Learning and Predictability via Technical Analysis: Evidence from

Bitcoin and Stocks with Hard-to-Value Fundamentals

Andrew Detzel
University of Denver

Hong Liu
Washington University in St. Louis

Jack Strauss
University of Denver

Guofu Zhou
Washington University in St. Louis

Yingzi Zhu
Tsinghua University

First draft: December 5, 2017

Current version: January 29, 2020

Authors emails: [email protected], [email protected], [email protected], [email protected], and

[email protected], respectively. We are grateful to helpful comments from William Brock,
Aaron Burt, Conrad Ciccotello, Ricardo Correa, Stephanie Curcuru, Mathieu Fournier (Discussant),
Campbell Harvey, Markus Ibert, Mohammad R. Jahan-Parvar, Raymond Kan, Tong Li (Discussant), Edith
Liu, Sai Ma, Bill McDonald, Friederike Niepmann, Peter Scholz (Discussant), Stephan Siegel, and Leifu
Zhang, along with seminar participants at the Federal Reserve Board, Beijing University, Colorado State
University, Washington University in St. Louis, Shanghai University of Finance and Economics, Renmin
University of China, and conference participants at 2018 Conference on Financial Predictability and Big
Data, 2018 China Finance Review International Conference, and 2018 EUROFIDAI December Paris
Meetings. Zhu thanks financial support by National Natural Science Foundation of China (# 71572091) for
her work on the project. We are especially grateful to Bing Han (Editor) and an anonymous referee for very
insightful and constructive comments that helped improve the paper substantially.

Electronic copy available at: https://ssrn.com/abstract=3115846

 Learning and Predictability via Technical Analysis: Evidence from
Bitcoin and Stocks with Hard-to-Value Fundamentals
What predicts returns on assets with “hard-to-value" fundamentals, such as Bitcoin and
stocks in new industries? We propose an equilibrium model that shows how rational learning can
generate return predictability through technical analysis. We document that ratios of prices to their
moving averages forecast daily Bitcoin returns in- and out-of-sample. Trading strategies based on
these ratios generate an economically significant alpha and Sharpe ratio gains relative to a buy-and-
hold position. Similar results hold for small-cap, young-firm, and low-analyst-coverage stocks as
well as NASDAQ stocks during the dotcom era.

JEL classification: G11, G12, G14

Keywords: Bitcoin, cryptocurrency, technical analysis, learning, return predictability

Electronic copy available at: https://ssrn.com/abstract=3115846

 Cryptocurrencies’ fundamental source of intrinsic value remains unclear. Market observers disagree

about their ability to serve as a currency and their currency status faces significant regulatory risk. Moreover,

unlike cash flows from more typical financial assets such as stocks and bonds, cryptocurrencies’ fundamentals

have few, if any, publicly available predictive signals, such as analyst coverage and accounting statements.

We refer to fundamentals with these characteristics of uncertainty, opacity, disagreement, and lack of

predictive information as “hard-to-value". In this paper, we theoretically and empirically examine the asset-

pricing implications of having such fundamentals. While cryptocurrencies are an ideal setting to investigate

this asset-pricing property, this property is more general as the fundamentals of most assets are hard-to-value

to varying degrees. For example, fundamentals of young small-cap stocks in new industries are harder to

value than those of large-cap stocks in established industries.

We propose a continuous-time equilibrium model in which two rational and risk-averse investors

costlessly trade a risky asset with hard-to-value fundamentals. This asset produces a stream of benefits called

a “convenience yield" that grows at an unobserved and stochastically evolving rate. Investors have different

priors and, aside from the convenience yield itself, observe no other signal about the yield’s latent growth

rate. The risky asset can be interpreted as a cryptocurrency where the convenience yield represents the flow

of benefits from usage as a medium of exchange or another asset such as a stock whose dividends or earnings

are hard-to-value. In the process of Bayesian learning, investors update their beliefs about the growth rate in

the direction of shocks to the convenience yield. However, the initial value of the growth rate is uncertain and

these shocks are only imperfectly correlated with unobservable shocks to this rate, causing investors to only

gradually move away from their priors when updating beliefs—and consequently valuations—resulting in

price drift. Specifically, returns are predictable by ratios of prices to their moving averages (MAs), which

summarize the beliefs of investors about the expected convenience yield growth rate. Moreover, because of

this predictability, it is optimal for investors to use the price-to-MA ratios in trading. As far as we know, this

is the first fully rational general equilibrium model with endogenous use of technical analysis. 1

1
Treynor and Ferguson (1985), Brown and Jennings (1989), Cespa and Vives (2012), Brock et al. (1992),
Hong and Stein (1999), Lo et al. (2000), Chiarella et al. (2006), Edmans et al. (2015), and Han et al. (2016),
among others, show that past prices can predict returns and trading based on technical indicators,
3

Electronic copy available at: https://ssrn.com/abstract=3115846

 Our model contributes to several strands of literature. It provides a fully rational and endogenous
justification for technical analysis and time-series return predictability by past prices, which most
practitioners and prior studies justify with irrational forces, such as sentiment, overconfidence, or under-
reaction (e.g., Barberis et al. ,1998; Daniel et al., 1998; Hong and Stein, 1999). 2 Moreover, prior
models with technical traders assume that a subset of investors use exogenously given technical
trading rules (e.g., Hong and Stein, 1999; Han et al., 2016). Our model also proposes a new
mechanism relative to the few rational models that generate return predictability by past prices. For
example, in prior rational expectations equilibrium models with learning, price drift can arise, but
only given higher-order disagreement between traders (e.g., Banerjee et al., 2009). Without this
disagreement, agents infer each other’s private signals immediately via the price, precluding a gradual
drift toward the fundamental value. In our model, drift does not require disagreement, only Bayesian
learning and the hard-to-value property of fundamentals. Johnson (2002) also generates price drift for
stocks with time-varying dividend growth rates, which are analogous to our convenience yield.
However, Johnson does not consider general equilibrium effects, learning, or endogenous technical
trading rules.
In the empirical portion of the paper, we investigate whether the predictability of returns by price-to-

MA ratios holds for Bitcoin and several equity portfolios with plausibly hard-to-value fundamentals. We find

that daily Bitcoin returns are predictable in- and out-of-sample by ratios of prices to their 1- to 20-week MAs.

Consistent with our model, this predictability strengthens when uncertainty decreases as investors learn about

the dynamics of the latent growth of the convenience yield. Indeed, we find a negative interaction between

the price-to-MA ratio and conditional return variance, a proxy for uncertainty, in return-forecasting

regressions. To assess the economic significance of this Bitcoin-return predictability to investors, we form a

trading strategy that goes long Bitcoin when the price is above the MA, and long cash otherwise. We find that

these trading strategies significantly outperform the buy-and-hold benchmark, producing large alphas and

increasing Sharpe ratios by 0.2 to 0.6. These results are similar across both halves of the sample. The MA

especially the moving averages of prices, can be profitable in the stock market. Schwager (1989) and Lo
and Hasanhodzic (2009) further provide insightful comments about the effectiveness of technical strategies
from top practitioners.
2
Perhaps the most widely used investments textbook, Bodie et al. (2014), places technical analysis in a
chapter on “behavioral finance”.
4

Electronic copy available at: https://ssrn.com/abstract=3115846

strategies also outperform the buy-and-hold benchmark when applied to two other cryptocurrencies, Ripple

and Ethereum, Bitcoin’s two largest competitors. Indeed, we find a negative interaction between the price-

to-MA ratio and conditional return variance, a proxy for uncertainty, in return-forecasting regressions.

To assess the economic significance of this Bitcoin-return predictability to investors, we form a trading

strategy that goes long Bitcoin when the price is above the MA, and long cash otherwise. We find that

these trading strategies significantly outperform the buy-and-hold benchmark, producing large alphas

and increasing Sharpe ratios by 0.2 to 0.6. These results are similar across both halves of the sample.

The MA strategies also outperform the buy-and-hold benchmark when applied to two other

cryptocurrencies, Ripple and Ethereum, Bitcoin’s two largest competitors.

Next, we evaluate whether returns on the NASDAQ portfolio are predictable by price-to-MA
ratios during five- and ten-year windows (1998–2002 and 1996–2005, respectively) that includes the
dot.com boom-and-bust of the early 2000’s. In this period, many emerging technologies associated with
the Internet introduced fundamentals that, at the time, were considered difficult to value. We show that
our MA trading strategies applied to the NASDAQ generate significant alpha and Sharpe ratio gains (of
0.2 to 0.5) relative to the buy-and-hold benchmark in this time period. Moreover, the gains of the MA
strategies steadily decline in the years following this period as fundamentals presumably became easier
to value. We also apply our MA strategies to portfolios formed on widely used proxies for information
availability: size, age, and analyst coverage. Consistent with our model, we find that over the 1963 to
2018 time period, the price-to-MA ratios positively and significantly forecast returns on small-cap and
young-stock portfolios, with negative and insignificant predictability for large-cap and old-stock
portfolios. Moreover, in the 1985 to 2018 time-period during which analyst forecasts are available, we
find that return predictability by the price-to-MA ratios decreases with both size and analyst coverage.

Our model also suggests that trading arises from differences in the MAs across investors.

Consistent with this implication, we show that proxies for disagreement across MA horizons and total

turnover implied by the various MA strategies are significantly and positively associated with Bitcoin

trading volume.

Overall, consistent with our model, the results in this paper demonstrate that for Bitcoin and

stocks with hard-to-value fundamentals, price drift exists and price-to-moving average ratios predict

Electronic copy available at: https://ssrn.com/abstract=3115846

returns.

The rest of the paper is organized as follows. Section II discusses the model and related literature.

Section III describes the data. Section IV reports empirical results, and Section V concludes.

2| THE MODEL AND RELATTED LITERAURE

2.1 The Model

In this section, we present a rational equilibrium asset-pricing model that examines the implications of
having hard-to-value fundamentals. In the model, investors continuously trade two assets for no cost:
a risky asset called “Bitcoin” with one unit of net supply and one risk-free asset with zero net
supply. 3
Assumption 1. Each unit of Bitcoin provides an observable stream of convenience yield 𝛿𝛿𝑡𝑡 , that

𝑑𝑑𝛿𝛿𝑡𝑡
grows according to: 4 = 𝑋𝑋𝑡𝑡 𝑑𝑑𝑑𝑑 + 𝜎𝜎𝛿𝛿 𝑑𝑑𝑍𝑍1𝑡𝑡 , (1)
𝛿𝛿𝑡𝑡

where the drift, 𝑋𝑋𝑡𝑡 , is unobservable, and evolves according to:

𝑑𝑑𝑋𝑋𝑡𝑡 = 𝜆𝜆(𝑋𝑋� − 𝑋𝑋𝑡𝑡 )𝑑𝑑𝑑𝑑 + 𝜌𝜌𝜎𝜎𝑋𝑋 𝑑𝑑𝑍𝑍1𝑡𝑡 + �1 − 𝜌𝜌2 𝜎𝜎𝑋𝑋 𝑑𝑑𝑍𝑍2𝑡𝑡 , (2)

where 𝜎𝜎𝛿𝛿 > 0, 𝜆𝜆 > 0, 𝑋𝑋� > 0, 𝜎𝜎𝑋𝑋 > 0, and 𝜌𝜌 ∈ [−1,1] are all known constants and (𝑍𝑍1𝑡𝑡 , 𝑍𝑍2𝑡𝑡 ) is a two-

dimensional standard Brownian motion.

While Bitcoin does not provide any cash flows, we assume it must offer some flow of benefits, which

we call “convenience yield", 𝛿𝛿𝑡𝑡 , to its owners, although there is significant and time-varying uncertainty

about how these benefits will evolve. For example, holding Bitcoin can facilitate transactions (particularly

illicit ones), hedge hyper-inflation risk caused by political turmoil, and serves as a store of value. As a result,

investors buy it trading off convenience yield and risks. For other financial assets like stocks and bonds, the

3 Obviously, in the real world, there are other risky assets in the economy. As a result, our model will not accurately
explain variation in the risk-free rate. In a single-risky-asset general equilibrium model, the risk-free rate will
necessarily have non-trivial volatility. However, adding other assets would not change our model's main point that
moving average rules can be optimal for trading Bitcoin due to learning. More generally, it is possible to add features
to correct counterfactual predictions about the risk-free rate, but only at the expense of parsimony.
4One of the features of Bitcoin that are different from traditional assets is that the supply is time varying,
because it depends on how much Bitcoin has been mined. However, incorporating this into the model would
make the analysis more complicated without changing the main mechanism behind our results.

Electronic copy available at: https://ssrn.com/abstract=3115846

convenience yield can be interpreted as a dividend, earnings, or interest paid to their owners. The state

variable 𝑋𝑋𝑡𝑡 is a catch-all for whatever state variable affects the convenience yield of an asset. For example,

in the case of Bitcoin, the state variable may capture uncertain regulatory risks, the likelihood of hyper-

inflation in some countries, the popularity of competing cryptocurrencies, and the related technology (e.g.,

block-chain update speed) advancement. In the case of stocks, the 𝑋𝑋𝑡𝑡 can represent the aggregate of all

variables that impact mean dividend growth.

On the investors, we make two assumptions:

Assumption 2 There are two types of investors who differ by their priors about the state variable

𝑋𝑋𝑡𝑡 and possibly initial endowment of Bitcoin. 5 Type 𝑖𝑖 investor is endowed with 𝜂𝜂𝑖𝑖 ∈ (0,1) units of

Bitcoin with 𝜂𝜂1 + 𝜂𝜂2 = 1 and has a prior that 𝑋𝑋0 is normally distributed with mean 𝑀𝑀𝑖𝑖 (0) and variance

𝑉𝑉 𝑖𝑖 (0), 𝑖𝑖 = 1,2.

Assumption 3 All investors have log preferences over the convenience yield provided by Bitcoin

with discount rate 𝛽𝛽 until time 𝑇𝑇. Specifically, the investor’s expected utility is

𝑇𝑇
𝐸𝐸 ∫0 𝑒𝑒 −𝛽𝛽𝛽𝛽 log𝐶𝐶𝑡𝑡𝑖𝑖 𝑑𝑑𝑑𝑑,

where 𝐶𝐶𝑡𝑡𝑖𝑖 denotes the convenience yield received by a Type 𝑖𝑖 investor from owning Bitcoin.

Intuitively, the heterogeneous priors across agents captures the fact that investors have significantly

different expectations about how Bitcoin’s fundamentals will grow in the future. Assuming log preferences

allows for relatively simple and transparent functional forms without altering what would be our main

prediction in the case of higher risk aversion.

Denote by 𝐹𝐹𝑡𝑡𝑖𝑖 the filtration at time 𝑡𝑡 generated by the Bitcoin price process {𝐵𝐵𝑠𝑠 } and the prior

(𝑀𝑀𝑖𝑖 (0), 𝑉𝑉 𝑖𝑖 (0)) for all 𝑠𝑠 ≤ 𝑡𝑡 and each investor 𝑖𝑖 = 1, 2. Because agents observe only one shock, 𝐹𝐹𝑡𝑡𝑖𝑖 turns

out to be equivalent to the filtration based on the convenience yield process {𝛿𝛿𝑠𝑠 } in lieu of {𝐵𝐵𝑠𝑠 }. Further let

𝑀𝑀𝑡𝑡𝑖𝑖 ≡ 𝐸𝐸[𝑋𝑋𝑡𝑡 |𝐹𝐹𝑡𝑡𝑖𝑖 ] be the conditional expectation of 𝑋𝑋𝑡𝑡 given 𝐹𝐹𝑡𝑡𝑖𝑖 . While many risky assets have uncertain and

5Since investors can continuously observe 𝛿𝛿𝑡𝑡 , they can directly calculate the volatility 𝜎𝜎𝛿𝛿 and therefore there is no
disagreement about the volatility.

Electronic copy available at: https://ssrn.com/abstract=3115846

stochastically evolving fundamental growth rates, the definition of the filtration and 𝑀𝑀𝑡𝑡 captures the essence

of the “hard-to-value" property. That is, Bitcoin investors must filter out the state variable from, at most, the

history of convenience yields, and they have initial disagreement about what the convenience yield growth

rate is. In contrast, assets with less hard-to-value fundamentals, such as large-cap stocks in established

industries, have other signals that help forecast fundamentals, such as thoroughly vetted accounting

statements, established correlations with macroeconomic conditions, and analyst coverage.

