ed

Are simple technical trading rules profitable in bitcoin markets?

Niek Deprez1, Michael Frömmel2,*

iew
Abstract

This paper examines the profitability of simple technical trading rules in bitcoin markets
comprehensively, by taking into account realistic investor behavior and transaction costs, and data

v
mining problems. Realistic investor behavior is replicated by first employing 75,360 simple technical
trading rules, divided over 6 commonly used trading rule classes and daily and intraday frequencies.

re
Next, we select the best performing rules after transaction costs using a multiple hypothesis procedure.
Finally, we form portfolios combining the selected rules and analyse their out-of-sample performance.
We find that, especially risk-return wise, simple technical trading rules can outperform a buy-and-hold
strategy in the bitcoin market out-of-sample.
er
Keywords: Bitcoin, Technical Analysis, False discovery rate, Intraday

JEL: G11, G14, G17

pe

1
Ghent University, Department of Economics, Sint Pietersplein 5, 9000 Ghent, Belgium,
[email protected]
2
Ghent University, Department of Economics, Sint Pietersplein 5, 9000 Ghent, Belgium,
ot

[email protected]

* Corresponding author
tn

Acknowledgements

For helpful suggestions we would like to thank José Manuel Carbó and participants of the Future Finance
rin

and Economics Association in Rennes (2022) and the 16th International on Computational and Financial
Econometrics in London (2022).
ep
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Are simple technical trading rules profitable in bitcoin markets?

we
Abstract
This paper examines the profitability of simple technical trading rules in bitcoin markets comprehensively, by

ie
taking into account realistic investor behavior and transaction costs, and data mining problems. Realistic
investor behavior is replicated by first employing 75,360 simple technical trading rules, divided over 6
commonly used trading rule classes and daily and intraday frequencies. Next, we select the best performing
rules after transaction costs using a multiple hypothesis procedure. Finally, we form portfolios combining

ev
the selected rules and analyse their out-of-sample performance. We find that, especially risk-return wise,
simple technical trading rules can outperform a buy-and-hold strategy in the bitcoin market out-of-sample.
Keywords: Bitcoin, Technical Analysis, False discovery rate, Intraday
JEL: G11, G14, G17

r
1. Introduction er
In 2008 in the midst of the global financial crisis and growing public distrust in financial institutions
Satoshi Nakamoto (2008) introduced a peer-to-peer electronic cash system, named Bitcoin. Two key charac-
teristics of Bitcoin are that it does not need any trusted third party in order to process electronic payments,
pe
and it has a supply regulated by a fixed, known algorithm, attempting to reduce the influence of financial
intermediaries and central authorities. More than a decade since its inception, Bitcoin is well known by the
general public, mainly due to its exponential increase in value. However, as a relatively new asset without
any apparent intrinsic value, the bitcoin market has been plagued by high volatility, market inefficiencies and
bubbles (Cheah and Fry, 2015; Urquhart, 2016; Bariviera et al., 2017; Nadarajah and Chu, 2017; Kristoufek,
2018; Sensoy, 2019). Following the efficient market hypothesis, these circumstances in the market should
theoretically make bitcoin returns more predictable and the use of technical analysis profitable.
ot

At the time of writing some research has already been done regarding profitability of technical trading
rules in bitcoin markets. First we take a look at the studies using daily price data. Gerritsen et al. (2019)
test 7 trend-following indicators between 2010 and 2019. They find that mainly support-and-resistance
(S&R) rules outperform a buy-and-hold strategy in the bitcoin market. Hudson and Urquhart (2021) test
tn

15,000 technical trading rules using multiple hypothesis procedures between 2010 and 2017, they find that
technical trading improves the risk-adjusted performance in cryptocurrency markets, however they do not
find out-of-sample (OOS) outperformance for bitcoin. Grobys et al. (2020) test a small set of MA rules
between 2016 and 2018 and find return predictability across different cryptocurrencies. Detzel et al. (2021)
test a limited sample of MA strategies between 2010 and 2018, and find return predictability OOS with
rin

significant lower risk than a buy-and-hold strategy. As far as we know, only one study looked at profitability
using high-frequency trading data: Corbet et al. (2019) test a small set of moving average (MA) rules and
S&R rules using different intraday price data between 2014 and 2018. They find that the variable MA
strategy is most profitable.
Together these studies form a comprehensive test for finding profitability of technical trading rules for
different trading frequencies and technical trading rule types separately. However, realistically, investors
ep

combine signals from these different trading rules and frequencies, usually selecting the best performing
rules after transaction costs and after good OOS performance. On top of that, some of these studies do not
take into account transaction costs or data mining, or do not test their results OOS. We add to the existing
literature by combining a wide range of technical trading rules across intraday and daily trading frequencies,
mimicking more realistic investor behavior, and testing their profitability after transaction costs and OOS,
Pr

while taking into account data mining.

