi and Shin, 2021, Conlon et al., 2021, Matkovskyy and
Jalan, 2020, Phochanachan et al., 2022, Smales, 2021) by identifying inflation shocks based on
the unexpected component of CPI and PCE releases – the difference between the actual data
and analysts’ expectations. By doing that, our innovations (or shocks) consider all observed
and unobserved factors that market agents consider in their projections of inflation, and not
only a few covariates included in typical VAR systems. As Nakamura and Steinsson (2017)
indicates, any factors left out as covariates in the regression will be part of the “shock”,
even if they are anticipated by market agents. Therefore, we believe that our identification
strategy provides a cleaner projection of the unexpected portion of inflation announcements
(i.e., inflation shocks). Furthermore, this research carries direct implications for managers9 ,
investors, and monetary authorities, contributing to the growing but still underexplored
literature about the inflation-hedging properties of Bitcoin.
While this study has unveiled numerous empirical findings, it is imperative to acknowl-
edge certain caveats that offer valuable insights for future research. Firstly, the dynamic
nature of the cryptocurrency market requires caution in interpreting results. Specifically,
the employed model does not account for the potential impacts of regulatory changes or the
proliferation of new cryptocurrencies on Bitcoin prices. Secondly, in comparison to prior
9
Not only asset managers, but also corporate managers since many corporations acquired bitcoin in recent
years diverse reasons (Gimenes et al., 2023).

17

Electronic copy available at: https://ssrn.com/abstract=4763347

analyses of other inflation-hedge assets, such as gold, the sampling period is confined to the
cryptocurrency market’s early stages. Consequently, pivotal aspects like trading volume or
liquidity may undergo abrupt changes in the future, posing challenges to the drawn con-
clusions. Thirdly, the study’s methodology relies on a linear model, potentially overlooking
nonlinear relationships and structural shifts in the analyzed variables. Lastly, it is crucial to
highlight that returns are calculated on a daily basis, which may lead to the loss of pertinent
intraday information and impact analysis accuracy. Therefore, future studies should explore
the feasibility of capturing these variations at a high frequency.

References

Al-Yahyaee, K. H., Mensi, W., and Yoon, S.-M. (2018). Efficiency, multifractality, and the
long-memory property of the bitcoin market: A comparative analysis with stock, currency,
and gold markets. Fin. Res. Lett., 27:228–234.

Alvarez, F., Lucas, Jr, R. E., and Weber, W. E. (2001). Interest rates and inflation. Am.
Econ. Rev., 91(2):219–225.

Armantier, O., Bruine de Bruin, W., Topa, G., van der Klaauw, W., and Zafar, B. (2011).
Inflation expectations and behavior: Do survey respondents act on their beliefs? SSRN
Electron. J.

Baum, C. F. (2013). Var, svar and vecm models. Boston College, Spring.

Beaudry, P. and Portier, F. (2006). Stock prices, news, and economic fluctuations. Am.
Econ. Rev., 96(4):1293–1307.

Binder, C. and Kamdar, R. (2022). Expected and realized inflation in historical perspective.
J. Econ. Perspect., 36(3):131–155.

Blau, B. M., Griffith, T. G., and Whitby, R. J. (2021). Inflation and bitcoin: A descriptive
time-series analysis. Econ. Lett., 203(109848):109848.

Bomfim, A. (2003). Pre-announcement effects, news effects, and volatility: Monetary policy
and the stock market. J. Bank. Financ., 27(1):133–151.

Bouri, E., Gupta, R., Tiwari, A. K., and Roubaud, D. (2017a). Does bitcoin hedge global
uncertainty? evidence from wavelet-based quantile-in-quantile regressions. Fin. Res. Lett.,
23:87–95.

18

Electronic copy available at: https://ssrn.com/abstract=4763347

Bouri, E., Molnár, P., Azzi, G., Roubaud, D., and Hagfors, L. I. (2017b). On the hedge
and safe haven properties of bitcoin: Is it really more than a diversifier? Fin. Res. Lett.,
20:192–198.