Proposition 1 In equilibrium in the economy defined by Assumptions 1–3:

𝑑𝑑𝐵𝐵𝑡𝑡 𝛿𝛿
𝐵𝐵𝑡𝑡
= �(𝛽𝛽 + 𝑀𝑀𝑡𝑡𝑖𝑖 ) − 𝐵𝐵𝑡𝑡 � 𝑑𝑑𝑑𝑑 + 𝜎𝜎𝛿𝛿 𝑑𝑑𝑍𝑍̂1𝑡𝑡
𝑖𝑖
, (3)
𝑡𝑡

the fraction of wealth invested in the Bitcoin by Investor 1 is

𝛼𝛼 𝑀𝑀𝑡𝑡1 −𝑀𝑀𝑡𝑡2
1 + 1+𝛼𝛼𝑡𝑡 𝜎𝜎𝛿𝛿2
, (4)
𝑡𝑡

and by Investor 2 is

1 𝑀𝑀𝑡𝑡1 −𝑀𝑀𝑡𝑡2
1− , (5)
1+𝛼𝛼𝑡𝑡 𝜎𝜎𝛿𝛿2

where
𝑡𝑡
𝐵𝐵𝑡𝑡 ∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢,𝑡𝑡)log𝐵𝐵𝑢𝑢 𝑑𝑑𝑑𝑑
𝑀𝑀𝑡𝑡𝑖𝑖 = ℎ𝑖𝑖 (𝑡𝑡) + 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡)log + (𝑓𝑓 𝑖𝑖 (𝑡𝑡, 𝑡𝑡) − 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡)) �log𝐵𝐵𝑡𝑡 − 𝑡𝑡 � (6)
𝐵𝐵0 ∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢,𝑡𝑡)𝑑𝑑𝑑𝑑

is the 𝑖𝑖th investor’s conditional expectation, 𝐸𝐸[𝑋𝑋𝑡𝑡 |𝐹𝐹𝑡𝑡𝑖𝑖 ], ℎ𝑖𝑖 (. ), 𝑓𝑓 𝑖𝑖 (. , . ), and 𝑔𝑔𝑖𝑖 (. , . ) are as defined in the

Appendix for 𝑖𝑖 = 1,2, and 𝛼𝛼𝑡𝑡 is as defined in Appendix (8), denoting the ratio of the marginal utility of type

1 investor to that of type 2 investor. In addition, if

2
𝜎𝜎𝑋𝑋
𝑉𝑉 𝑖𝑖 (0) ≤ 𝜌𝜌𝜎𝜎𝑋𝑋 𝜎𝜎𝛿𝛿 + 𝜆𝜆
, (7)

then

𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡) > 0, 𝑓𝑓 𝑖𝑖 (𝑡𝑡, 𝑡𝑡) − 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡) > 0, ∀𝑢𝑢 > 0, 𝑡𝑡 > 0. (8)

Proof. See Appendix.

2.2| Discussion and Related Literature

8

Electronic copy available at: https://ssrn.com/abstract=3115846

There are two important implications of Proposition 1. First, under Condition (7), the cum-yield Bitcoin

return is positively predictable by “ratios" of prices to their moving averages. Intuitively, because of

learning, investors only gradually update their priors in the direction of shocks to the convenience yield,

resulting in a price drift. If agents had access to other predictive signals, as they would with assets that are

not “hard to value", they would incorporate these signals and the predictability of prices by moving averages

would no longer be guaranteed. 6 Second, because of the predictive power of the price-to-MAs ratios, every

investor’s trading strategy depends on the moving averages. 7 The proof of Proposition 1 shows that learning

drives the entire return predictability in a way that is unaffected by the existence of multiple traders, which

serves only to create trading. As far as we know, this is the first equilibrium model that justifies the use of

moving averages of prices in guiding trading.

Existing models generate price drift via different mechanisms, most of which are based on
behavioral biases, such as under- or over-reaction. On the rational side, under certain circumstances,
rational expectations equilibrium (REE) models with learning, beginning with the seminal work of
Grossman (1976) and Hellwig (1980), also predict price drift. While agents in our model observe a
common signal, these models feature agents who receive private signals about fundamental value, and
they infer each other’s’ signals from the price. As discussed in Banerjee et al. (2009), price drift in these
models requires higher order disagreement to slow down the rate at which agents incorporate each
others’ signals in their private valuations. In contrast, agents in our model observe a common signal,
and price drift does not depend on multiple traders. Cochrane et al. (2008) also shows that price drift
can arise under certain conditions when multiple risky assets exist.

On the irrational side, many studies try to explain price drift, sometimes called time-series (TS)

6
One can easily extend this model to include an additional signal about the fundamental. In this extension, the traders
would still use MAs as a signal to trade. The less precise the additional fundamental signal, the more weight traders
would place on the price-to-MAs. Thus, the predictability of returns by the price-to-moving-average ratios should be
stronger the greater the degree to which fundamentals are hard-to-estimate.
2
7 𝜎𝜎𝑋𝑋
It can be shown that Condition (7) is guaranteed to hold eventually almost surely because 𝜌𝜌𝜎𝜎𝑋𝑋 𝜎𝜎𝛿𝛿 + is greater
𝜆𝜆
𝑖𝑖
than the steady-state level to which the conditional variance, 𝑉𝑉 (𝑡𝑡), monotonically converges. In addition, note that
even though 𝑉𝑉 𝑖𝑖 (0) can be large for assets with highly uncertain payoffs, the right hand side of the inequality can also
be large for these assets. So Condition (7) can be satisfied by even an asset that has a highly uncertain values before
reaching the steady state. With 𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡) > 0,
𝑡𝑡 𝑡𝑡
∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡)log𝐵𝐵𝑢𝑢 𝑑𝑑𝑑𝑑/ ∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡)𝑑𝑑𝑑𝑑,
is a weighted moving average of log Bitcoin prices.

Electronic copy available at: https://ssrn.com/abstract=3115846

momentum (e.g., Moskowitz et al., 2012, and Huang et al., 2019), and the related cross-sectional (CS)
momentum phenomenon of Jegadeesh and Titman (1993). One explanation proposed by this literature
is sustained over-reaction, which can be caused, for example, by positive feedback trading (DeLong et al.,
1990); Hong and Stein (1999), over-confidence and self-attribution confirmation biases (Daniel,
Hirshleifer, and Subrahmanyam, 1998), herding (Bikhchandani et al., 1992), or general sentiment
(Baker and Wurgler, 2006, 2007). A second explanation is under-reaction, which can be caused, for
example, by conservatism bias (Barberis et al., 1998), trend following (Hong and Stein, 1999), and
gradual diffusion of information (Hong and Stein, 1999; Hong et al., 2000). In our model, the reaction
of investors to signals resembles “under-reaction”, but not because of behavioral biases. With hard-
to-value fundamentals, investors rationally weight priors and incoming signals that feature
uncertainty, which leads to price drift, but this price drift does not represent mispricing.

Moskowitz et al. (2012) find that TS momentum can explain CS momentum empirically
and argue that many explanations for CS momentum are really explanations of TS momentum.
Thus, our model’s setting and prediction seem intuitively related to the finding that CS momentum
and post-earnings announcement drift are stronger when information uncertainty is higher (e.g,
Zhang, 2006). Presumably, information uncertainty is highly correlated with the degree to which
fundamentals are hard-to-value, and therefore we would expect to see greater price drift in
segments with high information uncertainty. Moreover, foreign investors, who presumably lack the
information of domestic investors, rely relatively heavily on CS momentum strategies (e.g., Choe
et al., 1999).

The second implication of Proposition 1 is that the optimal trading strategy is a function of
the MAs and trading is driven by a difference in investor beliefs. Thus, our model justifies one of the
most widely used class of technical analysis strategies, those based on MAs of prices. No prior rational
equilibrium endogenously generates such a practice. For example, early theoretical study of Zhu and
Zhou (2009) takes the MA rule as a given strategy for rational investors. Recently, Han et al. (2016)
propose a model with technical traders, but the traders use a price-to-moving average ratio rule
exogenously. In contrast, our paper seems the first that endogenizes MA trading in the model.

2.3| Cyptocurrency Literature

Our paper contributes to the growing literature on the economics of cryptocurrencies and the associated

10

Electronic copy available at: https://ssrn.com/abstract=3115846

blockchain technology. Relatively few papers in this vein study asset-pricing properties. Among them,
Liu and Tsyvinski (2018) document a TS momentum effect in cryptocurrency returns, but, unlike our
paper, do not provide a theory to rationalize this phenomenon. Using the Cagan model of
hyperinflation, Jermann (2018) empirically examines the relative contribution of shocks to volume and
velocity on variation in Bitcoin’s price. Jermann finds that most of the variation in Bitcoin’s price is
attributable to volume shocks, consistent with stochastic adoption dominating technology innovations.
Dwyer (2015) explains how cryptocurrencies can have positive value given limited supply. Athey et al.
(2016), Bolt and van Oordt (2016), and Pagnotta and Buraschi (2018) all provide models in which the
value of cryptocurrencies depends on some combination of (i) usage and the degree of adoption, (ii)
the scarcity of Bitcoin, and (iii) the value of anonymity.

Our model differs from those used by prior studies in at least two important respects. First, our

model does not require Bitcoin to be interpreted as a currency per se. We do not directly specify

currency-related determinants of its value (e.g. (i)–(iii) above). Rather, we model the flow of utility-

providing benefits as a random state variable, which we call a “convenience yield”, but admits a more

general interpretation. This generality is important because some market participants argue that Bitcoin

is better thought of as a speculative asset than a currency (e.g., Yermack, 2013). For example, Bitcoin’s

high volatility eliminates its use a store of value, a defining feature of money. Second, the papers cited

above all assume full information, however, our model features learning. This feature is critical given

the lack of agreement on what determines the value of Bitcoin. 8 The learning aspect of our model also

helps us to answer novel questions relative to the prior studies such as: What predicts Bitcoin returns?

Relative to asset pricing inquiries like ours, most of the literature on the economics of Bitcoin seeks

to identify problems, implementation issues, and uses of cryptocurrencies. Böhme et al. (2015) discuss the

virtual currency’s potential to disrupt existing payment systems and perhaps even monetary systems. Harvey

(2017) describes immense possibilities for the future for Bitcoin and its underlying blockchain technology.

Balvers and McDonald (2018) describe conditions and practical steps necessary for using blockchain

8
For instance, in a Bloomberg interview on December 4, 2013, Alan Greenspan stated: “You have to really stretch
your imagination to infer what the intrinsic value of Bitcoin is. I haven't been able to do it. Maybe somebody else can.”
11

Electronic copy available at: https://ssrn.com/abstract=3115846

technology as a global currency. Easley et al. (2019) provide a model of Bitcoin trading fees. Yermack

(2017) discusses use of Blockchain for trading equities and the corresponding governance implications.

Gandal et al. (2018) and Griffin and Shams (2018) document Bitcoin price manipulation. Biais et al. (2018)

model the reliability of the blockchain mechanism. Catalini and Gans (2017) discuss how blockchain

technology will shape the rate and direction of innovation. Chiu and Koeppl (2017) study the optimal design

of cryptocurrencies and assess quantitatively how well such currencies can support bilateral trade. Cong and

He (2018) model the impact of blockchain technology on information environments. Fernández-Villaverde

and Sanches (2017) model competition among privately issued currencies. Foley et al. (2018) document that

a large portion of Bitcoin transactions represents illegal activity. Huberman et al. (2017) model fees and

self-propagation mechanism of the Bitcoin payment system. Malinova and Park (2017) model the use of

blockchain in trading financial assets. Saleh (2017) examines economic viability of blockchain price-

formation mechanism. Prat and Walter (2016) show theoretically and empirically that Bitcoin prices forecast

Bitcoin production. Krueckeberg and Scholz (2018) argue that Bitcoin may qualify as a new asset class.

3| DATA

Bitcoin trades continuously on multiple exchanges around the world. We obtain daily Bitcoin prices
from the news and research site Coindesk.com, which is now standard in academic and professional
publications such as the Wall Street Journal, over the sample period July 18, 2010 (first day available)
through June 30, 2018. Starting July 1, 2013, Coindesk reports a Bitcoin price equal to the average of
those listed on large high-volume high-liquidity exchanges. Prior to July 2013, Coindesk reported the
price from Mt. Gox, an exchange that handled most of the trading volume in Bitcoin at the time. 9 We
also obtain data on two other cryptocurrencies, Ripple (XRP) and Ethereum (ETH), from
coinmarketcap.com. These two currencies are the largest competitors to Bitcoin by market cap, but
are only available over shorter samples (August 4, 2013–June 30, 2018 for XRP, and August 8, 2015–
June 30, 2018 for ETH).

We obtain daily risk-free rate, market excess return (𝑀𝑀𝑀𝑀𝑀𝑀), and returns on three Fama and French
(1993) size portfolios from the website of Kenneth French. The three size portfolios, “small", “medium",

9
For details on the history of the Bitcoin market, see Eha (2017).
12

Electronic copy available at: https://ssrn.com/abstract=3115846

and “big", are based on the NYSE 30%- and 70%-iles of market capitalization at the end of each month. To
measure the risk-free rate on weekends, we use the most recently available one-day risk-free rate. The
average risk free rate over this time (see below) is multiple orders of magnitude smaller than the average
Bitcoin return over this time so our risk-free rate assumptions can not have an economically meaningful
impact on our results. We obtain individual stock data from CRSP and analyst coverage from IBES. We
obtain daily prices and total returns on the NASDAQ total return index from Bloomberg. We obtain daily

levels of the CBOE implied volatility index (𝑉𝑉𝑉𝑉𝑉𝑉), 3-month and 10-year Treasury yields (𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 and 𝐿𝐿𝐿𝐿𝐿𝐿,

respectively), and Moody’s BAA- and AAA-bond index yields (𝐵𝐵𝐵𝐵𝐵𝐵 and 𝐴𝐴𝐴𝐴𝐴𝐴, respectively) from the St.
Louis Federal Reserve Bank website over the sample period July 18, 2010–June 30, 2018. We define
TERM = LTY −BILL and DEF = BAA−AAA. VIX, BILL, TERM, and DEF are commonly used
returns predictors and among the few available at the daily frequency (e.g., Ang and Bekaert, 2007;
Goyal and Welch, 2008; Brogaard and Detzel, 2015).
Table 1 presents summary statistics for select variables used in our predictability tests. Panel A

shows that Bitcoin earns an annualized daily excess return of 193.2% and a Sharpe ratio of 1.8 with an

annualized volatility of 106.2%. Moreover, Bitcoin has a modest positive autocorrelation. In contrast,

MKT has a modest negative autocorrelation and much lower average return and volatility over the

period of 13.7% and 14.8%, respectively. Although far less than the Sharpe ratio of Bitcoin, the

resulting MKT Sharpe ratio of 0.92 is relatively high by historical standards. Panel B presents summary

statistics for several benchmark return predictor variables used in the next section. All four are highly

persistent, with an autoregressive coefficient of 0.95–1.0. Moreover, Augmented Dickey-Fuller tests

fail to reject the null that any of the return predictors except VIX contain a unit root.

4| EMPIRICAL RESULTS

In this section, we test the predictions from our model in Section 2 that (i) short-horizon returns on

Bitcoin and other assets with hard-to-value fundamentals exhibit price drift and are predictable by

moving averages of price, and (ii) moving averages can explain Bitcoin trading volume. We begin by

examining the predictability of returns on Bitcoin, whose fundamentals are arguably “hardest" to value

13

Electronic copy available at: https://ssrn.com/abstract=3115846

among the assets we consider.

4.1 In-sample predictability of Bitcoin returns

Motivated by Eqs. (3) and (6), we test the predictability of one-day returns using the difference between

the log price of Bitcoin averages (instead of the 𝑔𝑔𝑖𝑖 -based weights). 10 and the moving average of these log

prices. For empirical work, we make two simplifications to the moving averages in Eq. (6). First, due to the

difficulties of estimating the exact functionals, we assume equal weighting in the moving Second, following

Brock et al. (1992), Lo et al. (2000), Han et al. (2013), Neely et al. (2014), and Han et al. (2016), we specify

fixed time horizons of L =1, 2, 4, 10, and 20 weeks for the moving averages even though these horizons are

endogenous in our model.

Specifically, letting 𝐵𝐵𝑡𝑡 denote the price of Bitcoin on day 𝑡𝑡, we define:

𝑏𝑏𝑡𝑡 = 𝑙𝑙𝑙𝑙𝑙𝑙(𝐵𝐵𝑡𝑡 ), (9)

and the moving averages by:

1
𝑚𝑚𝑎𝑎𝑡𝑡 (𝐿𝐿) = � � ∑𝑛𝑛⋅𝐿𝐿−1
𝑙𝑙=0 𝑏𝑏𝑡𝑡−𝑙𝑙 , (10)
𝑛𝑛⋅𝐿𝐿

where 𝑛𝑛 denotes the number of days per week in 𝐿𝐿 weeks. Bitcoin trades 7 days per week, however stock

returns and the macro predictors are only available on the 5 business days per week. Hence, for tests using

stock returns and the latter predictors, we use 𝑛𝑛 = 5. For tests using only Bitcoin returns and moving

averages, we use 7-day-per-week observations (𝑛𝑛 = 7). 11 The log price-to-moving average ratios, denoted

𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿), serve as our central predictor of interest in empirical tests and are defined as:

𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) = 𝑏𝑏𝑡𝑡 − 𝑚𝑚𝑎𝑎𝑡𝑡 (𝐿𝐿). (11)

Under condition (7), the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) should positively predict Bitcoin returns over short time

horizons. Table 2 evaluates in-sample predictive regressions of the form:

10 The exact functional form is determined by our exact setting, e.g., log utility. The functional form we use in
empirical tests still captures the essence of our model, which is robust to more general settings, that price drift exists
by functions of past prices, and these functions at least approximate price-to-moving average ratios.
11
To be clear, using 5-day (7-day) per week observations, the moving average horizons for 𝐿𝐿 =1, 2, 4, 10, and 20
weeks are, respectively, 5, 10, 20, 50, and 100 (7, 14, 28, 70, and 140) days.
14

Electronic copy available at: https://ssrn.com/abstract=3115846

 𝑟𝑟𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏′𝑋𝑋𝑡𝑡 + 𝜀𝜀𝑡𝑡+1 , (12)

where 𝑟𝑟𝑡𝑡+1 denotes the return on Bitcoin on day 𝑡𝑡 + 1. To facilitate comparison with predictability by

𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵, 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇, 𝐷𝐷𝐷𝐷𝐷𝐷, and 𝑉𝑉𝑉𝑉𝑉𝑉, we use 5-day “business" weeks throughout the table. Columns (1)–(5) of

Panel A present results with 𝑋𝑋𝑡𝑡 = 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) for each 𝐿𝐿. The 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) significantly predict 𝑟𝑟𝑡𝑡+1 for all

𝐿𝐿 with the positive sign predicted by our model. The moving averages of different horizons will

mechanically be highly correlated with each other. Hence, to test whether different horizons’ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿)

contain non-redundant predictive information, column (6) presents results in which the predictors are the

first three principal components of the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿), denoted 𝑋𝑋𝑡𝑡 = (𝑃𝑃𝑃𝑃1𝑡𝑡 , 𝑃𝑃𝑃𝑃2𝑡𝑡 , 𝑃𝑃𝑃𝑃3𝑡𝑡 )′. The second and

third principal components each load with at least marginal significance and the adjusted 𝑅𝑅 2 is roughly

three to four times as high as the specifications in columns (1)–(5). Hence, it appears the set of all

𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) contain at least two distinct predictive signals, consistent with our model in which different

traders use different MA horizons.