Preprint submitted to Elsevier December 13, 2022

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
2. Data

we
We use tick-by-tick transaction data from Bitstamp1 , a online cryptocurrency exchange, which contains
the volume and BTC-USD exchange rate of each trade between 2012 and 2022. After cleaning the data
with the method described in Brownlees and Gallo (2006)2 , we re-sample it at a daily, 60, 30 and 10 minute
frequency. We use the period before 2013 to initialize our technical trading rules and the period afterwards
for our analysis. We convert the prices, Pt , in our sample to logarithmic returns such that:

ie
Pt
rt = ln (1)
Pt−1

ev
Price average maximum minimum

65,000
30,000

r
20,000
10,000

3,000
er
BTC-USD

1,000

250
150
pe
50

10

2012 2014 2016 2018 2020 2022

Time
ot

Figure 1: BTC-USD exchange rate on Bitstamp from 2012 until 2022. We report the monthly average, minimum, and maximum
closing price. The dashed line indicates the start of our sample of technical trading rules. The dotted line indicates the start
of our sample of OOS portfolios.
tn

Next to transaction data we also appropriate ’realistic’ transaction costs by using the average daily
logarithmic bid-ask spread3 , and the minimum transaction fee4 on the Bitstamp exchange. Figure 1 shows
the average, minimum, and maximum monthly BTC-USD exchange rate during our sample period. During
the sample bitcoin has known an astronomical increase from 10 USD to 65,000 USD. This increase, however,
is paired with volatile price swings, most notably at the end of 2013, 2017, and at the end of our sample.
rin

Table 1 shows the summary statistics of the returns for the 4 sampling frequencies and the 0- and 100-BTC-
deep bid-ask spreads. Looking at the returns, we see that the standard deviation is multiples of the mean
return, showing the extreme volatiliy in the BTC market is. More shockingly, however, we see the extreme
minimum and maximum returns for every sampling frequency. We also note that there exist some significant
ep

1 The data is obtained via [dataset] Bitcoincharts.com (2020)

2 First we delete all trades with a 0 or negative price or volume. Then we substitute the price of simultaneous trades with
the median price at that time. Lastly we run their data cleaning algorithm in order to detect outliers, with with δ = 10%,
γ = 1, k = 40.
3 The data is obtained via [dataset] Bitcoinity.org (2020)
4 Bitstamp charges a uniform trading fee of minimum 0,20% per trade before March 2nd 2015 and 0,10% afterwards. In
Pr

2019 they lowered their fees for exceptionally high volumes even further to 0%, but we don’t take this decrease into account in
our paper (Bitstamp, 2012, 2015b, 2019).

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Day 60 Min. 30 Min. 10 Min. 0-spread 100-spread

we
N 3287 78888 157776 473328 3252 3252
Mean 0.248 0.01 0.005 0.002 0.134 1.495
Std. Dev. 4.61 1.053 0.798 0.5 0.124 1.316
Min -66.395 -34.729 -44.287 -29.725 0.014 0.3
Max 33.749 29.888 59.83 73.691 1.505 12.886
Skewness -1.368 -1.372 1.037 7.087 3.136 3.128
Kurtosis 23.171 107.405 408.931 1317.227 16.007 12.249

ie
ρ(1) -0.023 −0.087∗∗∗ −0.105∗∗∗ −0.09∗∗∗
ρ(2) 0.001 −0.006∗ −0.007∗∗∗ −0.041∗∗∗
ρ(5) 0.068∗∗∗ 0.011∗∗∗ -0.004 −0.006∗∗∗
ρ(10) 0.056∗∗∗ 0.007∗ -0.003 -0.002

ev
Table 1: Descriptive statistics of the log return of the BTC/USD exchange rate per sampling frequency and the log bid-ask
spread at the 0- and 100-BTC order book depth, on Bitstamp between 2013 and 2022. The returns and spreads are given in
percentages (e.g. the average of 0.263 for the daily log return = 0.263%). We show the amount of observations (N), the mean,
standard deviation, minimum, maximum skewness and kurtosis. Lastly, ρ(k) is the autocorrelation on the kth lag.