Bureau of Labor Statistics (BLS) (2011). Consumer price in-

dex: First quarter 2011. https://www.bls.gov/opub/btn/archive/
differences-between-the-consumer-price-index-and-the-personal-consumption-expenditures
pdf. Accessed on February 1, 2024.

Choi, S. and Shin, J. (2021). Bitcoin: An inflation hedge but not a safe haven. Fin. Res.
Lett., 46(102379):102379.

Christiano, L. J., Eichenbaum, M., and Evans, C. L. (2005). Nominal rigidities and the
dynamic effects of a shock to monetary policy. J. Polit. Econ., 113(1):1–45.

Cochrane, J. H. and Piazzesi, M. (2002). The fed and interest rates—a high-frequency
identification. Am. Econ. Rev., 92(2):90–95.

Coibion, O. (2012). Are the effects of monetary policy shocks big or small? Am. Econ. J.
Macroecon., 4(2):1–32.

Colombo, J., Cruz, F., Paese, L., and Cortes, R. (2021). The diversification benefits of
cryptocurrencies in multi-asset portfolios: cross-country evidence. Available at SSRN
3776260.

Colombo, J. A. and Yarovaya, L. (2024). Are crypto and non-crypto investors alike? evidence
from a comprehensive survey in brazil. Technology in Society, page 102468.

Conlon, T., Corbet, S., and McGee, R. J. (2021). Inflation and cryptocurrencies revisited:
A time-scale analysis. Econ. Lett., 206(109996):109996.

Corbet, S., Larkin, C., Lucey, B. M., Meegan, A., and Yarovaya, L. (2020). The impact of
macroeconomic news on bitcoin returns. Eur. J. Fin., 26(14):1396–1416.

Enders, W. (2008). Applied econometric time series. John Wiley & Sons, New York, USA.

Farias, H. P. (2008). Função resposta a impulso e decomposição da variância do erro de

previsão aplicados às principais bolsas de valores. Dissertação de mestrado, Universidade
Federal de Lavras, Lavras.

Farka, M. (2009). The effect of monetary policy shocks on stock prices accounting for
endogeneity and omitted variable biases. Rev. Financ. Econ., 18(1):47–55.

Fleming, M. J. and Remolona, E. M. (1999). Price formation and liquidity in the us treasury
market: The response to public information. J. Finance, 54(5):1901–1915.

19

Electronic copy available at: https://ssrn.com/abstract=4763347

Furlanetto, F., Ravazzolo, F., and Sarferaz, S. (2019). Identification of financial factors in
economic fluctuations. Econ. J. (London), 129(617):311–337.

Gimenes, A. D., Colombo, J. A., and Yousaf, I. (2023). Store of value or speculative in-
vestment? market reaction to corporate announcements of cryptocurrency acquisition.
Financial Innovation, 9(1):123.

Hanson, S. G. and Stein, J. C. (2015). Monetary policy and long-term real rates. J. Financ.
Econ., 115(3):429–448.

Jordà, Ò., Schularick, M., and Taylor, A. M. (2017). Macrofinancial history and the new
business cycle facts. NBER Macroecon. Annu., 31(1):213–263.

Kurov, A., Olson, E., and Zaynutdinova, G. R. (2022). When does the fed care about stock
prices? J. Bank. Financ., 142(106556):106556.

Kuttner, K. N. (2001). Monetary policy surprises and interest rates: Evidence from the fed
funds futures market. J. Monet. Econ., 47(3):523–544.

Li, X. and Wang, C. A. (2017). The technology and economic determinants of cryptocurrency
exchange rates: The case of bitcoin. Decis. Support Syst., 95:49–60.

Ljung, G. M. and Box, G. E. P. (1978). On a measure of lack of fit in time series models.
Biometrika, 65(2):297–303.

Lowenkron, A. and Garcia, M. G. P. (2007). Monetary policy credibility and inflation

risk premium: a model with application to Brazilian data. Textos para discussão 543,
Department of Economics PUC-Rio (Brazil).