Panel B presents predictive regressions of the form Eq. (12) using the common “macro" return

predictors 𝑋𝑋𝑡𝑡 = 𝑉𝑉𝑉𝑉𝑋𝑋𝑡𝑡 , 𝐵𝐵𝐵𝐵𝐿𝐿𝐿𝐿𝑡𝑡 , 𝑇𝑇𝑇𝑇𝑇𝑇𝑀𝑀𝑡𝑡 , or 𝐷𝐷𝐷𝐷𝐹𝐹𝑡𝑡 . Columns (1)–(5) show that none of these variables

significantly predict Bitcoin returns in Eq. (12) either individually or jointly. Moreover, column (6), which

uses predictors 𝑋𝑋𝑡𝑡 = (𝑉𝑉𝑉𝑉𝑋𝑋𝑡𝑡 , 𝐵𝐵𝐵𝐵𝐵𝐵𝐿𝐿𝑡𝑡 , 𝑇𝑇𝑇𝑇𝑇𝑇𝑀𝑀𝑡𝑡 , 𝐷𝐷𝐷𝐷𝐹𝐹𝑡𝑡 , 𝑃𝑃𝑃𝑃1𝑡𝑡 , 𝑃𝑃𝑃𝑃2𝑡𝑡 , 𝑃𝑃𝑃𝑃3𝑡𝑡 )′, shows that the macro return

predictors do not subsume the predictive power of the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿).

Condition (7) will hold after enough time elapses with probability one as agents learn and posterior

variance decreases. However, at times when the variance of the conditional expectation is relatively high,

the predictive coefficient (analogous to the 𝑓𝑓 𝑖𝑖 (𝑡𝑡, 𝑡𝑡) − 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡) in Eq. (6)) on the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) should be

relatively low. When this variance is high enough to violate condition (7), which is most likely to happen at

the beginning of the sample, the predictive coefficient will even become negative. To test these patterns, we

proxy for variance of the state variable using a measure of the conditional variance of the Bitcoin return.

Specifically, we use the exponentially weighted moving average variance of Bitcoin returns, denoted 𝜎𝜎𝑡𝑡2 . 12

12 We use the smoothing parameter of 0.94 which is the default from RiskMetrics for computing conditional variances

of daily returns. 𝜎𝜎𝑡𝑡2 is defined recursively as 𝜎𝜎𝑡𝑡2 = (0.94) ∗ 𝜎𝜎𝑡𝑡−1

2
+ (0.06) ∗ 𝑟𝑟𝑡𝑡2 , where 𝑟𝑟𝑡𝑡 is the day-𝑡𝑡 return on
2
Bitcoin. 𝜎𝜎0 is defined to be the sample variance over the first 140 days of our sample that are not used in our return-
15

Electronic copy available at: https://ssrn.com/abstract=3115846

 Table 3 presents predictive regressions of the form:

𝑟𝑟𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) + 𝑐𝑐 ⋅ 𝜎𝜎𝑡𝑡2 + 𝑑𝑑 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) ⋅ 𝜎𝜎𝑡𝑡2 + 𝜀𝜀𝑡𝑡+1 . (13)

For these regressions, we use the whole sample period 12/06/2010–06/30/2018 and 7-day-per-week

observations. The 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) load significantly for all moving average horizons. Consistent with our model,

the interaction terms between 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) and 𝜎𝜎𝑡𝑡2 are all negative, so high variance attenuates the predictive

coefficients on the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿). Moreover, the interaction terms are significant for three of the five moving

average horizons.

The top graph in Figure 1 plots the conditional variance of the Bitcoin returns over time. Consistent

with the role of learning in our model, the variability of the conditional variance decreases over time. The

bottom graph in Figure 1 plots the coefficient on the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (4) conditional on variance (𝑏𝑏 + 𝑑𝑑 ⋅ 𝜎𝜎𝑡𝑡2 ).

Consistent with our model, this coefficient is positive most of the time, especially later in the sample,

however it is negative when conditional variance is high enough, early in the sample.

Overall, the in-sample predictability evidence in Tables 2 and 3 is consistent with our model. The

𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) positively predict Bitcoin returns on average. However, high conditional variance that exists

before agents have a chance to “learn it away" can reverse this predictive relationship.

4.2| Out-of-sample predictability of Bitcoin returns

It is well established that highly persistent regressors such as V IX, BILL, TERM, and DEF can
generate spuriously high in-sample return predictability (e.g., Stambaugh, 1999; Ferson et al., 2003;
Campbell and Yogo, 2006). These biases, parameter instability, and look-ahead biases imply that in-
sample estimates can overstate true real-time predictability, which directly impacts investors (e.g.,
Goyal and Welch, 2008). Hence, we next assess the out-of-sample predictability of Bitcoin returns.
2
Table 4 presents out-of-sample 𝑅𝑅 2 (𝑅𝑅𝑂𝑂𝑂𝑂 ) of forecasts from recursively estimated regressions similar

2
to those estimated in-sample in Table 2. 13 The first five columns of Panel A report 𝑅𝑅𝑂𝑂𝑂𝑂 based on

prediction tests because they are required to compute the initial 140-day moving average. In particular, the 𝜎𝜎𝑡𝑡2 is not
based on any “in-sample" data used in the predictive regressions. In our model, it can be shown that the conditional
variance of discretized returns is tightly linked to the posterior variance.
13 All out-of-sample regressions in this Table use expanding (not rolling) windows using all data available through 𝑡𝑡

to make the forecast for day 𝑡𝑡 + 1.

16

Electronic copy available at: https://ssrn.com/abstract=3115846

 2
regressions of the form Eq. (12). For robustness, we report 𝑅𝑅𝑂𝑂𝑂𝑂 using several split dates between the in-

sample and out-of-sample periods that include both relatively large in- and out-of-sample periods (e.g.,
2
Kelly and Pruitt, 2013). The last column (denoted MEAN), follows Rapach et al. (2010) and presents 𝑅𝑅𝑂𝑂𝑂𝑂

for the MEAN combination forecast, which is the simple average of the forecasts from the first five columns.

Prior studies find that the MEAN combination forecasts are robust, frequently outperforming more

sophisticated combination methods (that have more estimation error) in forecasting returns and other

macroeconomic time-series out-of-sample (e.g., Timmermann, 2006; Rapach et al., 2010; Detzel and

Strauss, 2018). Moreover, with diffuse priors about which MA horizon is optimal, technical traders would

presumably give equal-weight to the different forecasts.

2
Panel A shows that several of the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) individually predict returns out-of-sample with 𝑅𝑅𝑂𝑂𝑂𝑂 >

0. Moreover, for each split date, the MEAN forecasts forecasts predict returns with at least marginal

2
significance and 𝑅𝑅𝑂𝑂𝑂𝑂 of 0.83%–1.42%, which are high for the daily horizon. For comparison, Pettenuzzo

et al. (2014) find out-of-sample 𝑅𝑅 2 ranging from -0.08% to 0.55% for monthly stock returns. Panel B

presents results from similar tests as Panel A, but using 𝑉𝑉𝑉𝑉𝑉𝑉, 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵, 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇, and 𝐷𝐷𝐷𝐷𝐷𝐷 as predictors.

Unlike the forecasts based on the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿), those based on the macro predictors generally have negative

2
𝑅𝑅𝑂𝑂𝑂𝑂 . Prior evidence show that predicting returns out-of-sample is challenging, especially at short horizons.

Hence, it is already remarkable that we observe one-day out-of-sample predictability of Bitcoin returns by

the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿). It should also be the case that this predictability increases with horizon. Thus, in Panel C, we

2
present 𝑅𝑅𝑂𝑂𝑂𝑂 based on recursively estimated regressions of one-week (7-day) Bitcoin returns on the

𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿):

𝑟𝑟𝑡𝑡+1,𝑡𝑡+7 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) + 𝜀𝜀𝑡𝑡+1,𝑡𝑡+7 . (14)

Consistent with prior evidence on stock and bond return predictability, Panel C shows that for each
2
out-of-sample window and each 𝐿𝐿, the 𝑅𝑅𝑂𝑂𝑂𝑂 generally increase in both magnitude and significance relative

2 2
to the analogous one-day-return 𝑅𝑅𝑂𝑂𝑂𝑂 in Panel A. The MEAN forecast, for example, has 𝑅𝑅𝑂𝑂𝑂𝑂 that are

2
statistically significance and large for weekly returns. For comparison, Rapach et al. (2010) find 𝑅𝑅𝑂𝑂𝑂𝑂 of

1%–3.5% for quarterly stock returns. It is also worth noting that the predictability is not confined to the
17

Electronic copy available at: https://ssrn.com/abstract=3115846

 2
early part of the sample, the most recent 10% of the sample still has large and statistically significant 𝑅𝑅𝑂𝑂𝑂𝑂 .

Overall, the out-of-sample evidence shows that the in-sample predictability of Bitcoin returns does not

represent small-sample biases and evinces that investors can take advantage of Bitcoin predictability by

moving averages of log prices.

4.3| Performance of Bitcoin technical analysis strategies

The results above show that the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) predict Bitcoin returns with statistical significance. Next, we

evaluate the associated economic significance by assessing the performance of trading strategies based on

this predictability predictability (e.g., Pesaran and Timmermann, 1995; Cochrane, 2008; Rapach et al.,

2010). We define the buy indicator (buy=1) associated with each MA strategy, MA(𝐿𝐿), as:

1, if 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) > 0

𝑆𝑆𝐿𝐿,𝑡𝑡 = � (15)
0, otherwise

The return on the Bitcoin MA(𝐿𝐿) strategy on day 𝑡𝑡 is given by:

𝑀𝑀𝑀𝑀(𝐿𝐿)
𝑟𝑟𝑡𝑡 = 𝑆𝑆𝐿𝐿,𝑡𝑡 ⋅ 𝑟𝑟𝑡𝑡 + (1 − 𝑆𝑆𝐿𝐿,𝑡𝑡 ) ⋅ 𝑟𝑟𝑓𝑓𝑓𝑓 , (16)

where 𝑟𝑟𝑡𝑡 and 𝑟𝑟𝑓𝑓𝑓𝑓 denote, respectively, the return on Bitcoin and the risk-free rate on day 𝑡𝑡. Intuitively,

the trading strategy defined by Eq. (16) captures the short-term trends predicted by our model by going long

Bitcoin when its price is expected to trend upward, and vice versa. We denote the excess return of the buy-

𝑀𝑀𝑀𝑀(𝐿𝐿)
and-hold position in Bitcoin as 𝑟𝑟𝑥𝑥𝑡𝑡 and the excess return on the MA(𝐿𝐿) strategies by 𝑟𝑟𝑥𝑥𝑡𝑡 .

Table 5 presents summary statistics for the buy-and-hold and MA(𝐿𝐿) strategies. Panel A, which uses

the full sample (12/06/2010–6/30/2018), shows that all strategies are right-skewed and have fat tails. The

Sharpe ratio of Bitcoin is 1.8, which is about four times the historical Sharpe ratio of the stock market (e.g.,

Cochraine, 2005). All of the MA(𝐿𝐿) strategies further increase this ratio to 2.0 to 2.5. Moreover, all but one

of these Sharpe ratio gains are at least marginally significant using the heteroskedasticity and autocorrelation

(HAC) robust test for equality of Sharpe ratios of Ledoit and Wolf (2008). The maximum drawdown of

Bitcoin is 89.5%, while those of the MA(𝐿𝐿) strategies are all lower, ranging from 64.4% to 77.9%.

18

Electronic copy available at: https://ssrn.com/abstract=3115846

Comparing Panels B and C indicates that the performance of Bitcoin was higher during the first half of the

sample, although the Sharpe ratio gains of the MA(𝐿𝐿) strategies relative to the buy-and-hold position are

similar in both subsamples.

Panel A of Figure 0 plots the cumulative value of $1 invested in Bitcoin and the MA(2) (two-week)

strategy at the beginning of the sample. At the end of our sample, the $1 in Bitcoin grew to $33,617 while

the $1 in the MA(2) strategy grew to approximately $148,549, a difference of about $114,932 over 7.5

years! Panel B plots the drawdowns of Bitcoin and the MA(2) strategy. As Panel B shows, the out-

performance of the MA strategies relative to the buy and hold largely stems from the MA strategy having

both shorter and less severe drawdowns than the buy-and-hold. For example, Bitcoin prices hit an all-time

high in December 2017 at $19,343 and subsequently fell to $6,343 by the end of our sample. Panel B shows

investors using the MA(2) strategy would have been spared most of the losses from this price decline.

Table 6 further tests the performance of MA strategies relative to the buy-and-hold. Specifically, we

regress the excess returns of the MA strategies on the buy-and-hold benchmark:

𝑀𝑀𝑀𝑀(𝐿𝐿)
𝑟𝑟𝑥𝑥𝑡𝑡 = 𝛼𝛼 + 𝛽𝛽 ⋅ 𝑟𝑟𝑥𝑥𝑡𝑡 + 𝜀𝜀𝑡𝑡 . (17)

The MA(𝐿𝐿) strategies are long the risk-free rate up to about 40% of days, so 𝛽𝛽 < 1 and 𝛽𝛽 ⋅ 𝑟𝑟𝑥𝑥𝑡𝑡

is the natural benchmark return for evaluating the average returns of 𝑟𝑟𝑥𝑥 𝑀𝑀𝑀𝑀(𝐿𝐿) . Moreover, the arguments in

Lewellen and Nagel (2006) show that 𝛼𝛼 will increase with the quantity cov(𝑆𝑆𝐿𝐿,𝑡𝑡 , 𝐸𝐸𝑡𝑡 (𝑟𝑟𝑥𝑥𝑡𝑡 )) , which

measures the degree to which the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) positively predicts Bitcoin returns.

𝑀𝑀𝑀𝑀(𝐿𝐿)
A positive alpha also indicates that access to 𝑟𝑟𝑥𝑥𝑡𝑡 increases the maximum possible Sharpe ratio

relative to that of a buy-and-hold Bitcoin position (𝑟𝑟𝑥𝑥𝑡𝑡 ). Thus, a measure of the economic size of alpha is

the degree to which it expands the mean-variance frontier. Intuitively, this expansion depends on the alpha

relative to the residual risk investors must bear to capture it. The maximum Sharpe ratio (𝑆𝑆𝑅𝑅𝑁𝑁𝑒𝑒𝑒𝑒 ) attainable

𝑀𝑀𝑀𝑀(𝐿𝐿)
from access to 𝑟𝑟𝑥𝑥𝑡𝑡 and 𝑟𝑟𝑥𝑥𝑡𝑡 is given by:

𝛼𝛼 2
2
𝑆𝑆𝑅𝑅𝑁𝑁𝑁𝑁𝑁𝑁 = �� � + 𝑆𝑆𝑅𝑅𝑂𝑂𝑂𝑂𝑂𝑂 , (18)
𝜎𝜎(𝜀𝜀𝑡𝑡 )

where 𝑆𝑆𝑅𝑅𝑂𝑂𝑂𝑂𝑂𝑂 is the Sharpe ratio of 𝑟𝑟𝑥𝑥𝑡𝑡 (e.g., Bodie et al., 2014). The percentage increase in mean-

19

Electronic copy available at: https://ssrn.com/abstract=3115846

variance utility, which, for any level of risk aversion, is equal to:
2 2
𝑆𝑆𝑅𝑅𝑁𝑁𝑁𝑁𝑁𝑁 −𝑆𝑆𝑅𝑅𝑂𝑂𝑂𝑂𝑂𝑂
Utility gain = 2 , (19)
𝑆𝑆𝑅𝑅𝑂𝑂𝑂𝑂𝑂𝑂

measures the economic significance of the frontier expansion achieved by 𝛼𝛼. In Table 6, we report both the

𝛼𝛼
appraisal ratio � � and mean-variance utility gains along with the 𝛼𝛼. For comparison, Campbell and
𝜎𝜎(𝜀𝜀𝑡𝑡 )

Thompson (2008) find that timing expected returns on the stock market increases mean-variance utility by

approximately 35%, providing a useful benchmark utility gain.