r
autocorrelation on different lags, which may point to return predictability in our sample. Looking at the bid-
ask spreads, we see that the average 0-BTC-deep spread is 0.134%, while the average of the 100-BTC-deep
spread is more than 1%-point higher. These spreads can be quite volatile as well5 .
er
Using data from the Bitstamp exchange has 2 distinct advantages. First Bitstamp is one of the oldest
online trading platform for cryptocurrencies. It was founded in 2011 and still exists today , resulting in
one of the longest available datasets of bitcoin transaction data. Second, Bitstamp is associated with few
incidents and trade disruptions6 , resulting in a relatively continuous data set with little missing data and a
pe
good representation for the overall market price fluctuations. The platform does not provide margin trading
and short-selling during our sample period, which will be taken into account in our study.

3. Methodology

In this paper we attempt to mimic realistic investor behaviour. Technical traders usually use various
ot

types technical trading rules and various trading horizons and frequencies concurrently. By combining the
input from these two dimensions they decide whether to buy or sell a certain asset, giving more importance
to rules with high historical outperformance. They aim to optimize their profits net of costs and OOS.
Keeping this in mind, we use 4 steps in order to determine if technical trading rules are profitable in
tn

bitcoin markets. (1) We determine the investment universe, a set of possible technical trading rules an
investor may consider. (2) For each sub-sample and for every technical trading rule we then calculate the
daily return and appropriate performance measures (3) Using these performance measures, we determine
which rules statistically outperform the benchmark during each sub-period (4) We form a portfolio from the
outperforming technical trading rules and test if it is profitable OOS.
rin

3.1. Technical trading rules

We construct 18,840 (75,360 over all 4 sampling frequencies) individual technical trading rules, subdivided
under 6 commonly used classes of technical trading rules, namely: moving averages (MA), filter rules,
support and resistance (S&R) trading rules, channel breakout (CB) trading rules, on-balance volume averages
(OBV), and the relative strength indicator (RSI). Our selection of technical trading rules is largely based
ep

on the selection of Sullivan et al. (1999) and Hsu et al. (2016). A more in depth discussion of the different
parameterizations can be found in Appendix B.

5 The bid-ask spreads are much higher in the beginning of the sample, but they slowly decrease over time.
Pr

6 In its long existence Bitstamp was only hacked once in 2015, when almost 19,000 Bitcoins were stolen(Bitstamp, 2015a)

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
3.2. Returns and performance measures

we
As mentioned in section 2, Bitstamp does not support short and margin trading. Therefore, we opt for
a in-or-out trading strategy where we take a long position in bitcoin if a buy signal is given, and invest in
the risk-free rate7 otherwise. The return of the j-th technical trading rule is then calculated as follows:
rf
rj,t = Dj,t−1 rb,t + Dj,t−1 rrf,t − Cj,t (2)
(
1, if Dj,t−1 = 0

ie
rf
with Dj,t−1 ∈ {1, 0, 0} and Dj,t−1 = , for j = 1, . . . , J.
0, otherwise

Where rb,t and rrf,t are the logarithmic returns of the BTC-USD exchange rate and the risk-free rate

ev
rf
respectively. Dj,t−1 and Dj,t−1 are the 1-period position in BTC-USD or the risk-free asset of the j-th
technical trading rule respectively. Cj,t is the (logarithmic) trading cost at time t. It consists of a flat
percentage fee and half of the bid-ask spread. After calculating the returns for every rule at every sampling
frequency, all returns are re-sampled to a daily frequency. We then calculate two commonly used performance
metrics φj,k for every j-th rule, namely: the mean excess return (f¯j ) and the sharpe ratio (SRj ):

r
f¯j =
1X
T

T t=1
(rj,t − rb,t ) (3a)
er SRj =
r̄j
σ rj
−
r̄b
σ rb
(3b)
pe
Where rj,t is the return of the j-th technical trading rule, rb,t is the return of the buy-and-hold benchmark,
and σRj , σrb and, σfj are the standard deviation of the return of the j-th rule, the benchmark and the excess
return respectively.