Lütkepohl, H. (2005). New introduction to multiple time series analysis. Springer, Berlin,
Germany.

Lütkepohl, H. (2011). Vector Autoregressive Models. Economics Working Papers

ECO2011/30, European University Institute.

Maneejuk, P., Pirabun, N., Singjai, S., and Yamaka, W. (2021). Currency hedging strategies
using histogram-valued data: Bivariate markov switching GARCH models. Mathematics,
9(21):2773.

Matkovskyy, R. and Jalan, A. (2020). Bitcoin vs inflation: Can bitcoin be a macro hedge?
evidence from a quantile-on-quantile model. SSRN Electron. J.

Morettin, P. A. and de Castro Toloi, C. M. (2006). Análise de séries temporais. Blucher,

São Paulo, Brasil.

Nakamura, E. and Steinsson, J. (2017). Identification in Macroeconomics. NBER Working

Papers 23968, National Bureau of Economic Research, Inc.

20

Electronic copy available at: https://ssrn.com/abstract=4763347

Phochanachan, P., Pirabun, N., Leurcharusmee, S., and Yamaka, W. (2022). Do bitcoin and
traditional financial assets act as an inflation hedge during stable and turbulent markets?
evidence from high cryptocurrency adoption countries. Axioms, 11(7):339.

Pyo, S. and Lee, J. (2020). Do FOMC and macroeconomic announcements affect bitcoin
prices? Fin. Res. Lett., 37(101386):101386.

Rigobon, R. and Sack, B. (2003). Measuring the reaction of monetary policy to the stock
market. Q. J. Econ., 118(2):639–669.

Rigobon, R. and Sack, B. (2004). The impact of monetary policy on asset prices. J. Monet.
Econ., 51(8):1553–1575.

Roley, V. V. and Sellon, G. H. (1998). Market reaction to monetary policy nonannounce-

ments. Research Working Paper 98-06, Federal Reserve Bank of Kansas City.

Romer, C. D. and Romer, D. H. (2004). A new measure of monetary shocks: Derivation and
implications. Am. Econ. Rev., 94(4):1055–1084.

Schwert, G. W. (1989). Tests for unit roots: A monte carlo investigation. J. Bus. Econ.
Stat., 7(2):147–159.

Sims, C. A. (1980). Macroeconomics and reality. Econometrica, 48(1):1–48.

Smales, L. A. (2021). Cryptocurrency as an alternative inflation hedge? SSRN Electron. J.

Stock, J. H. and Watson, M. W. (2001). Vector autoregressions. J. Econ. Perspect.,

15(4):101–115.

Tarbert, H. (1996). Is commercial property a hedge against inflation? a cointegration ap-

proach. Journal of Property Finance, 7(1):77–98.

Weber, M., Gorodnichenko, Y., and Coibion, O. (2022). The expected, perceived, and
realized inflation of u.s. households before and during the COVID19 pandemic. Technical
Report 29640, National Bureau of Economic Research, Cambridge, MA.

World Economic Forum (2023). Global risks report 2023. https://www.weforum.org/

publications/global-risks-report-2023/in-full/. Accessed on February 1, 2023.

21

Electronic copy available at: https://ssrn.com/abstract=4763347

A Appendix
A Descriptive statistics - Surprise inflation measures

Figure A1. Frequency distribution of surprise inflation

Note: This figure shows the frequency distribution graphs of the unexpected component of
inflation calculated in one-day windows around monthly CPI MoM (top left graph), CPI YoY
(top right graph), and Core PCE MoM (bottom graph) announcements. The sample period
covers from 2010:08 to 2023:01. The horizontal axes represent different intervals/values of
surprise inflation.