Panel A shows that over the entire sample period, the MA(𝐿𝐿) strategies earn significant 𝛼𝛼 with

respect to 𝑟𝑟𝑥𝑥𝑡𝑡 of 0.09% to 0.24% per day. These alphas lead to economically large utility gains of 19.7%

to 85.5%. Panel B shows these results remain strong in the second half of the sample. With the turnovers in

the Table, it would take large transaction costs of (1.38%–6.85% one-way) to eliminate the alphas of the

MA(𝐿𝐿) strategies. These figures are large relative to actual one-way transaction costs in Bitcoin. For

example, market orders on the Bitcoin exchange GDAX have fees of 0.10%–0.30% for market orders and

0% for limit orders. Even the most expensive market order fees are an order of magnitude too small to

meaningfully impact the 𝛼𝛼s of the MA(𝐿𝐿) strategies.

A naive alternative to our discrete buy-or-sell strategies defined by Eq. (16) would be estimating

mean-variance weights using our Bitcoin-return forecasts, and then testing whether the resulting strategy

out-performs the buy-and-hold benchmark (e.g., Marquering and Verbeek, 2004; Campbell and Thompson,

2008; Huang et al., 2015). However, this approach has several theoretical and empirical shortcomings

relative to our simple MA(𝐿𝐿) strategies. First, the mean-variance weights assume the investor is choosing

between the market return and the risk-free asset. However, Bitcoin is a poor theoretical proxy to the market

portfolio of risky assets. Second, prior to 2017, investors could not short-sell Bitcoin or buy it on margin or

via futures contracts. Hence, the weights on Bitcoin should be constrained between zero and one. Thus, the

mean-variance-weights approach could only outperform the MA strategies by choosing optimal variation

between zero and one. This in turn exacerbates the following two problems: (i) that the mean-variance

weights require at least two estimated forecasts, and therefore come with substantial estimation error, and

20

Electronic copy available at: https://ssrn.com/abstract=3115846

(ii) the mean-variance weights assume for tractability the mean-variance functional form of investor utility.

While a common assumption, mean-variance utility is unlikely to precisely capture the behavior of a

representative investor. In contrast, our discrete MA(𝐿𝐿) strategies are based on a directly observable out-

of-sample signals and require no estimation error. They also make no assumption about the utility of

underlying investors. Overall, the strong performance of our MA(𝐿𝐿) strategies relative to the buy-and-hold

precludes the need for more sophisticated methods to demonstrate the economic significance of out-of-

sample predictability by MAs.

4.4| Performance of trading strategies applied to other cryptocurrencies

To examine the robustness of our trading strategy performance, Table 7 presents performance results similar

to those above for Ripple (XRP, Panels A and B) and Etheruem (ETH, Panels C and D), which are the two

largest digital currencies by market capitalization beside Bitcoin. Panel A shows that all the MA strategies

except MA(4) increase Sharpe ratios relative to the buy-and-hold strategy by up to 0.54 (from 1.05). This

difference is significant for the MA(1) and MA(2) strategies and marginally significant for the equal-

weighted portfolio of MA strategies (EW). Each strategy reduces the maximum drawdown of the buy-and-

hold Ripple strategy by about 4.7%-32.3%. Panel B shows that the MA(1), MA(2), MA(4), and EW

strategies also earn significant alphas with respect to the buy-and-hold XRP strategy, generating large utility

gains (92.9%–180.1%) in the process.

Panels C and D present similar results as Panels A and D, respectively, but for strategies based on

ETH instead of XRP. The ETH sample is only two and a half years long, leading to relatively low statistical

power, but qualitatively similar inferences as for the Bitcoin and Ethereurm MA strategies. The MA

strategies earn higher Sharpe ratios than the buy-and-hold ETH strategy. Panel C shows the ETH MA-

strategy alphas are significant for three horizons (1, 2, and 4 weeks) as well as the EW strategy and the

associated utility gains are economically large.

4.5| Performance of strategies applied to dotcom-era NASDAQ portfolio

Next, we apply each of the MA strategies defined by Eq. (16) to the NASDAQ total return index

using daily data over the sample 1996–2005, a ten year window approximately centered around the peak of

21

Electronic copy available at: https://ssrn.com/abstract=3115846

the NASDAQ “bubble" in March 2000. During this time, fundamentals of tech stocks were difficult to

interpret, widely disagreed upon, and had great forecast uncertainty. For example, Ofek and Richardson

(2003) document that during this period, aggregate earnings of internet stocks were negative and price-

earnings ratios frequently exceeded 1000. In particular, NASDAQ fundamentals plausibly qualify as hard-

to-interpret in the 1996–2005 sample.

Table 8 documents the performance of the MA strategies applied to the NASDAQ. Results in Panel

A show that over 1996–2005, the MA(2), MA(4) and MA(10) methods possess mean returns more than

four percent greater than the 7.3% of the buy-and-hold NASDAQ strategy. Further, all five methods

substantially boost the NASDAQ Sharpe ratio of 0.29. For example, MA(2), MA(4) and MA(10) possess

Sharpe ratios of 0.73 to 0.79. The last column documents that the MA strategies also greatly reduce the

maximum drawdown of NASDAQ (77.9%) to 25.7%–45.6%. 14 Panel B of Table 8 presents the alphas,

appraisal and utility gain for the NASDAQ. Results document significant alpha for MA(2) to MA(10)

strategies. It also reveals high utility gains for all five strategies, ranging from 137%–688%.

Panel C of Table 8 presents results over a tighter window around the NASDAQ peak. A number of

new internet companies entered the NASDAQ around this period, and the fundamentals of many other firms

were questioned after several large earnings misstatements. The NASDAQ peaked in March 2000, and by

the end of 2002 had lost 78% of its value. In contrast, the MA strategies have maximum drawdowns from

34%-43% and the equal-weighted MA strategy lost only 34% of its value—less than half the buy-and-hold

position. Further, the Sharpe ratios for the MA strategies during this five year window ranged from 0.36-

0.60, compared to zero for the buy-and-hold. The equal-weighted strategy generates a Sharpe ratio of 0.57

and is significantly greater than the buy-and-hold Sharpe ratio. Overall, the performance of the trading

strategies documented in Table 8 is consistent with the predictions of our model.

Figure 3 depicts the performance of the buy-and-hold position in NASDAQ relative to the MA(4)

strategy. Panel A shows the MA(4) increases more steadily than NASDAQ. The MA(4) returns $3.66 at

14 We also applied our strategies to the NASDAQ over the past ten years, which follows the maturation of internet-
based technologies and an increased understanding of fundamentals. These untabulated results show that the MA
strategies no longer earn significant alpha or produce Sharpe ratio gains.
22

Electronic copy available at: https://ssrn.com/abstract=3115846

the end of 2005 to an investor with a $1 investment at the beginning of 1996. Conversely, a buy-and-hold

investor in NASDAQ would have about half ($1.85) of the accumulated value. Panel A shows that much of

the performance gains from the MA strategy come from avoiding most of NASDAQ’s large crash in the

early 2000’s. Panel B further shows that the MA(4) strategy, similar to when applied to Bitcoin, derives

much of its performance from avoiding the major NASDAQ drawdowns during this time period. Following

the dotcom era, tech companies become more established and the availability of value-relevant information

presumably increases. Hence, according to our model, the MA strategies performance should decline after

this time period. Panel C confirms that, indeed, the Sharpe ratio improvements of the NASDAQ MA(4)

strategy steadily decline post-2001.

4.6| Performance of strategies applied to small-cap stocks, young stocks, and stocks with low

analyst coverage

The NASDAQ results exploit time-series variation in the degree to which fundamentals are hard to
forecast. Next we exploit cross-sectional variation. Specifically, we examine the predictability of
returns by price-to-MA ratios across portfolios sorted on size, age, and analyst coverage, which are three
common proxies for the availability of value relevant information (e.g., Bhushan, 1989; Hong et al., 2000;
Zhang, 2006).
In Panels A and B of Table 9, we apply our MA strategies to each of the three value-weighted Fama

and French (1993) size portfolios, “Small”, “Medium” and, “Big”. The sample period is July 1,

1963 through June 30, 2018. 15 Panel A presents heteroskedasticity-robust t-statistics from regressions of

daily excess portfolio returns on the price-to-moving average ratios:

𝑟𝑟𝑥𝑥𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) + 𝜀𝜀𝑡𝑡+1 . (20)

For each MA, these 𝑡𝑡-statistics are positive and highly significant (3.96–6.03) for the portfolio of small-

cap stocks. The 𝑡𝑡-statistics fall for mid-cap stocks (1.67–4.43) and become negative and insignificant for

large-cap stocks ((-0.96)–(-0.10)). Hence, consistent with our model, the predictability of returns by the

𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) is greater for small-cap stocks than large-caps.

15
Untabulated results show that inferences are robust across each half of this sample as well.
23

Electronic copy available at: https://ssrn.com/abstract=3115846

 Panel B presents Sharpe ratios for the buy-and-hold (BH) return on each portfolio as well as each of

the MA strategies and the equal-weighted portfolio (EW) of the MA strategies. The figures parallel those in

Panel A: Sharpe ratio gains of the MA strategies are highest for small-cap stocks, followed by mid-caps,

and then large-caps. For example, in small-caps, the Sharpe ratio gain of the EW strategy are 1.91 from 0.5,

relative to 0.65 from 0.39 in large caps. While the performance of the MA strategies is higher for small-caps

than large caps, the gains are still nontrivial for large-cap stocks. Coupled with the fact that large-cap stocks

returns are unpredictable by the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿), this finding indicates that the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) are useful volatility-

timing signals for large-caps. 16 Panels C and D repeat the analysis of Panels A and B, respectively, for

value-weighted tercile portfolios formed at the end of each June based on firm age, which, following Zhang

(2006), is defined to be the number of years a firm is in the CRSP database. Consistent with our theory and

the results in Panels A and B, Panels C and D show that younger firms, which are presumably harder to

value, have relatively high predictive coefficients in Eq. (20) and relatively high Sharpe ratios from the

𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) strategies. For example, the predictive coefficients for young firms in Panel C are significant for

all 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) with 𝑡𝑡-statistics of 2.89 or higher, while they are never significant, and often negative, for

old firms. Moreover, while the BH Sharpe ratios are the same for old and young firms in Panel D, they MA

Sharpe ratios of young firm are about twice or more than those of the old firms.

Table 10 presents 𝑡𝑡-statistics for predictive coefficients from regressions of the form Eq. (20) for

portfolios formed by independent sorts on size and analyst coverage. Due to IBES data availability, the

sample period for these tests is January 1985 through June 2018. Consistent with our model, the 𝑡𝑡 statistics

decrease with both size and analyst coverage. The 𝑡𝑡-statistics are large and positive in small caps, and within

small caps, predictive coefficients are significantly higher for stocks with the lowest analyst coverage.

Conversely, the predictive coefficients are all negative for large caps, and significantly moreso for the large-

cap portfolios with high analyst coverage.

Overall, the results from Tables 9 and 10 show that, consistent with our model, the price-to-MA

16
Sharpe ratio gains and alphas increase with (positive) return predictability of the timing signal (which, in our case,
are the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿)) and decrease with (positive) predictability of volatility (e.g., Lewellen and Nagel, 2006). That is, these
performance metrics measure the combined effect of “market timing” and “volatility timing” benefits of a given strategy.
24

Electronic copy available at: https://ssrn.com/abstract=3115846

ratios predict returns of small-cap, young-firm, and low-analyst-coverage stocks, all of which have a relative

lack of value-relevant information, and fail to predict returns of old-firm and large-cap stocks.

4.7| Volume implications of our model

In our model, trading results from differences in the moving averages used by different traders. Testing this

refutable implication provides an opportunity to validate our model’s mechanism in explaining the

predictability of Bitcoin by moving averages of multiple horizons.

We test for volume generated by technical trading in two ways. First, we evaluate whether increases

in total turnover implied by different MA signals also leads to higher Bitcoin volume. We measure this total

turnover by the sum of the turnover generated by each moving-average buy-sell indicator 𝑆𝑆𝐿𝐿,𝑡𝑡 : ∑𝐿𝐿 |Δ𝑆𝑆𝐿𝐿,𝑡𝑡 |.

Second, we evaluate whether disagreement among MA buy-sell indicators (𝑆𝑆𝐿𝐿,𝑡𝑡 ) is associated with higher

trading volume. Intuitively, if technical traders disagree, they will trade with each other. As a measure of

disagreement, we use the cross-sectional standard deviation of the signed turnover implied by each MA

strategy, denoted 𝜎𝜎𝐿𝐿 (Δ𝑆𝑆𝐿𝐿,𝑇𝑇 ). For each measure, Table 11 presents estimations of regressions of the form:

Δlog(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑒𝑒𝑡𝑡 ) = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑋𝑋𝑡𝑡 + 𝑐𝑐 ⋅ |𝑟𝑟𝑡𝑡 | + 𝑑𝑑 ⋅ Δlog(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑒𝑒𝑡𝑡−1 ) + 𝜀𝜀𝑡𝑡 , (21)

where the 𝑋𝑋𝑡𝑡 denotes one or both of our two volume-inducing variables. Because large price shocks are

the main empirical determinant of volume and are likely correlated with our price-based indicators, we

control for the absolute value of returns (see, e.g., Karpoff, 1987). We use change in log volume as the

dependent variable because the level of volume is not stationary. Volume is from coinmarketcap.com, which

began reporting on 12/27/2013, so these regressions use the 12/27/2013–6/30/2018 (𝑛𝑛 = 1,647). In Panel

A, we restrict 𝑑𝑑 = 0. However, to avoid any possibility of results being driven by autocorrelation in volume,

we do not make this restriction in Panel B.

Results in column (1) of both Panels demonstrate that increases in turnover across MA horizons lead

to increases in volume, controlling for price shocks. Similarly, column (2) of each panel shows that increases

in disagreement among MA traders also leads to significant increases in volume. Finally, comparing column

(3) of each Panel shows that the MA-implied turnover and disagreement jointly and positively correlate with

25

Electronic copy available at: https://ssrn.com/abstract=3115846

volume, however the statistical inference varies with specification. 17 Overall, the results in Table 11 are

consistent with traders using MA strategies significantly impacting trading volume in Bitcoin.

5.| CONCLUSION

In this paper, we theoretically and empirically examine dynamics of the prices of assets with “hard-

to-value" fundamentals, such as Bitcoin. We propose a new equilibrium theory that shows that when

fundamentals are hard to value, rational learning causes price drift and ratios of prices to their moving

averages to forecast returns. This in turn provides a fully rational motivation for common technical analysis

strategies that use price-to-moving average ratios, which are typically justified by mispricing-based

arguments. Our empirical results strongly confirm the predictions of our model. Bitcoin and stocks with

hard-to-value fundamentals are predictable by price-to-moving average ratios and simple real-time

strategies based on this predictability significantly outperform the buy-and-hold strategy. Given that the key

assumption underlying our model is the difficulty of forecasting fundamentals, a potentially fruitful avenue

for future research is extending our results to other assets well-described by this assumption, perhaps new

asset classes, and to examine the degree to which price drift is concentrated within these assets.

17 This finding is likely not due to multicollinearity, the correlation between 𝜎𝜎𝐿𝐿 �Δ𝑆𝑆𝐿𝐿,𝑡𝑡 � and Σ𝐿𝐿 �Δ𝑆𝑆𝐿𝐿,𝑡𝑡 � is 0.53.
26

Electronic copy available at: https://ssrn.com/abstract=3115846

 Appendix: Proof of Proposition 1

In this appendix, we present the proof of Proposition 1.

First we provide the evolution equations for conditional expectation and the conditional
variance. Following the standard continuous-time filtering theory, ∀𝑖𝑖 = 1, 2, 𝑀𝑀𝑡𝑡𝑖𝑖 satisfies
𝑑𝑑𝑀𝑀𝑡𝑡𝑖𝑖 = 𝜆𝜆(𝑋𝑋� − 𝑀𝑀𝑡𝑡𝑖𝑖 )𝑑𝑑𝑑𝑑 + 𝜎𝜎𝑀𝑀
𝑖𝑖
(𝑡𝑡)𝑑𝑑𝑍𝑍̂1𝑡𝑡
𝑖𝑖
, 𝑀𝑀0𝑖𝑖 = 𝑀𝑀𝑖𝑖 (0), (1)
where 𝑍𝑍̂1𝑡𝑡
𝑖𝑖
is the (observable) innovation processes satisfying
𝑡𝑡 𝑋𝑋𝑠𝑠 −𝑀𝑀𝑠𝑠𝑖𝑖
𝑍𝑍̂1𝑡𝑡
𝑖𝑖
= ∫0 𝑑𝑑𝑑𝑑 + 𝑍𝑍1𝑡𝑡 ,
𝜎𝜎𝛿𝛿

𝑖𝑖 𝑉𝑉 𝑖𝑖 (𝑡𝑡)
𝜎𝜎𝑀𝑀 (𝑡𝑡) = 𝜎𝜎𝛿𝛿
+ 𝜌𝜌𝜎𝜎𝑋𝑋 , 𝑉𝑉 𝑖𝑖 (𝑡𝑡) ≡ 𝐸𝐸[(𝑋𝑋𝑡𝑡 − 𝑀𝑀𝑡𝑡𝑖𝑖 )2 |𝐹𝐹𝑡𝑡𝑖𝑖 ] is the conditional variance of 𝑋𝑋𝑡𝑡 satisfying

𝑑𝑑𝑉𝑉 𝑖𝑖 (𝑡𝑡) 1 2
𝑑𝑑𝑑𝑑
= −2𝜆𝜆𝑉𝑉 𝑖𝑖 (𝑡𝑡) + 𝜎𝜎𝑋𝑋2 − �𝜎𝜎 𝑉𝑉 𝑖𝑖 (𝑡𝑡) + 𝜌𝜌𝜎𝜎𝑋𝑋 � . (2)
𝛿𝛿

This implies that

𝑑𝑑𝛿𝛿𝑡𝑡
= 𝑀𝑀𝑡𝑡𝑖𝑖 𝑑𝑑𝑑𝑑 + 𝜎𝜎𝛿𝛿 𝑑𝑑𝑍𝑍̂1𝑡𝑡
𝑖𝑖
, 𝑖𝑖 = 1, 2. (3)
𝛿𝛿𝑡𝑡

Next we derive the optimal trading strategy and equilibrium Bitcoin price. As in Detemple
(1986), Gennotte (1986), and Detemple (1991). 18 In particular, given the initial endowment
𝜂𝜂𝑖𝑖 > 0 and the prior (𝑀𝑀𝑖𝑖 (0− ), 𝑉𝑉𝑖𝑖 (0− )), Investor 𝑖𝑖’s portfolio selection problem is equivalent to
𝑇𝑇
max
𝑖𝑖 𝑖𝑖
𝐸𝐸 ∫0 𝑒𝑒 −𝛽𝛽𝛽𝛽 log𝐶𝐶𝑡𝑡𝑖𝑖 𝑑𝑑𝑑𝑑,
𝜃𝜃 ,𝐶𝐶

subject to
𝑑𝑑𝑊𝑊𝑡𝑡 = 𝑟𝑟𝑡𝑡 𝑊𝑊𝑡𝑡 𝑑𝑑𝑑𝑑 + 𝜃𝜃𝑡𝑡𝑖𝑖 (𝜇𝜇𝑡𝑡𝑖𝑖 − 𝑟𝑟𝑡𝑡 )𝑑𝑑𝑑𝑑 + 𝜃𝜃𝑡𝑡𝑖𝑖 𝜎𝜎𝛿𝛿 𝑑𝑑𝑍𝑍̂1𝑡𝑡
𝑖𝑖
− 𝐶𝐶𝑡𝑡𝑖𝑖 𝑑𝑑𝑑𝑑. (4)
Define 𝜋𝜋𝑡𝑡𝑖𝑖 as the state price density for investor 𝑖𝑖. Then
𝑑𝑑𝜋𝜋𝑡𝑡𝑖𝑖 = −𝑟𝑟𝑡𝑡 𝜋𝜋𝑡𝑡𝑖𝑖 𝑑𝑑𝑑𝑑 − 𝜅𝜅𝑡𝑡𝑖𝑖 𝜋𝜋𝑡𝑡𝑖𝑖 𝑑𝑑𝑍𝑍̂1𝑡𝑡
𝑖𝑖
, (5)
where 𝜅𝜅𝑡𝑡𝑖𝑖 is the price of risk perceived by investor 𝑖𝑖, i.e.,
𝜇𝜇𝑡𝑡𝑖𝑖 −𝑟𝑟𝑡𝑡
𝜅𝜅𝑡𝑡𝑖𝑖 = . (6)
𝜎𝜎𝛿𝛿

Using the standard dual approach (e.g., Cox and Huang, 1989) to solve Investor 𝑖𝑖’s problem, we
have
𝑒𝑒 −𝛽𝛽𝛽𝛽 (𝐶𝐶𝑡𝑡𝑖𝑖 )−1 = 𝜉𝜉𝑖𝑖 𝜋𝜋𝑡𝑡𝑖𝑖 , 𝑖𝑖 = 1, 2, (7)

18
The separation principle applies because the objective function is independent of the unobservable state variable
(see, e.g., Fleming and Rishel (1975, Chap. 4, Sec. 11) .
27

Electronic copy available at: https://ssrn.com/abstract=3115846

 where 𝜉𝜉𝑖𝑖 is the corresponding Lagrangian multiplier. Define
𝜉𝜉 𝜋𝜋1
𝛼𝛼𝑡𝑡 = 𝜉𝜉1 𝜋𝜋𝑡𝑡2 (8)
2 𝑡𝑡

to be the ratio of the marginal utilities. Then 𝛼𝛼𝑡𝑡 evolves as

𝑑𝑑
𝜇𝜇 𝜂𝜂
𝑑𝑑𝛼𝛼𝑡𝑡 = −𝛼𝛼𝑡𝑡 𝜎𝜎𝑡𝑡 𝑑𝑑𝑍𝑍̂1𝑡𝑡
1
, 𝜇𝜇𝑡𝑡𝑑𝑑 = 𝜇𝜇𝑡𝑡1 − 𝜇𝜇𝑡𝑡2 , 𝛼𝛼0 = 𝜂𝜂2 , (9)
𝛿𝛿 1

where the first equality is from Ito’s lemma and the consistency condition (i.e., the Bitcoin price is
the same across all investors), and the last equality follows from the budget constraints.
By market clearing condition 𝐶𝐶𝑡𝑡1 + 𝐶𝐶𝑡𝑡2 = 𝛿𝛿𝑡𝑡 , Equation (7) and Equation (8), we have
𝛿𝛿 𝛼𝛼 𝛿𝛿
𝐶𝐶𝑡𝑡1 = 1+𝛼𝛼𝑡𝑡 , 𝐶𝐶𝑡𝑡2 = 1+𝛼𝛼
𝑡𝑡 𝑡𝑡
. (10)
𝑡𝑡 𝑡𝑡

Then applying Ito’s lemma to (7) and compare with equation (5), we have
𝛼𝛼 𝜇𝜇𝑡𝑡𝑑𝑑 1 𝜇𝜇𝑡𝑡𝑑𝑑
𝜅𝜅𝑡𝑡1 = 𝜎𝜎𝛿𝛿 + 1+𝛼𝛼𝑡𝑡 , 𝜅𝜅𝑡𝑡2 = 𝜎𝜎𝛿𝛿 − 1+𝛼𝛼 ,
𝑡𝑡 𝜎𝜎𝛿𝛿 𝑡𝑡 𝜎𝜎𝛿𝛿

and
1 𝛼𝛼
𝑟𝑟𝑡𝑡 = 𝛽𝛽 + 1+𝛼𝛼 𝑀𝑀𝑡𝑡1 + 1+𝛼𝛼𝑡𝑡 𝑀𝑀𝑡𝑡2 − 𝜎𝜎𝛿𝛿2 . (11)
𝑡𝑡 𝑡𝑡

Therefore, the fraction of wealth invested in the Bitcoin by Investor 1 is

𝜅𝜅𝑡𝑡1 /𝜎𝜎𝛿𝛿 ,
i.e.,
𝛼𝛼 𝜇𝜇𝑡𝑡𝑑𝑑
1 + 1+𝛼𝛼𝑡𝑡 , (12)
𝑡𝑡 𝜎𝜎𝛿𝛿2

and by Investor 2 is
1 𝜇𝜇𝑡𝑡𝑑𝑑
1 − 1+𝛼𝛼 . (13)
𝑡𝑡 𝜎𝜎𝛿𝛿2

So if 𝜇𝜇𝑡𝑡𝑑𝑑 > 0, i.e., Investor 1 is more optimistic than Investor 2, then Investor 1 borrows to buy the
Bitcoin, and Investor 2 sells the Bitcoin and lends.
Using the expression for 𝐶𝐶𝑡𝑡1 and equation (7), we have the Bitcoin price
𝑇𝑇 𝜋𝜋𝑠𝑠1 1−𝑒𝑒 −𝛽𝛽(𝑇𝑇−𝑡𝑡)
𝐵𝐵𝑡𝑡 = 𝐸𝐸𝑡𝑡1 ∫𝑡𝑡 𝛿𝛿𝑠𝑠 𝑑𝑑𝑑𝑑 = 𝛿𝛿𝑡𝑡 ,
𝜋𝜋𝑡𝑡1 𝛽𝛽

which implies that

𝑑𝑑𝐵𝐵𝑡𝑡 = ((𝛽𝛽 + 𝑀𝑀𝑡𝑡𝑖𝑖 )𝐵𝐵𝑡𝑡 − 𝛿𝛿𝑡𝑡 )𝑑𝑑𝑑𝑑 + 𝜎𝜎𝛿𝛿 𝐵𝐵𝑡𝑡 𝑑𝑑𝑍𝑍̂1𝑡𝑡
𝑖𝑖
,

𝜇𝜇𝑡𝑡𝑖𝑖 = 𝛽𝛽 + 𝑀𝑀𝑡𝑡𝑖𝑖 . (14)

This implies that

28

Electronic copy available at: https://ssrn.com/abstract=3115846

 1 𝛽𝛽 1
𝑑𝑑𝑍𝑍̂1𝑡𝑡
𝑖𝑖
= 𝜎𝜎 �𝑑𝑑log𝐵𝐵𝑡𝑡 − �𝑀𝑀𝑡𝑡𝑖𝑖 − 1−𝑒𝑒 −𝛽𝛽(𝑇𝑇−𝑡𝑡) − 2 𝜎𝜎𝛿𝛿2 � 𝑑𝑑𝑑𝑑�.
𝛿𝛿

Investor 1’s wealth is

𝑇𝑇 𝜋𝜋𝑠𝑠1 1 1−𝑒𝑒 −𝛽𝛽(𝑇𝑇−𝑡𝑡) 1
𝑊𝑊1𝑡𝑡 = 𝐸𝐸𝑡𝑡1 ∫𝑡𝑡 𝐶𝐶 𝑑𝑑𝑑𝑑 = 𝐶𝐶𝑡𝑡1 = 1+𝛼𝛼 𝐵𝐵𝑡𝑡
𝜋𝜋𝑡𝑡1 𝑠𝑠 𝛽𝛽 𝑡𝑡

and Investor 2’s wealth is

𝑇𝑇 𝜋𝜋𝑠𝑠2 1−𝑒𝑒 −𝛽𝛽(𝑇𝑇−𝑡𝑡) 𝛼𝛼
𝑊𝑊2𝑡𝑡 = 𝐸𝐸𝑡𝑡1 ∫𝑡𝑡 𝐶𝐶𝑠𝑠2 𝑑𝑑𝑑𝑑 = 𝐶𝐶𝑡𝑡2 = 1+𝛼𝛼𝑡𝑡 𝐵𝐵𝑡𝑡 .
𝜋𝜋𝑡𝑡2 𝛽𝛽 𝑡𝑡

The number of Bitcoin Investor 1 holds is equal to

𝑑𝑑
𝛼𝛼𝑡𝑡 𝜇𝜇𝑡𝑡
(1+ )𝑊𝑊1𝑡𝑡
1+𝛼𝛼𝑡𝑡 𝜎𝜎2 1 𝛼𝛼 𝜇𝜇𝑡𝑡𝑑𝑑
𝑁𝑁1𝑡𝑡 = 𝛿𝛿
= 1+𝛼𝛼 (1 + 1+𝛼𝛼𝑡𝑡 ).
𝐵𝐵𝑡𝑡 𝑡𝑡 𝑡𝑡 𝜎𝜎𝛿𝛿2

The number of Bitcoin Investor 2 holds is equal to

𝑑𝑑
1 𝜇𝜇𝑡𝑡
(1− )𝑊𝑊2𝑡𝑡
1+𝛼𝛼𝑡𝑡 𝜎𝜎2 𝛼𝛼 1 𝜇𝜇𝑡𝑡𝑑𝑑
𝑁𝑁2𝑡𝑡 = 𝛿𝛿
= 1+𝛼𝛼𝑡𝑡 (1 − 1+𝛼𝛼 ).
𝐵𝐵𝑡𝑡 𝑡𝑡 𝑡𝑡 𝜎𝜎𝛿𝛿2

We have thus
𝜕𝜕𝑁𝑁1𝑡𝑡 −(1+𝛼𝛼𝑡𝑡 )+(1−𝛼𝛼𝑡𝑡 )𝜇𝜇𝑡𝑡𝑑𝑑 /𝜎𝜎𝛿𝛿2
= ,
𝜕𝜕𝛼𝛼𝑡𝑡 (1+𝛼𝛼𝑡𝑡 )3

which is < 0 if and only if

𝜇𝜇𝑑𝑑 /𝜎𝜎2 −1
𝛼𝛼𝑡𝑡 > 𝜇𝜇𝑡𝑡𝑑𝑑 /𝜎𝜎𝛿𝛿2 +1.
𝑡𝑡 𝛿𝛿

Next we derive the expression of the conditional expectation 𝑀𝑀𝑡𝑡𝑖𝑖 in the form of moving averages.
We have
𝑑𝑑𝑀𝑀𝑡𝑡𝑖𝑖 = (𝑎𝑎𝑖𝑖 (𝑡𝑡) − 𝑏𝑏 𝑖𝑖 (𝑡𝑡)𝑀𝑀𝑡𝑡𝑖𝑖 )𝑑𝑑𝑑𝑑 + 𝑐𝑐 𝑖𝑖 (𝑡𝑡)𝑑𝑑log𝐵𝐵𝑡𝑡 , (15)
where
𝛽𝛽 1
𝑎𝑎𝑖𝑖 (𝑡𝑡) = 𝜆𝜆𝑋𝑋� + �1−𝑒𝑒 −𝛽𝛽(𝑇𝑇−𝑡𝑡) + 2 𝜎𝜎𝛿𝛿2 � 𝑐𝑐 𝑖𝑖 (𝑡𝑡),

𝑖𝑖
𝜎𝜎𝑀𝑀 (𝑡𝑡)
𝑏𝑏 𝑖𝑖 (𝑡𝑡) = 𝜆𝜆 + 𝑐𝑐 𝑖𝑖 (𝑡𝑡), 𝑐𝑐 𝑖𝑖 (𝑡𝑡) = 𝜎𝜎𝛿𝛿
.

Equation (15) implies that

𝑡𝑡
𝑀𝑀𝑡𝑡𝑖𝑖 = ℎ𝑖𝑖 (𝑡𝑡) + ∫0 𝑓𝑓 𝑖𝑖 (𝑢𝑢, 𝑡𝑡)𝑑𝑑log𝐵𝐵𝑢𝑢 ,
where
𝑡𝑡 𝑖𝑖 𝑡𝑡 𝑢𝑢 𝑖𝑖
ℎ𝑖𝑖 (𝑡𝑡) = 𝑒𝑒 − ∫0 𝑏𝑏 (𝑠𝑠)𝑑𝑑𝑑𝑑 ∫0 𝑎𝑎𝑖𝑖 (𝑢𝑢)𝑒𝑒 ∫0 𝑏𝑏 (𝑠𝑠)𝑑𝑑𝑑𝑑
𝑑𝑑𝑑𝑑,.
and

29

Electronic copy available at: https://ssrn.com/abstract=3115846

 𝑢𝑢 𝑖𝑖
𝑏𝑏 (𝑠𝑠)𝑑𝑑𝑑𝑑
𝑓𝑓 𝑖𝑖 (𝑢𝑢, 𝑡𝑡) = 𝑐𝑐 𝑖𝑖 (𝑢𝑢)𝑒𝑒 ∫𝑡𝑡 . (16)
Then by integration by parts, we have
𝑡𝑡
𝑀𝑀𝑡𝑡𝑖𝑖 = ℎ𝑖𝑖 (𝑡𝑡) − 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡)log𝐵𝐵0 + 𝑐𝑐 𝑖𝑖 (𝑡𝑡)log𝐵𝐵𝑡𝑡 − ∫0 log𝐵𝐵𝑢𝑢 𝑑𝑑𝑓𝑓 𝑖𝑖 (𝑢𝑢, 𝑡𝑡)
𝑡𝑡
𝐵𝐵 ∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢,𝑡𝑡)log𝐵𝐵𝑢𝑢 𝑑𝑑𝑑𝑑
= ℎ𝑖𝑖 (𝑡𝑡) + 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡)log 𝐵𝐵𝑡𝑡 + (𝑓𝑓 𝑖𝑖 (𝑡𝑡, 𝑡𝑡) − 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡)) �log𝐵𝐵𝑡𝑡 − 𝑡𝑡 �,
0 ∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢,𝑡𝑡)𝑑𝑑𝑑𝑑

where
𝜕𝜕𝑓𝑓 𝑖𝑖 (𝑢𝑢,𝑡𝑡)
𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡) = 𝜕𝜕𝜕𝜕
. (17)