22

Electronic copy available at: https://ssrn.com/abstract=4763347

Figure A2. CPI MoM, CPI YoY, and PCE MoM - Actual, Forecast, and Surprise compar-
isons
(a) CPI MoM – Actual, Forecast, and Surprise

1.50
Observed CPI (%, MoM)
Forecasted CPI (%, MoM)
1.00

0.50

Avg. = 0.21
0.00

-0.50

-1.00

1.50
Surprise CPI (p.p., MoM)

1.00

0.50

0.00
Avg. = -0.00

-0.50

-1.00
0 011 3 4 5 6 8 9 0 1 3
201 2, 2 201 201 201 201 201 201 202 202 202
13, 1 10, 12, 11, 09, 08, 10, 08, 07, 06,
Aug Nov Feb May Aug Nov Feb May Aug Nov Feb

(b) PCE MoM – Actual, Forecast, and Surprise

1.00
Observed PCE (%, MoM)
Forecasted PCE (%, MoM)

0.50

Avg. = 0.16
0.00

-0.50

1.00
Surprise PCE (p.p., MoM)

0.50

0.00

Avg. = -0.02

-0.50
0 011 3 4 5 6 8 9 0 1 3
201 2, 2 201 201 201 201 201 201 202 202 202
13, 1 10, 12, 11, 09, 08, 10, 08, 07, 06,
Aug Nov Feb May Aug Nov Feb May Aug Nov Feb

23

Electronic copy available at: https://ssrn.com/abstract=4763347

Figure A2. CPI MoM, CPI YoY, and PCE MoM - Actual, Forecast, and Surprise compar-
isons (cont.)
(a) CPI YoY – Actual, Forecast, and Surprise

10.00
Observed CPI (%, YoY)
Forecasted CPI (%, YoY)
8.00

6.00

4.00

2.00 Avg. = 2.46

0.00

Surprise CPI (p.p., YoY)

0.50

0.00
Avg. = 0.02

-0.50
0 011 3 4 5 6 8 9 0 1 3
201 2, 2 201 201 201 201 201 201 202 202 202
13, 1 10, 12, 11, 09, 08, 10, 08, 07, 06,
Aug Nov Feb May Aug Nov Feb May Aug Nov Feb

Note: Subfigures (a), (b), and (c) show the observed and the forecasted U.S. Consumer Price
Index (CPI, %, Year-over-Year) for each release of the U.S. Bureau of Labor Statistics (BLS).
Actual data is seasonally adjusted for MoM comparisons (not seasonally adjusted for YoY
comparisons) and comes from the BLS, and forecasts are obtained at Investing.com. The lower
part of the figure shows the time series of the Surprise component (difference between the
observed, actual inflation, and the analysts’ forecast). Covered CPI (PCE) announcements
range from August 13, 2010 (August 30, 2010), to January 12, 2023 (Jan 27, 2023).

24

Electronic copy available at: https://ssrn.com/abstract=4763347

Figure A3. Time series and correlation of Surprises – CPI MoM, CPI YoY, and PCE MoM

1.00
Surprise CPI (p.p., MoM)
Surprise PCE (p.p., MoM)
Surprise CPI (p.p., YoY)

0.50

0.00

Corr. [CPI_MoM, PCE_MoM] = 0.12

Corr. [CPI_YoY, PCE_MoM] = 0.02
-0.50 Corr. [CPI_MoM, CPI_YoY] = 0.63

-1.00
0 1 3 4 5 6 8 9 0 1 3
201 201 201 201 201 201 201 201 202 202 202
Aug Nov Feb May Aug Nov Feb May Aug Nov Feb

Note: The upper part of this figure shows the observed and forecasted U.S. Consumer Price
Index (CPI, %, Year-over-Year) for each release of the U.S. Bureau of Labor Statistics (BLS).
Actual data is seasonally adjusted for MoM comparisons (not seasonally adjusted for YoY
comparisons) and comes from the BLS, and forecasts are obtained at Investing.com. The lower
part of the figure shows the time series of the Surprise component (difference between the
observed, actual inflation, and the analysts’ forecast). Covered CPI (PCE) announcements
range from August 13, 2010 (August 30, 2010), to January 12, 2023 (Jan 27, 2023).