We show next that if Condition (7) is satisfied, then 𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡) > 0. To prove this, we
substitute 𝑓𝑓 𝑖𝑖 (𝑢𝑢, 𝑡𝑡) in Equation (16) into (17) to obtain
𝑑𝑑𝑐𝑐 𝑖𝑖 (𝑢𝑢) 𝑢𝑢 𝑖𝑖
𝑏𝑏 (𝑠𝑠)𝑑𝑑𝑑𝑑
𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡) = � 𝑑𝑑𝑑𝑑
+ 𝑐𝑐 𝑖𝑖 (𝑢𝑢)𝑏𝑏 𝑖𝑖 (𝑢𝑢)� 𝑒𝑒 ∫𝑡𝑡 , (18)
𝑑𝑑𝑐𝑐 𝑖𝑖 (𝑢𝑢)
thus we need to find condition for 𝑑𝑑𝑑𝑑
+ 𝑐𝑐 𝑖𝑖 (𝑢𝑢)𝑏𝑏 𝑖𝑖 (𝑢𝑢) > 0. Note that
𝑑𝑑𝑐𝑐 𝑖𝑖 (𝑡𝑡) 𝜎𝜎 𝜎𝜎2
𝑑𝑑𝑑𝑑
= −2𝜆𝜆 �𝑐𝑐 𝑖𝑖 (𝑡𝑡) − 𝜌𝜌 𝜎𝜎𝑋𝑋� + 𝜎𝜎𝑋𝑋2 − (𝑐𝑐 𝑖𝑖 (𝑡𝑡))2 , (19)
𝛿𝛿 𝛿𝛿

and 𝑏𝑏 𝑖𝑖 (𝑡𝑡) = 𝜆𝜆 + 𝑐𝑐 𝑖𝑖 (𝑡𝑡), we need to have the following condition

𝜎𝜎 𝜎𝜎2
𝑐𝑐 𝑖𝑖 (𝑡𝑡) < 2𝜌𝜌 𝜎𝜎𝑋𝑋 + 𝜆𝜆𝜎𝜎𝑋𝑋2 . (20)
𝛿𝛿 𝛿𝛿
𝜎𝜎
Due to the dynamics of 𝑐𝑐 𝑖𝑖 (𝑡𝑡) given in Equation (19), it can be proven that at 𝑐𝑐 𝑖𝑖 (𝑡𝑡) = 2𝜌𝜌 𝜎𝜎𝑋𝑋 +
𝛿𝛿
2
𝜎𝜎𝑋𝑋 𝑑𝑑𝑐𝑐 𝑖𝑖 (𝑡𝑡)
, < 0. This implies that as long as
𝜆𝜆𝜎𝜎𝛿𝛿2 𝑑𝑑𝑑𝑑

𝜎𝜎 𝜎𝜎2
𝑐𝑐 𝑖𝑖 (0) ≤ 2𝜌𝜌 𝜎𝜎𝑋𝑋 + 𝜆𝜆𝜎𝜎𝑋𝑋2 , (21)
𝛿𝛿 𝛿𝛿

or equivalently,
2
𝜎𝜎𝑋𝑋
𝑉𝑉 𝑖𝑖 (0) ≤ 𝜌𝜌𝜎𝜎𝑋𝑋 𝜎𝜎𝛿𝛿 + , (22)
𝜆𝜆

Condition (7) holds.

Under Condition (7), the expression
𝑡𝑡
∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢,𝑡𝑡)log𝐵𝐵𝑢𝑢 𝑑𝑑𝑑𝑑
𝑡𝑡 (23)
∫0 𝑔𝑔𝑖𝑖 (𝑢𝑢,𝑡𝑡)𝑑𝑑𝑑𝑑

is a weighted average of log(𝐵𝐵𝑢𝑢 ) over the interval [0, 𝑡𝑡]. In addition, by the definition of
𝑔𝑔𝑖𝑖 (𝑢𝑢, 𝑡𝑡), this implies that
𝑓𝑓 𝑖𝑖 (𝑡𝑡, 𝑡𝑡) − 𝑓𝑓 𝑖𝑖 (0, 𝑡𝑡) > 0
for any 𝑡𝑡. This completes the proof of Proposition 1.

30

Electronic copy available at: https://ssrn.com/abstract=3115846

Figure 1: Conditional variance of Bitcoin returns

Panel A depicts the condition variance of Bitcoin returns using an exponentially

weighted moving average variance (σ2 ) of daily Bitcoin returns.
Panel B depicts the predictive coefficient of pmat(4) conditional on (b + d· σ2 ) using the following
regression estimated in Table 3:
𝒓𝒓𝒕𝒕+𝟏𝟏 = 𝒂𝒂 + 𝒃𝒃 ⋅ 𝒑𝒑𝒑𝒑𝒂𝒂𝒕𝒕 (𝑳𝑳) + 𝒄𝒄 ⋅ 𝝈𝝈𝟐𝟐𝒕𝒕 + 𝒅𝒅 ⋅ 𝒑𝒑𝒑𝒑𝒂𝒂𝒕𝒕 (𝑳𝑳) ⋅ 𝝈𝝈𝟐𝟐𝒕𝒕 + 𝜺𝜺𝒕𝒕+𝟏𝟏 . (1)

31

Electronic copy available at: https://ssrn.com/abstract=3115846

Figure 2: Performance of investment in Bitcoin buy-and-hold and MA strategies

Panel A presents cumulative returns to $1 invested in the buy-and-hold and MA(4)

Bitcoin strategies over 7/18/2010–6/30/2018. Panel B presents drawdowns of each
strategy in Panel A.

32

Electronic copy available at: https://ssrn.com/abstract=3115846

Figure 3: Performance of investment in NASDAQ buy-and-hold and MA(4) strategies

Panel A plots cumulative returns to $1 invested in the buy-and-hold and MA(4) NASDAQ
strategies on 1/2/1996 through 12/30/2005. Panel B plots drawdowns of each strategy. Panel C
plots Sharpe ratios for each strategy estimated, for each date t, using the sample period 1/2/1996
through t.

33

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 1: Summary statistics

Panel A presents summary statistics of the returns in excess of the 1-day risk-free rate on Bitcoin (BTC) and the CRSP
value-weighted index (𝑀𝑀𝑀𝑀𝑀𝑀). Means, standard deviations, and Sharpe ratios are annualized. Panel B presents summary
statistics of other relevant variables. AR1 denotes the first-order autoregressive coefficient and 𝑝𝑝𝑑𝑑𝑑𝑑 denotes the p-value
from an augmented Dickey-Fuller test for the null of a unit root. The sample period is daily from 12/06/2010−6/30/2018.
Bitcoin returns trade 7 days a week and have 2,766 observations during the sample period. Other variables are available
5 days a week and have 1,976 observations during this period.

Panel A: Returns
Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis AR1
𝐵𝐵𝐵𝐵𝐵𝐵 193.18 106.20 1.82 -38.83 52.89 0.78 14.97 0.05
𝑀𝑀𝑀𝑀 13.65 14.76 0.92 -6.97 4.97 -0.52 8.04 -0.08
Panel B: Predictor variables
Mean(%) SD(%) Min(%) Max(%) Skewness Kurtosis AR1 𝑝𝑝𝑑𝑑𝑑𝑑
𝑉𝑉𝑉𝑉𝑉𝑉 16.30 5.53 9.14 48.00 2.05 8.36 0.95 0.00
𝐵𝐵𝐵𝐵𝐵𝐵 0.32 0.48 -0.02 1.91 1.92 5.61 1.00 1.00
𝑇𝑇𝑇𝑇𝑇𝑇 1.99 0.61 0.87 3.60 0.45 2.50 1.00 0.89
𝐷𝐷𝐷𝐷𝐷𝐷 0.95 0.25 0.53 1.54 0.62 2.33 1.00 0.31

34

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 2: In-sample predictability of Bitcoin returns

This table presents estimates of predictive regressions of the form: 𝑟𝑟𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏′𝑋𝑋𝑡𝑡 + 𝜖𝜖𝑡𝑡+1 , where 𝑟𝑟𝑡𝑡+1 denotes the
return on Bitcoin on business day 𝑡𝑡 + 1 . In Panel A, the predictors are the log price/moving average
ratios, 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿), where L is the number of weeks, with 5 business days per week. We also extract the first three
principal components (𝑃𝑃𝑃𝑃1, 𝑃𝑃𝑃𝑃2, 𝑜𝑜𝑜𝑜 𝑃𝑃𝑃𝑃3) from the 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿). In columns (1) to (5) of Panel B, the predictors
include these principal components along with the other return predictors (𝑉𝑉𝑉𝑉𝑉𝑉, 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵, 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇, and 𝐷𝐷𝐷𝐷𝐷𝐷). Column (6)
of Panel B adds the three principal components. The sample period are business days 12/06/2010–6/30/2018 (𝑛𝑛 =
1,976). Heteroscedasticity-robust t-statistics are presented in parentheses.

Panel A: 1-Day Predictability of Bitcoin returns by log moving average/price ratios

(1) (2) (3) (4) (5) (6)
𝑝𝑝𝑝𝑝𝑝𝑝(1) 0.39 𝑃𝑃𝑃𝑃1 -0.01
(2.61) (-1.46)
𝑝𝑝𝑝𝑝𝑝𝑝(2) 0.40 𝑃𝑃𝑃𝑃2 2.64
(2.67) (2.86)
𝑝𝑝𝑝𝑝𝑝𝑝(4) 0.42 𝑃𝑃𝑃𝑃3 -5.01
(2.81) (1.82)
𝑝𝑝𝑝𝑝𝑝𝑝(10) 0.46
(3.02)
𝑝𝑝𝑝𝑝𝑝𝑝(20) 0.45
(3.00)
2
Adj-𝑅𝑅
(%) 0.41 0.45 0.51 0.62 0.64 1.88
Panel B: 1-Day Predictability of Bitcoin returns by log moving average/price ratios
(1) (2) (3) (4) (5) (6)
𝑉𝑉𝑉𝑉𝑉𝑉 -0.04 -0.05 -0.02
(-1.33) (-1.30) (-0.34)
𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 -0.11 0.05 0.02
(-0.71) (1.08) (1.91)
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 0.34 0.34 -0.04
(1.01) (-0.49) (-0.59)
𝐷𝐷𝐷𝐷𝐷𝐷 -0.43 -0.75 -0.01
(-1.68) (-1.85) (1.53)
𝑃𝑃𝑃𝑃1 -0.01
(2.32)
𝑃𝑃𝑃𝑃3 2.24
(2.50)
𝑃𝑃𝑃𝑃3 5.30
(1.84)
2
Adj-𝑅𝑅
0.01 0.00 0.00 0.00 0.20 1.82
(%)

35

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 3: In-sample predictability of Bitcoin returns conditional on variance

This table presents estimates of predictive regressions of the form:

𝑟𝑟𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) + 𝑐𝑐 ⋅ 𝜎𝜎𝑡𝑡2 + 𝑑𝑑 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) ⋅ 𝜎𝜎𝑡𝑡2 + 𝜖𝜖𝑡𝑡+1 ,
where 𝑟𝑟𝑡𝑡+1 denotes the return on Bitcoin on day 𝑡𝑡 + 1, 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) denotes the log price-to-𝐿𝐿-week moving average
ratio, and 𝜎𝜎𝑡𝑡2 denotes the exponential weighted moving average variance of Bitcoin returns. The 𝜎𝜎𝑡𝑡2 are defined
recursively as 𝜎𝜎𝑡𝑡2 = 0.94 ⋅ 𝜎𝜎𝑡𝑡−1
2
+ 0.06 ⋅ 𝑟𝑟𝑡𝑡2 . The sample period for the regression is 12/06/2010−6/30/2018 (n=2,766
using 7-day-per week observations). The initial 𝜎𝜎02 is estimated as the sample variance of 𝑟𝑟𝑡𝑡 over 7/28/2010–
12/05/2010. Heteroskedasticity-robust t-statistics are presented in parentheses. *, **, *** denotes 10%, 5%, 1%
significance levels.

(1) (2) (3) (4) (5)

𝑝𝑝𝑝𝑝𝑝𝑝(1)
(2.13)
2
𝑝𝑝𝑝𝑝𝑝𝑝(1) ⋅ 𝜎𝜎 -5.03
(-1.31)
𝑝𝑝𝑝𝑝𝑝𝑝(2) 7.52***
(3.29)
𝑝𝑝𝑝𝑝𝑝𝑝(2) ⋅ 𝜎𝜎 2 -5.57**
(-2.07)
𝑝𝑝𝑝𝑝𝑝𝑝(4) 5.81***
(4.01)
𝑝𝑝𝑝𝑝𝑝𝑝(4) ⋅ 𝜎𝜎 2 -3.95**
(-2.43)
𝑝𝑝𝑝𝑝𝑝𝑝(10) 2.87***
(3.55)
𝑝𝑝𝑝𝑝𝑝𝑝(10) ⋅ 𝜎𝜎 2 -1.83**
(-2.08)
𝑝𝑝𝑝𝑝𝑝𝑝(20) 1.26***
(2.59)
𝑝𝑝𝑝𝑝𝑝𝑝(20) ⋅ 𝜎𝜎 2 -0.98
(-1.56)
𝜎𝜎 2 1.04** 1.28*** 1.50*** 1.51** 1.56**
(2.12) (2.70) (2.97) (2.51) (2.24)
Adj-𝑅𝑅 2 (%) 0.93 1.62 2.10 1.77 1.09

36

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 4: Out-of-sample predictability of Bitcoin returns
2
Panels A and B present 𝑅𝑅𝑂𝑂𝑂𝑂 (out-of-sample 𝑅𝑅2 ) in percent for recursively estimated predictive regressions of the form:
𝑟𝑟𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏′𝑋𝑋𝑡𝑡 + 𝜖𝜖𝑡𝑡+1 , where 𝑟𝑟𝑡𝑡+1 denotes day-𝑡𝑡 + 1 return on Bitcoin. Both panels use 5-day-per week observations.
In Panel A, the predictors are the 𝑝𝑝𝑝𝑝𝑝𝑝(𝐿𝐿) and in Panel B, they are 𝑉𝑉𝑉𝑉𝑉𝑉, 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵, 𝐷𝐷𝐷𝐷𝐷𝐷, and 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇. Panel C uses the 7-
day-per-week-observations and forecasts one week returns (𝑟𝑟𝑡𝑡+1,𝑡𝑡+7 ) using recursively estimated regressions of the form:
𝑟𝑟𝑡𝑡+1,𝑡𝑡+7 = 𝑎𝑎 + 𝑏𝑏 ′ 𝑋𝑋𝑡𝑡 + 𝜖𝜖𝑡𝑡+1,𝑡𝑡+7 . 𝑇𝑇0 denotes the in-sample period as a percentage of the total sample. The MEAN is a
simple combination forecast that averages the five moving average forecasts. The sample is 12/06/2010−6/30/2018
(n=1976 in Panels A and B; n=2766 for Panel C). *, **, *** denotes 10%, 5%, 1% significance levels using the Clark-West
2
(2007) MSFE-adjusted statistic that tests the null of equal MSFE (𝑅𝑅𝑂𝑂𝑂𝑂 =0) against the competing model that has a lower
2
MSFE (𝑅𝑅𝑂𝑂𝑂𝑂 >0).

Panel A: 1-day horizon, 5-day-per-week observations

𝑇𝑇0 𝑝𝑝𝑝𝑝𝑝𝑝(1) 𝑝𝑝𝑝𝑝𝑝𝑝(2) 𝑝𝑝𝑝𝑝𝑝𝑝(4) 𝑝𝑝𝑝𝑝𝑝𝑝(10) 𝑝𝑝𝑝𝑝𝑝𝑝(20) MEAN
25% -0.32 -0.11 0.70** 1.01** 0.72** 0.83*
50% -0.73 -0.27 0.08 0.31** 0.86*** 0.38*
90% 0.94 1.13* 1.51* 0.81 0.70 1.42*
Panel B: 1-day horizon, 5-day-per-week observations
𝑇𝑇0 VIX BILL TERM DEF MEAN
25% -0.87 -0.21 -0.27 -0.01 -0.14
50% -1.07 0.06 -0.06 -0.03 -0.13
90% -0.77 -0.08 -0.02 0.17 -0.10
Panel C: 1-week horizon, 7-day-per-week observations
𝑇𝑇0 𝑝𝑝𝑝𝑝𝑝𝑝(1) 𝑝𝑝𝑝𝑝𝑝𝑝(2) 𝑝𝑝𝑝𝑝𝑝𝑝(4) 𝑝𝑝𝑝𝑝𝑝𝑝(10) 𝑝𝑝𝑝𝑝𝑝𝑝(20) MEAN
25% -0.11 0.84** 3.67** 3.71** 1.66** 3.08**
50% 1.05** 1.06** 2.38** 2.87** 4.13** 3.62**
90% 2.70** 2.32* 2.57** 1.07* 1.66** 1.92**

37

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 5: Performance of Bitcoin trading strategies

This table presents summary statistics of the returns in excess of the 1-day risk-free rate on Bitcoin (BTC) and each of the
MA(𝐿𝐿) Bitcoin strategies, which take a long position in Bitcoin if 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) > 0, and the risk-free rate otherwise. EW
denotes an equal-weighted portfolio of the individual 𝑀𝑀𝑀𝑀(𝐿𝐿) strategies. Means, standard deviations, and Sharpe ratios are
annualized. The sample period is daily from 12/06/2010−6/30/2018 (7 days per week). Panel A presents full sample results
(𝑛𝑛=2,764). Panels B and C, respectively, present results for the first (9/17/2010-8/28/2014, n=1,383) and second halves
(8/29/2014-6/30/2018, 𝑛𝑛=1,383) of the sample. MDD denotes maximum drawdown. We use Ledoit and Wolf (2008) test
of equality of Sharpe ratios that is robust to heteroskedasticity and serial correlation. *, **, *** denotes significance at the
10%, 5%, and 1% confidence levels, respectively.

Panel A: Full-sample
Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis MDD(%)
BTC 193.18 106.20 1.82 -38.83 52.89 0.78 14.97 89.48
**
MA(1) 196.54 79.33 2.48 -38.83 52.89 2.12 31.47 71.65
MA(2) 187.38 79.20 2.37** -38.83 52.89 2.07 31.27 64.43
*
MA(4) 187.34 82.67 2.27 -38.83 52.89 1.63 29.21 69.66
*
MA(10) 195.70 88.50 2.21 -38.83 52.89 1.59 24.99 70.28
MA(20) 188.77 94.96 1.99 -38.83 52.89 1.27 21.11 77.87
EW 191.15 78.72 2.43*** -38.83 52.89 2.09 31.31 64.60
Panel B: First-half
Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis MDD(%)
BTC 288.46 129.50 2.23 -38.83 52.89 0.77 12.47 89.48
MA(1) 279.04 98.77 2.83* -38.83 52.89 1.85 24.58 71.65
MA(2) 272.33 99.01 2.75 -38.83 52.89 1.88 24.32 64.43
MA(4) 277.63 104.11 2.67 -38.83 52.89 1.45 22.16 69.66
MA(10) 309.57 110.78 2.79** -38.83 52.89 1.48 19.20 70.28
MA(20) 271.54 118.56 2.29* -38.83 52.89 1.17 16.26 77.87
EW 282.02 99.41 2.84** -38.83 52.89 1.85 23.54 64.60
Panel C: Second-half
Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis MDD(%)
BTC 97.89 75.80 1.29 -21.90 25.41 0.10 8.14 69.77
MA(1) 114.03 52.90 2.16** -11.13 25.41 1.68 16.37 32.82
MA(2) 102.43 52.02 1.97* -11.13 22.97 1.02 12.81 37.27
MA(4) 97.05 52.81 1.84 -14.24 22.97 0.66 12.24 43.85
MA(10) 81.83 57.67 1.42 -16.73 22.97 0.13 11.34 56.33
MA(20) 106.01 62.82 1.69 -16.73 25.41 0.41 12.03 58.44
EW 100.27 49.70 2.02** -11.13 22.97 1.00 12.10 40.05

38

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 6: Alphas of MA Bitcoin strategies relative to buy-and-hold benchmark
MA(𝐿𝐿)
Panels A and B present regressions of the form: 𝑟𝑟𝑟𝑟𝑡𝑡 = 𝛼𝛼 + 𝛽𝛽 ⋅ 𝑟𝑟𝑥𝑥𝑡𝑡 + 𝜖𝜖𝑡𝑡 , where 𝑟𝑟𝑥𝑥𝑡𝑡 denotes the day-𝑡𝑡 buy-and-hold
MA(𝐿𝐿)
excess return on Bitcoin and 𝑟𝑟𝑥𝑥𝑡𝑡 denotes the excess return on the MA(𝐿𝐿) Bitcoin strategy. Beneath each regression
MA(𝐿𝐿)
is the appraisal ratio of the MA strategy as well as the utility gain from access to 𝑟𝑟𝑟𝑟𝑡𝑡 in addition to 𝑟𝑟𝑥𝑥𝑡𝑡 . EW denotes
an equal-weighted portfolio of the MA strategies. Panel A also reports the average daily turnover (TO) of the MA strategies
and the one-way transaction cost (FEE) that would be required to eliminate the alpha of the MA strategy. Panel A presents
results for the full sample period (12/06/2010−6/30/2018, 𝑛𝑛=2,766). Panel B presents results for the second half of the
sample (n=1,383). Heteroskedasticity-robust t-statistics are below point estimates in parentheses. *, **, *** denotes
significance at the 10%, 5%, and 1% confidence levels, respectively.

Panel A: Full-sample
MA(1) MA(2) MA(4)MA(20) MA(10) EW
𝛽𝛽 0.56*** 0.56*** 0.61***
0.80***0.69*** 0.64***
(15.65) (15.52) (17.70)
(34.57)(23.02) (24.08)
𝛼𝛼(%) 0.24*** 0.22*** 0.19***
0.09** 0.17*** 0.18***
(4.66) (4.21) (3.75)
(2.12) (3.37) (4.71)
2
𝑅𝑅 0.56 0.56 0.610.80 0.69 0.75
Appraisal 1.68 1.52 1.350.81 1.26 1.71
Utility gain(%) 85.53 69.39 55.45
19.70 47.84 88.15
TO(%) 17.62 10.35 6.151.37 2.93 7.68
FEE(%) 1.38 2.12 3.136.85 5.75 2.39
Panel B: Second-half subsample
MA(1) MA(2) MA(4) MA(10) MA(20) EW
𝛽𝛽 0.49*** 0.47*** 0.49*** 0.58*** 0.69*** 0.54***
(12.84) (12.51) (12.76) (14.70) (19.56) (19.29)
𝛼𝛼(%) 0.18*** 0.15*** 0.14** 0.07 0.11** 0.13***
(3.41) (2.90) (2.54) (1.31) (2.10) (3.27)
𝑅𝑅 2 0.49 0.47 0.48 0.58 0.69 0.68
Appraisal 1.75 1.49 1.31 0.67 1.10 1.69
Utility gain(%) 183.90 132.81 102.43 27.17 72.90 170.74

39

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 7: Performance of trading strategies applied to Ripple and ETH

Panel A presents summary statistics of the returns in excess of the 1-day risk-free rate on Ripple (XRP) and each of the
MA(𝐿𝐿) strategies applied to Ripple. Means, standard deviations, and Sharpe ratios are annualized. MDD denotes
maximum drawdown. EW denotes an equal-weighted portfolio of the MA strategies. Panel B presents regressions of the
MA(𝐿𝐿) MA(𝐿𝐿)
form: 𝑟𝑟𝑟𝑟𝑡𝑡 = 𝛼𝛼 + 𝛽𝛽 ⋅ 𝑟𝑟𝑥𝑥𝑡𝑡 + 𝜖𝜖𝑡𝑡 , where 𝑟𝑟𝑥𝑥𝑡𝑡 denotes the day-𝑡𝑡 buy-and-hold excess return on XRP and 𝑟𝑟𝑥𝑥𝑡𝑡
denotes the excess return on the MA(𝐿𝐿) XRP strategy. Beneath each regression is the appraisal ratio of the MA strategy
MA(𝐿𝐿)
and the utility gain from access to 𝑟𝑟𝑟𝑟𝑡𝑡 . In Panels A and B, the sample is 12/24/2013–6/30/2018 (𝑛𝑛=1,650). Panels
C and D, presents similar statistics as Panels A and B, respectively, but for strategies applied to Ethereum (ETH) instead
of XRP. In Panels C and D, the sample is 12/28/2015–6/30/2018 (𝑛𝑛=916). We use the Ledoit and Wolf (2008) test of
equality of Sharpe ratios. Heteroskedasticity robust t-statistics are below point estimates in parentheses. *, **, ***
denotes significance at the 10%, 5%, and 1% confidence levels, respectively.

Panel A: Summary Statistics for XRP strategies

Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis MDD(%)
XRP 167.15 159.61 1.05 -46.01 179.37 7.54 141.85 90.22
MA(1) **
222.36 140.17 1.59 -46.01 179.37 10.90 234.94 77.60
MA(2) **
222.98 140.98 1.58 -46.01 179.37 10.85 230.83 72.41
MA(4) 196.70 140.42 1.40 -46.01 179.37 10.93 234.77 57.93
MA(10) 164.11 141.39 1.16 -46.01 179.37 10.76 228.61 81.63
MA(20) 135.51 143.88 0.94 -46.01 179.37 10.24 213.68 85.50
EW *
188.33 136.34 1.38 -46.01 179.37 11.86 262.17 65.72
Panel B: Strategy alphas for XRP strategies
(1) (2) (3) (4) (5) (6)
MA(1) MA(2) MA(4) MA(10) MA(20) EW
𝛽𝛽 0.77*** 0.78*** 0.77*** 0.78*** 0.81*** 0.79***
(10.99) (11.87) (11.36) (11.87) (13.83) (12.34)
𝛼𝛼(%) 0.26*** 0.25*** 0.18** 0.09 -0.00 0.16**
(3.24) (3.20) (2.33) (1.15) (-0.01) (2.49)
𝑅𝑅 2 0.77 0.78 0.78 0.78 0.81 0.84
Appraisal 1.40 1.41 1.01 0.50 0.00 1.06
Utility gain(%) 178.29 180.09 92.91 22.97 0.00 102.93

40

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 7: (Cont’d)

Panel C: Summary Statistics for ETH strategies

Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis MDD(%)
ETH 334.16 132.00 2.53 -27.06 35.36 0.81 6.90 73.48
MA(1) 319.09 107.37 2.97 -27.06 35.36 1.68 12.49 42.36
MA(2) 310.48 108.48 2.86 -27.06 35.36 1.61 12.15 43.11
MA(4) *
345.23 111.45 3.10 -27.06 35.36 1.41 11.44 56.81
MA(10) 247.73 116.86 2.12 -27.06 35.36 1.10 10.13 75.23
MA(20) 302.12 122.65 2.46 -27.06 35.36 1.05 8.77 69.92
EW 304.93 107.28 2.84 -27.06 35.36 1.57 12.23 50.41
Panel D: Strategy alphas for ETH strategies
(1) (2) (3) (4) (5) (6)
MA(1) MA(2) MA(4) MA(10) MA(20) EW
𝛽𝛽 0.66*** 0.67*** 0.71*** 0.78*** 0.86*** 0.74***
(17.83) (18.49) (21.39) (26.36) (39.89) (27.89)
𝛼𝛼(%) 0.27** 0.23** 0.29*** -0.04 0.04 0.16**
(2.55) (2.23) (2.90) (-0.41) (0.50) (2.17)
2
𝑅𝑅 0.66 0.67 0.71 0.78 0.86 0.82
Appraisal 1.57 1.37 1.79 0.00 0.31 1.30
Utility gain(%) 38.39 29.39 50.07 0.00 1.48 26.25

41

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 8: Performance of trading strategies applied to NASDAQ over 1996–2005

Panel A presents summary statistics of the returns in excess of the 1-day risk-free rate on NASDAQ and each of the
MA(𝐿𝐿) NASDAQ strategies. Means, standard deviations, and Sharpe ratios are annualized. EW denotes an equal-
weighted portfolio of the MA strategies. MDD denotes maximum drawdown. Panel B presents regressions of the form:
MA(𝐿𝐿) MA(𝐿𝐿)
𝑟𝑟𝑟𝑟𝑡𝑡 = 𝛼𝛼 + 𝛽𝛽 ⋅ 𝑟𝑟𝑥𝑥𝑡𝑡 + 𝜖𝜖𝑡𝑡 , where 𝑟𝑟𝑥𝑥𝑡𝑡 denotes the day-𝑡𝑡 buy-and-hold excess return on NASDAQ and 𝑟𝑟𝑥𝑥𝑡𝑡
denotes the excess return on the MA(𝐿𝐿) NASDAQ strategy. Beneath each regression is the appraisal ratio of the MA
MA(𝐿𝐿)
strategy and the utility gain from access to 𝑟𝑟𝑟𝑟𝑡𝑡 . The sample period is 1/2/1996−12/30/2005 (𝑛𝑛=2,519). Panel C
present results similar to Panel A using over the 1998–2002 subsample (𝑛𝑛=1,256). We use the Ledoit and Wolf (2008)
test of equality of Sharpe ratios. Heteroskedasticity-robust t-statistics are below point estimates in parentheses. *, **,
*** denotes significance at the 10%, 5%, and 1% confidence levels, respectively.

Panel A: Summary Statistics of NASDAQ strategies

Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis MDD(%)
NASDAQ 8.53 29.01 0.29 -9.69 14.15 0.19 7.14 77.93
MA(1) 9.32 18.34 0.51 -6.23 8.10 0.07 9.55 42.73
MA(2) 12.80 17.63 0.73* -6.23 8.10 0.04 9.17 42.08
*
MA(4) 13.34 17.41 0.77 -5.59 8.10 -0.07 8.36 25.66
*
MA(10) 13.84 17.45 0.79 -7.66 4.92 -0.40 7.37 33.81
MA(20) 7.78 17.20 0.45 -7.66 4.28 -0.50 7.68 45.62
EW 11.42 15.09 0.76** -5.58 4.86 -0.16 5.97 34.49
Panel B: Strategy alphas
(1) (2) (3) (4) (5) (6)
MA(1) MA(2) MA(4) MA(10) MA(20) EW
𝛽𝛽 0.40*** 0.37*** 0.36*** 0.36*** 0.35*** 0.37***
(16.66) (16.23) (16.28) (16.53) (16.32) (19.97)
𝛼𝛼(%) 0.02 0.04** 0.04** 0.04** 0.02 0.03**
(1.32) (2.18) (2.33) (2.44) (1.09) (2.47)
2
𝑅𝑅 0.40 0.37 0.36 0.36 0.35 0.50
Appraisal 0.42 0.69 0.74 0.77 0.34 0.78
Utility gain(%) 199.85 548.18 627.66 687.82 137.42 105.47
Panel C: Summary Statistics of NASDAQ strategies over 1998–2002
Mean(%) SD(%) Sharpe Min(%) Max(%) Skewness Kurtosis MDD(%)
NASDAQ -0.13 37.18 0.00 -9.69 14.15 0.23 5.08 77.93
MA(1) 9.62 23.07 0.42 -6.23 8.10 0.06 7.17 42.73
MA(2) 10.69 21.93 0.49* -6.23 8.10 0.05 7.14 42.08
MA(4) 12.84 21.47 0.60* -5.59 8.10 -0.07 6.62 25.66
MA(10) 11.89 21.12 0.56* -7.66 4.92 -0.43 6.10 33.81
MA(20) 7.11 19.98 0.36 -7.66 4.28 -0.48 6.59 39.75
EW 10.43 18.23 0.57** -5.58 4.86 -0.15 4.83 34.49

42

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 9: Predictability of returns and performance of MA strategies across size portfolios

In Panels A and B, we apply our MA strategies to each of the three value-weighted Fama and French (1993) size
portfolios, “Small”, “Medium”, and “Big”. Panel A presents heteroskedasticity-robust t-statistics from regressions of
daily excess portfolio returns on the 𝐿𝐿-week price-to-moving average ratios:
𝑟𝑟𝑥𝑥𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) + 𝜖𝜖𝑡𝑡+1 .

Panel B presents Sharpe ratios for the buy-and-hold (BH) return on each portfolio as well as each of the MA strategies
and the equal-weighted portfolio (EW) of the MA strategies. Panels C and D repeat the analysis of Panels A and B,
respectively, using age-sorted tercile portfolios, “Young”, “Medium”, and “Old”. The sample period is July 1, 1963
through June 30, 2018.

Panel A: t-statistics from predictive regressions for returns of size portfolios

MA(1) MA(2) MA(4) MA(10) MA(20)
Small 6.03 5.90 5.87 5.23 3.96
Medium 4.43 3.42 3.06 2.44 1.67
Big -0.10 -0.92 -0.96 -0.83 -0.67
Panel B: Sharpe ratios of MA strategies applied to size portfolios
BH MA(1) MA(2) MA(4) MA(10) MA(20) EW
Small 0.50 2.05 1.90 1.75 1.52 1.16 1.91
Mediu
0.51 1.73 1.52 1.27 1.10 0.87 1.49
m
Large 0.39 0.73 0.62 0.53 0.48 0.47 0.65
Panel C: t-statistics from predictive regressions for returns of age portfolios
MA(1) MA(2) MA(4) MA(10) MA(20)
Young 6.56 4.97 4.49 3.70 2.89
Medium 3.96 2.59 2.14 1.82 1.33
Old 0.62 -0.25 -0.33 -0.34 -0.32
Panel D: Sharpe ratios of MA strategies applied to age portfolios
BH MA(1) MA(2) MA(4) MA(10) MA(20) EW
Young 0.48 2.00 1.75 1.57 1.35 1.12 1.79
Mediu
0.46 1.53 1.22 1.16 0.95 0.85 1.31
m
Old 0.47 0.91 0.73 0.61 0.59 0.63 0.80

43

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 10: Predictability of returns and performance of MA strategies across portfolios
formed on size and analyst coverage

At the end of each year, we form nine value-weighted portfolios formed as the intersections of two independent tercile
sorts of all U.S. common stocks into three portfolios based on each of market capitalization (“Small”, “Medium”, and
“Big”) and the number of analyst forecasts over the year (“Low”, “Medium”, and “High”). The first three columns of
the table present heteroskedasticity-robust t-statistics from regressions of daily excess portfolio returns on the 𝐿𝐿-week
price-to-moving average ratios:
𝑟𝑟𝑥𝑥𝑡𝑡+1 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑝𝑝𝑝𝑝𝑎𝑎𝑡𝑡 (𝐿𝐿) + 𝜖𝜖𝑡𝑡+1 .
The fourth column of the table presents heteroskedasticity-robust GMM-based t-statistics of the difference between
the 𝑏𝑏 from to the “Low” portfolio in the row minus the 𝑏𝑏 from the “High” portfolio in the row. The sample period is
January 2, 1985 through June 29, 2018.