25

Electronic copy available at: https://ssrn.com/abstract=4763347

Figure A4. Cross-correlation between surprise inflation measures: CPI MoM, PCE MoM,
and CPI YoY
(a) CPI MoM x PCE MoM (b) CPI YoY x PCE MoM
Cross-correlation Cross-correlation
Surprise CPI (p.p., MoM) -> Surprise PCE (p.p., MoM) Surprise CPI (p.p., YoY) -> Surprise PCE (p.p., MoM)
1.00 1.00 1.00 1.00

0.50 0.50 0.50 0.50

Cross-correlations of Shock_CPI_MoM and Shock_PCE_MoM

Cross-correlations of Shock_CPI_YoY and Shock_PCE_MoM

0.00 0.00 0.00 0.00

-0.50 -0.50 -0.50 -0.50

-1.00 -1.00 -1.00 -1.00

-10 -5 0 5 10 -10 -5 0 5 10
Lag Lag

(c) CPI MoM x CPI YoY

Cross-correlation
Surprise CPI (p.p., MoM) -> Surprise CPI (p.p., YoY)
1.00 1.00

0.50 0.50
Cross-correlations of Shock_CPI_MoM and Shock_CPI_YoY

0.00 0.00

-0.50 -0.50

-1.00 -1.00
-10 -5 0 5 10
Lag

Note: This Figure shows the cross-correlation among each pair of computed inflation surprise
(difference between the observed, actual inflation, and the analysts’ forecast): CPI MoM-PCE
MoM (Panel A), CPI YoY-PCE MoM (Panel B), and CPI MoM-CPI YoY (Panel C). The Y-
axis shows the linear correlation coefficient that ranges from -1 to +1. The X-axis presents the
lags and leads that are symmetrical and range from one to ten. The red vertical line emphasizes
the contemporaneous linear correlation coefficient (lead/lag equals zero). Covered CPI (PCE)
announcements range from August 13, 2010 (August 30, 2010), to January 12, 2023 (Jan 27,
2023).

26

Electronic copy available at: https://ssrn.com/abstract=4763347

B Model Checking

Table B1. Augmented Dickey-Fuller unit root test

Variable Order Lags Test stat. p-value Assessment

Panel A: Model with drift
Ret SP500 I(0) 4 -5.868 * < 0.01 Reject H0
Dif VIX I(0) 0 -13.547 * < 0.01 Reject H0
Surp Inf I(0) 0 -10.074 * < 0.01 Reject H0
Dif US 1Y I(0) 10 -4.508 * < 0.01 Reject H0
Ret Gold I(0) 0 -10.243 * < 0.01 Reject H0
Ret BTC I(0) 0 -11.729 * < 0.01 Reject H0
Panel B: Model without drift or trend
Ret SP500 I(0) 4 -5.894 * < 0.01 Reject H0
Dif VIX I(0) 0 -13.527 * < 0.01 Reject H0
Surp Inf I(0) 0 -10.108 * < 0.01 Reject H0
Dif US 1Y I(0) 10 -4.488 * < 0.01 Reject H0
Ret Gold I(0) 0 -10.075 * < 0.01 Reject H0
Ret BTC I(0) 0 -11.745 * < 0.01 Reject H0
Panel C: Model with drift and trend
Ret SP500 I(0) 4 -5.836 * < 0.01 Reject H0
Dif VIX I(0) 0 -13.554 * < 0.01 Reject H0
Surp Inf I(0) 0 -10.345 * < 0.01 Reject H0
Dif US 1Y I(0) 10 -4.872 * < 0.01 Reject H0
Ret Gold I(0) 0 -10.898 * < 0.01 Reject H0
Ret BTC I(0) 0 -11.688 * < 0.01 Reject H0
Panel D: Model with drift and linear and quadratic trends
Ret SP500 I(0) 4 -5.895 * < 0.01 Reject H0
Dif VIX I(0) 0 -13.523 * < 0.01 Reject H0
Surp Inf I(0) 0 -10.669 * < 0.01 Reject H0
Dif US 1Y I(0) 10 -5.308 * < 0.01 Reject H0
Ret Gold I(0) 0 -10.876 * < 0.01 Reject H0
Ret BTC I(0) 0 -11.801 * < 0.01 Reject H0

Note: This table shows the results of the Augmented Dickey-Fuller (ADF) test for the Model
with different specifications. The sample period covers from 2010:08 to 2023:01. The Ret (Dif)
prefix indicates that the series was constructed from returns (differences) calculated in one-day
windows around monthly CPI MoM announcements. The maximum number of lags considered
for implementing the test is 13 and is defined by the Schwert criterion (1989). *, **, and ***
denote statistical significance at 1, 5, and 10%, respectively, and lead to the rejection of the null
hypothesis indicating the non-existence of a unit root. The Lags column reports the optimal
number of lags following the Akaike information criterion (AIC).

27

Electronic copy available at: https://ssrn.com/abstract=4763347

 Table B2. Portmanteau residual autocorrelation test

Model Test stat. Critical value (5%) p-value Degrees of freedom Assessment
VAR(6) 264.8 290.0 0.277 252 Accept H0
VAR(7) 265.5 251.3 0.012 216 Reject H0
VAR(8) 265.5 212.3 0.000 180 Reject H0

Note: This table shows the results of the Ljung and Box test (1978) for the Model with different
VAR orders. The sample period covers from 2010:08 to 2023:01. The maximum number of
lags considered for implementing the test is 13 and is defined by the Schwert criterion (1989).
QLB statistic values lower than the critical values lead to the acceptance of the null hypothesis
and indicate that the residual autocorrelation up to the maximum number of lags is zero,
considering a 95% confidence interval.

Table B3. Model stability

Lag Ret SP500 Dif VIX Surp Inf Dif US 1Y Ret Gold Ret BTC
1 0.120 0.233 0.233 0.235 0.077 0.077
2 0.207 0.185 0.185 0.152 0.068 0.008
3 0.631 0.157 0.157 0.158 0.158 0.032
4 0.242 0.223 0.223 0.206 0.206 0.169
5 0.197 0.197 0.273 0.120 0.107 0.034
6 0.271 0.334 0.076 0.076 0.011 0.093

Note: This table shows the results of the stability diagnosis test applied to the VAR(6) model.
The sample period covers from 2010:08 to 2023:01. The Ret (Dif) prefix indicates that the
series was constructed from returns (differences) calculated in one-day windows around monthly
CPI MoM announcements. The values reported correspond to the eigenvalues (in modulus)
associated with each model lag. Values lower than 1 mean that they are inside the unit circle,
which highlights the stability of the model.

28

Electronic copy available at: https://ssrn.com/abstract=4763347

C Full results - COIRFs of the baseline model

29

Electronic copy available at: https://ssrn.com/abstract=4763347

 Figure C1. COIRF Panel - Baseline model

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of the six
variables and their 95% confidence bands for the sample period between 2010:08 and 2023:01.
The Ret (Dif) prefix indicates that the series was constructed from returns (differences) cal-
culated in one-day windows around monthly CPI MoM announcements. The units of the
horizontal axes are months.

30

Electronic copy available at: https://ssrn.com/abstract=4763347

D Robustness Tests

Figure D1. RT 01 - Alternative VAR order

VAR(9)

VAR(12)

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of Bitcoin
and gold to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2010:08 and 2023:01 for the Model with alternative lags. The Ret
prefix indicates that the series was constructed from returns calculated in one-day windows
around monthly CPI MoM announcements. The units of the horizontal axes are a month.

31

Electronic copy available at: https://ssrn.com/abstract=4763347

 Figure D2. RT 02 - Alternative identification

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of Bitcoin
and gold to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2010:08 and 2023:01 for the Model with the following temporal
ordering: Dif U S 1Y , Surp Inf , Ret SP 500, Dif V IX, Ret Gold and Ret BT C. The Ret
(Dif) prefix indicates that the series was constructed from returns (differences) calculated in
one-day windows around monthly CPI MoM announcements. The units of the horizontal axes
are a month.