MA(1)
Low Med High Low-High
Small 5.24 5.04 4.93 2.95
Med 1.48 1.83 2.72 -1.75
Big 0.19 0.00 -1.94 4.08
MA(2)
Low Med High Low-High
Small 5.18 4.87 4.39 2.95
Med 1.16 1.34 1.94 -1.17
Big -0.45 -0.63 -2.16 3.58
MA(4)
Low Med High Low-High
Small 5.19 4.71 4.19 2.78
Med 1.18 1.33 1.68 -0.66
Big -0.45 -0.64 -1.83 2.97
MA(10)
Low Med High Low-High
Small 4.98 4.40 3.32 2.80
Med 1.01 1.16 1.44 -0.47
Big -0.08 -0.39 -1.39 2.76
MA(20)
Low Med High Low-High
Small 3.94 3.10 2.29 2.35
Med 0.43 0.47 0.71 -0.31
Big 0.10 -0.43 -1.14 2.56

44

Electronic copy available at: https://ssrn.com/abstract=3115846

Table 11: Volume and technical trading indicators

This table presents regressions of the form:

Δ log(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣)𝑡𝑡 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑋𝑋𝑡𝑡 + 𝑐𝑐 ⋅ |rt |+ 𝜖𝜖𝑡𝑡 ,
Δ log(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣)𝑡𝑡 = 𝑎𝑎 + 𝑏𝑏 ⋅ 𝑋𝑋𝑡𝑡 + 𝑐𝑐 ⋅ |rt | + 𝑑𝑑 ⋅ Δ log(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣)𝑡𝑡−1 + 𝜖𝜖𝑡𝑡 ,
where 𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑒𝑒𝑡𝑡 denotes the trading volume in Bitcoin on day 𝑡𝑡, |𝑟𝑟𝑡𝑡 | denotes the absolute return on Bitcoin on day
𝑡𝑡, and 𝑋𝑋𝑡𝑡 denotes one of two predictors. The second equation introduces lagged volume to accommodate for possible
serial correlation. In column (1), 𝑋𝑋𝑡𝑡 is the sum �∑𝐿𝐿�Δ𝑆𝑆𝐿𝐿,𝑡𝑡 �� of the absolute turnovers �Δ𝑆𝑆𝐿𝐿,𝑡𝑡 � from each of the
MA(𝐿𝐿) strategies. In column (2), 𝑋𝑋𝑡𝑡 is the cross-sectional standard deviation �𝜎𝜎𝐿𝐿 �Δ𝑆𝑆𝐿𝐿,𝑡𝑡 �� of Δ𝑆𝑆𝐿𝐿,𝑡𝑡 , a measure of
the “disagreement” among technical traders using the different MA strategies (𝐿𝐿 = 1, 2, 4, 10, or 20 weeks). In
column (3), 𝑋𝑋𝑡𝑡 includes ∑𝐿𝐿�Δ𝑆𝑆𝐿𝐿,𝑡𝑡 � and 𝜎𝜎𝐿𝐿 �Δ𝑆𝑆𝐿𝐿,𝑡𝑡 �. The sample is 12/27/2013–6/30/2018 (𝑛𝑛=1,647).
Heteroskedasticity-robust t-statistics are in parentheses.

Panel A: Determinants of Volume, without controlling for lagged volume

(1) (2) (3)
∑𝐿𝐿�|Δ𝑆𝑆𝐿𝐿,𝑡𝑡 |� 0.03 0.07
(4.38) (1.47)
𝜎𝜎𝐿𝐿 �Δ𝑆𝑆𝐿𝐿,𝑡𝑡 � 0.15 0.05
(3.65) (3.46)
|𝑟𝑟𝑡𝑡 | 4.90 5.03 4.73
(11.94) (11.76) (11.58)
2
Adj-𝑅𝑅 0.17 0.16 0.17
Panel B: Determinants of Volume, controlling for lagged Volume
(1) (2) (3)
∑𝐿𝐿�|Δ𝑆𝑆𝐿𝐿,𝑡𝑡 |� 0.03 0.04
(3.56) (3.85)
𝜎𝜎𝐿𝐿 �Δ𝑆𝑆𝐿𝐿,𝑡𝑡 � 0.12 0.16
(2.78) (1.80)
|𝑟𝑟𝑡𝑡 | 5.07 5.18 5.04
(13.05) (12.91) (12.69)
Δ log(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣)𝑡𝑡−1 -0.23 -0.23 -0.23
(10.13) (-10.21) (10.81)
2
Adj-𝑅𝑅 0.22 0.22 0.22

45

Electronic copy available at: https://ssrn.com/abstract=3115846

 REFERENCES
Ang, A. and Bekaert, G. (2007). Stock return predictability: is it there?. Review of Financial Studies,
20, 651–707.

Athey, S., Parashkevov, I., Sarukkai, V. and Xia, J. (2016). Bitcoin pricing, adoption, and usage:
Theory and evidence. Stanford University working paper.

Baker, M. and Wurgler, J. (2006). Investor sentiment and the cross-section of stock returns. Journal
of Finance, 61, 1645–1680.

Baker, M. and Wurgler, J. (2007). Investor sentiment in the stock market. Journal of Economic
Perspectives, 21, 129–152.

Balvers, R. and McDonald, B. (2018). Designing a global digital currency. Working paper.

Banerjee, S., Kaniel, R. and Kremer, I. (2009). Price drift as an outcome of differences in higher -

order beliefs. Review of Financial Studies, 22, 3707–3734.

Barberis, N., Shleifer, A. and Vishny, R. (1998). A model of investor sentiment. Journal of Financial
Economics, 49, 307–343.

Bhushan, R. (1989). Firm characteristics and analyst following. Journal of Accounting and Eco-

nomics, 11, 255–274.

Biais, B., Bisi`ere, C., Bouvard, M. and Casamatta, C. (2018). The blockchain folk theorem.

Toulouse School of Economics working paper.

Bikhchandani, S., Hirshleifer, D. and Welch, I. (1992). A theory of fads, fashion, custom, and

cultural change as informational cascades. Journal of Political Economy, 100, 992–1026.

Bodie, Z., Kane, A. and Marcus, A. J. (2014). Investments, 10th ed., McGraw Hill, Ohio.

Böhme R., Christin, N., Edelman, B. and Moore, T. (2015). Bitcoin: Economics, technology, and

governance. Journal of Economic Perspectives, 29, 213–238.

Bolt, W. and van Oordt, M. R. C. (2016). On the value of virtual currencies. Bank of Canada

working paper.

Brock, W., Lakonishok, J. and LeBaron, B. (1992). Simple technical trading rules and the stochastic

properties of stock returns. Journal of Finance, 47, 1731–1764.

Brogaard, J. and Detzel, A. (2015). The asset-pricing implications of government economic policy

uncertainty. Management Science, 61, 3–18.

Brown, D. P. and Jennings, R. H. (1989). On technical analysis. Review of Financial Studies, 2, 527–

551.
46

Electronic copy available at: https://ssrn.com/abstract=3115846

Campbell, J. Y. and Thompson, S. B. (2008). Predicting excess stock returns out of sample: can

anything beat the historical average? Review of Financial Studies, 21, 1509–1531.

Campbell, J. Y. and Yogo, M. (2006). Efficient tests of stock return predictability. Journal of

Financial Economics, 81, 27–60.

Catalini, C. and Gans, J. S. (2017). Some simple economics of the blockchain. MIT working paper.

Cespa, G. and Vives, X. (2012). Dynamic trading and asset prices: Keynes vs. Hayek. Review of

Economic Studies ,79, 539–580.

Chiarella, C., He, X.-Z. and Hommes, C. (2006). A dynamic analysis of moving average rules,

Journal of Economic Dynamics and Control, 30, 1729–1753.

Chiu, J. and Koeppl, T. V. (2017). The economics of cryptocurrencies—bitcoin and beyond. Bank

of Canada working paper.

Choe, H., Kho, B.-C. and Stulz, R. M. (1999). Do foreign investors destabilize stock markets? the

Korean experience in 1997. Journal of Financial Economics, 54, 227 – 264.

Cochrane, J. H. (2005). Asset Pricing. Princeton University Press, Princeton, N.J.

Cochrane, J. H. (2008). The dog that did not bark: a defense of return predictability. Review of

Financial Studies, 21, 1533–1575.

Cochrane, J. H., Longstaff, F. A. and Santa-Clara, P. (2008). Two trees. The Review of Financial

Studies, 21, 347–385.

Cong, L. W. and He, Z. (2018). Blockchain disruption and smart contracts. University of Chicago

working paper.

Cox, J. C. and Huang, C. (1989). Optimal consumption and portfolio policies when asset prices

follow a diffusion process. Journal of Economic Theory, 49, 33–83.

Daniel, K., Hirshleifer, D. and Subrahmanyam, A. (1998). Investor psychology and security market

under- and overreactions. The Journal of Finance, 53, 1839–1885.

DeLong, J. B., Shleifer, A., Summers, L. H. and Waldmann, R. J. (1990). Noise trader risk in

financial markets. Journal of Political Economy, 89, 703–738.

Detemple, J. B. (1986). Asset pricing in a production economy with incomplete information. Journal

of Finance, 41, 383–391.

47

Electronic copy available at: https://ssrn.com/abstract=3115846

Detemple, J. B. (1991). Further results on asset pricing with incomplete information. Journal of

Economic Dynamics and Control, 15, 425–453.

Detzel, A. and Strauss, J. (2018). Combination return forecasts and portfolio allocation with the

cross-section of book-to-market ratios. Review of Finance, 22, 1949–1973.

Dwyer, G. P. (2015). The economics of Bitcoin and similar private digital currencies. Journal of

Financial Stability, 17 (April), 81–91.

Easley, D., O’Hara, M. and Basu, S. (2019). From mining to markets: The evolution of Bitcoin

transaction fees. Journal of Financial Economics, 134(1), 91-109.

Edmans, A., Goldstein, I. and Jiang, W. (2015). Feedback effects, asymmetric trading, and the

limits to arbitrage. American Economic Review, 105 (12), 3766–3797.

Eha, B. P. (2017). How Money Got Free. Oneworld Publications Ltd, London, UK.

Fama, E. F. and French, K. R. (1993). Common risk factors in the returns on stocks and bonds.

Journal of Financial Economics 33, 3–56.

Fernández-Villaverde, J. and Sanches, D. (2017). Can currency competition work? University

of Pennsylvania working paper.

Ferson, W. E., Sarkissian, S. and Simin, T. T. (2003). Spurious regressions in financial economics?

Journal of Finance, 58, 1393–1413.

Foley, S., Karlsen, J. R. and Putnins, T. J. (2018). Sex, drugs, and bitcoin: How much illegal

activity is financed through cryptocurrencies? Working paper.

Gandal, N., Hamrick, J., Moore, T. and Oberman, T. (2018). Price manipulation in the bitcoin

ecosystem, Journal of Monetary Economics, 95 (May), 86-96..

Gennotte, G. (1986). Optimal portfolio choice under incomplete information. Journal of Finance,

41, 733–746.

Goyal, A. and Welch, I. (2008). A comprehensive look at the empirical performance of equity

premium prediction. Review of Financial Studies 21, 1455–1508.

Griffin, J. M. and Shams, A. (2018). Is Bitcoin really un-tethered? University of Texas working

paper.

Grossman, S. (1976). On the efficiency of competitive stock markets where trades have diverse

48

Electronic copy available at: https://ssrn.com/abstract=3115846

 information. The Journal of Finance 31, 573–585.

Han, Y., Yang, K. and Zhou, G. (2013). A new anomaly: The cross-sectional profitability of

technical analysis. Journal of Financial and Quantitative Analysis, 48, 1433–1461.

Han, Y., Zhou, G. and Zhu, Y. (2016). A trend factor: Any economic gains from using information

over investment horizons? Journal of Financial Economics, 122, 352–375.

Harvey, C. R. (2017). Bitcoin myths and facts. Duke University working paper.

Hellwig, M. F. (1980). On the aggregation of information in competitive markets. Journal of

Economic Theory, 22, 477 – 498.

Hong, H., Lim, T. and Stein, J. C. (2000). Bad news travels slowly: Size, analyst coverage, and the

profitability of momentum strategies, The Journal of Finance, 55, 265–295.

Hong, H. and Stein, J. C. (1999). A unified theory of underreaction, momentum trading, and

overreaction in asset markets. Journal of Finance, 54, 2143–2184.

Huang, D., Jiang, F., Tu, J. and Zhou, G. (2015). Investor sentiment aligned: A powerful

predictor of stock returns. Review of Financial Studies, 28, 791–837.

Huang, D., Li, J., Wang, L. and Zhou, G. (2019). Time-series momentum: Is it there?. Journal of

Financial Economics, forthcoming.

Huberman, G., Leshno, J. D. and Moallemi, C. C. (2017). Monopoly without a monopolist: An

economic analysis of the bitcoin payment system. Columbia Business School working paper.

Jegadeesh, N. and Titman, S. (1993). Returns to buying winners and selling losers: Implications

for stock market efficiency. Journal of Finance, 48, 65–91.

Jermann, U. J. (2018). Bitcoin and Cagan’s model of hyperinflation. University of Pennsylvania

SSRN working paper.

Johnson, T. C. (2002). Rational momentum effects. The Journal of Finance, 57, 585–608.

Karpoff, J. M. (1987). The relation between price changes and trading volume: A survey. Journal of

Financial and Quantitative Analysis, 22, 109–126.

Kelly, B. and Pruitt, S. (2013). Market expectations in the cross-section of present values. Journal of

Finance, 68, 1721–1756.

Krueckeberg, S. and Scholz, P. (2018). Cryptocurrencies as an asset class? University of Texas

49

Electronic copy available at: https://ssrn.com/abstract=3115846

 working paper.

Ledoit, O. and Wolf, M. (2008). Robust performance hypothesis testing with the Sharpe ratio.

Journal of Empirical Finance, 15, 850 – 859.

Lewellen, J. and Nagel, S. (2006). The conditional CAPM does not explain asset-pricing anomalies.

Journal of Financial Economics, 82, 289–314.

Liu, Y. and Tsyvinski, A. (2018). Risks and returns of cryptocurrency. NBER working paper.

Lo, A. W. and Hasanhodzic, J. (2009). The heretics of finance: Conversations with leading practi-

tioners of technical analysis. Bloomberg Press.

Lo, A. W., Mamaysky, H. and Wang, J. (2000). Foundations of technical analysis: Computational

algorithms, statistical inference, and empirical implementation. Journal of Finance, 55, 1705–

1770.

Malinova, K. and Park, A. (2017). Market design with Blockchain technology. University of Toronto

working paper.

Marquering, W. and Verbeek, M. (2004). The economic value of predicting stock index returns and

volatility. Journal of Financial and Quantitative Analysis, 39, 407–429.

Moskowitz, T. J., Ooi, Y. H. and Pedersen, L. H. (2012). Time series momentum. Journal of

Financial Economics, 104, 228 – 250.

Neely, C. J., Rapach, D. E., Tu, J. and Zhou, G. (2014). Forecasting the equity risk premium: The

role of technical indicators. Management Science 60, 1772–1791.

Ofek, E. and Richardson, M. (2003). Dotcom mania: The rise and fall of internet stock prices, The

Journal of Finance, 58, 1113 – 1137.

Pagnotta, E. S. and Buraschi, A. (2018). An equilibrium valuation of bitcoin and decentralized

network assets. Imperial College working paper.

Pesaran, M. H. and Timmermann, A. (1995). Predictability of stock returns: Robustness and

economic significance. Journal of Finance, 50, 1201–1228.

Pettenuzzo, D., Timmermann, A. and Valkanov, R. (2014). Forecasting stock returns under economic

constraints, Journal of Financial Economics, 114, 517 – 553.

Prat, J. and Walter, B. (2016). An equilibrium model of the market for bitcoin mining. CESifo

50

Electronic copy available at: https://ssrn.com/abstract=3115846

 working paper.

Rapach, D. E., Strauss, J. K. and Zhou, G. (2010). Out-of-sample equity premium prediction:

combination forecasts and links to the real economy. Review of Financial Studies, 23, 821–862.

Saleh, F. (2017). Blockchain without waste: Proof-of-stake. NYU working paper.

Schwager, J. D. (1989). Market Wizards. John Wiley & Sons, Hoboken, New Jersey.

Stambaugh, R. F. (1999). Predictive regressions. Journal of Financial Economics 54, 375–421.

Timmermann, A. (2006). Forecast combinations, in G. Elliott, C. W. J. Granger and A.

Timmermann, eds. Handbook of Economic Forecasting, Vol. 1, Elsevier, North-Holland,

Amsterdam, chapter 4, 135–196.

Treynor, J. L. and Ferguson, R. (1985). In defense of technical analysis. Journal of Finance 40, 757–

773.

Yermack, D. (2013). Is bitcoin a real currency? An economic appraisal. NBER working paper.

Yermack, D. (2017). Corporate governance and Blockchains. Review of Finance 21 (1), 7–31.

Zhang, X. F. (2006). Information uncertainty and stock returns. Journal of Finance, 61, 105–137.

Zhu, Y. and Zhou, G. (2009). Technical analysis: An asset allocation perspective on the use of

moving averages. Journal of Financial Economics 92, 519–544.

51

Electronic copy available at: https://ssrn.com/abstract=3115846