Figure D3. RT 03 - Alternative specification

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of Bitcoin
and gold to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2010:08 and 2023:01 for the Model with the addition of a linear
trend component in the specification. The Ret prefix indicates that the series was constructed
from returns calculated in one-day windows around monthly CPI MoM announcements. The
units of the horizontal axes are a month.

32

Electronic copy available at: https://ssrn.com/abstract=4763347

 Figure D4. RT 04 - Before COVID-19

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of Bitcoin
and gold to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2010:08 and 2020:01 (114 observations). The Ret prefix indicates
that the series was constructed from returns calculated in one-day windows around monthly
CPI MoM announcements. The units of the horizontal axes are a month.

Figure D5. RT 05 - After BTC structural break

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) of Bitcoin
and gold to the one-standard-deviation shock in inflation and their 95% confidence bands for
the sample period between 2014:01 and 2023:01 (109 observations). The Ret prefix indicates
that the series was constructed from returns calculated in one-day windows around monthly
CPI MoM announcements. The units of the horizontal axes are a month.

33

Electronic copy available at: https://ssrn.com/abstract=4763347

 Figure D6. RT 06 - Including actual inflation

Note: This figure shows cumulative orthogonal impulse response functions (COIRFs) and their
95% confidence bands for the sample period between 2010:08 and 2023:01 for the Model with
the inclusion of the actual inflation series (Act Inf ). The Ret (Dif) prefix indicates that the
series was constructed from returns (differences) calculated in one-day windows around monthly
CPI MoM announcements. The units of the vertical axes correspond to the intensity of the
series’ response to the one-standard-deviation shock in inflation (actual inflation). The units
of the horizontal axes are a month.

34

Electronic copy available at: https://ssrn.com/abstract=4763347

Table D1. Summary statistics: Full period, Before COVID-19 and After BTC structural
break

Variable Mean p50 SD Asymmetry Kurtosis Min Max Obs.

Panel A: Full period (2010m8 to 2023m1)
Ret SP500 -0.000 0.001 0.012 -0.680 6.758 -0.050 0.054 150
Dif VIX -0.123 -0.175 1.900 1.794 9.214 -5.710 11.090 150
Surp Inf -0.000 0.000 0.136 0.769 2.588 -0.400 0.600 150
Dif US 1Y 0.003 0.000 0.045 2.607 13.714 -0.160 0.230 150
Ret Gold 0.001 0.001 0.009 -0.210 1.341 -0.029 0.028 150
Ret BTC -0.004 0.000 0.091 -5.094 49.918 -0.849 0.288 150
Panel B: Before COVID-19 (2010m8 to 2020m1)
Ret SP500 0.001 0.001 0.009 -1.467 6.692 -0.046 0.020 114
Dif VIX -0.167 -0.140 1.703 2.134 17.015 -5.710 11.090 114
Surp Inf -0.021 0.000 0.114 0.024 1.200 -0.400 0.300 114
Dif US 1Y -0.000 0.000 0.018 1.701 11.132 -0.050 0.110 114
Ret Gold 0.001 0.000 0.009 -0.190 1.667 -0.029 0.028 114
Ret BTC -0.006 0.000 0.101 -4.892 43.346 -0.849 0.288 114
Panel C: After BTC structural break (2014m1 to 2023m1)
Ret SP500 0.000 0.001 0.012 -0.441 6.968 -0.050 0.054 109
Dif VIX -0.230 -0.300 1.757 0.876 3.839 -5.710 6.600 109
Surp Inf 0.006 0.000 0.138 0.925 3.378 -0.400 0.600 109
Dif US 1Y 0.005 0.000 0.053 2.199 9.413 -0.160 0.230 109
Ret Gold 0.002 0.003 0.009 -0.289 1.938 -0.029 0.028 109
Ret BTC -0.010 -0.000 0.095 -6.607 58.196 -0.849 0.104 109

Note: This table shows statistics for all series covering different sampling periods. The Ret
(Dif) prefix indicates that the series was derived from returns (differences) calculated within
one-day windows around monthly CPI MoM announcements.

35

Electronic copy available at: https://ssrn.com/abstract=4763347