3.3. Outperformance and data snooping

While testing our extensive trading rule universe, there is always a chance that we find highly profitable
rules as a result of luck rather technical trading rule the forecasting ability of that particular rule (i.e. data
ot

mining). Various multiple hypothesis testing methods have been developed to mitigate this problem. These
methods can generally be subdivided in 2 types, namely: controlling the Family wise error rate (FWER) or
the false discovery rate (FDR).
The FWER is defined as the probability of rejecting at least one true null hypothesis. These methods
tn

are useful if you want to limit the probability of erroneously selecting at least one ”lucky” technical trading
rule. However, Romano et al. (2008) argue that controlling FWER can be too conservative, especially when
the universe of rules is large. In practice, the strict control of the FWER is not necessary, as investors
typically combine trading signals of multiple strategies.
As a result it is more beneficial for them to identify a large number of outperforming rules, while
rin

allowing a certain amount of ”lucky” rules. This is exactly what the FDR does. It tries to control the
expected proportion of falsely rejected null hypotheses (Benjamini and Hochberg, 1995). In this paper we
will use the FDR+/− methodology introduced by Barras et al. (2010) and adapted for the estimation of
technical trading rules by Bajgrowicz and Scaillet (2012). After computing the performance metrics φj for
every rule j, 1 ≤ j ≤ J, we test whether it has a statistically positive or negative performance compared to
ep

the benchmark:

H0,j : φj = 0; Ha,j : φj 6= 0 (4)

7 We
Pr

use the US 3-month treasury bill rate obtained via [dataset] OECD (2020) as risk-free rate.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Among the significant rules8 , we define R+ as the amount of rules that have a positive performance,
φj > 0. Among these, F + is the number of significantly positive rules which where chosen by luck. The

we
FDR among the rules yielding positive returns, denoted by FDR+ , is defined as the expected proportion of
the trading rules falsely determined as outperforming. The FDR+ can be estimated as FDR ˆ + = F̂ + , with
R̂+
F̂ + and R̂+ being estimators of F + and R+ (see Appendix C). A FDR+ of k% means that on average k%
of the selected rules do not genuinely outperform the benchmark.

ie
3.4. Out-of-sample portfolios
Starting from January 2014 onwards, every month we determine the best performing rules of the past
year and hold them in our portfolios until the end of the month. This results in a OOS period of 2,922
trading days (8 years). We construct 3 different types of portfolios, namely a portfolio containing: only the

ev
best performing rule, the 50 best performing rules, and a 10%-FDR+ portfolio, where the FDR+ level is
fixed at 10%. With the selected technical trading rules we construct equally weighted portfolios9 .

4. Results

r
Table 2 shows the OOS-performance of our in-or-out strategy portfolios after transaction costs. Panel
(a) shows the general performance of portfolios containing the best performing rule (Best), the 50 best
performing rules (Best 50), and the rules selected by the FDR+ -methodology (FDR-10%). Starting on the
er
left, looking at the total return we see that the 3 mean-return portfolios perform slightly better, in absolute
terms, than the benchmark portfolio. The subdivision between positive and negative periods shows us that
all rules decrease the size of the returns in both directions by about half. This could be an indication
that the trading rules limit risk. Looking at the risk measures we see that this is indeed the case. More
pe
specifically, the maximum drawdown (MDD) and the 95%-VaR (and the skewness) show that our portfolios
are successful in decreasing downside risk. The SR portfolios have the lowest risk, paired with a lower return.
As a result of the lower risk most portfolios also outperform the risk-return measures of the benchmark.
All portfolios have a comparable amount of transactions between 450 and 750 (more than 55 per year) and
average transaction costs of around -0.14% per trade. The break-even-costs (BEC) show us how high a %-
point increase in the transaction costs can be until the strategy has a return of 0 (or equal to the benchmark
ot

between brackets). This ranges from 0.43% to 0.74%. Our outperforming rules only need an increase of
less than 0.05%-points before breaking even with the benchmark. Lastly, the POR shows the percentage
of outperforming rules (in sample). We see that 4.09% of the trading rules, on average, outperformed the
benchmark in terms of mean excess return and only 3.29% in terms of the SR.
tn

Panel (b) shows the performance of the FDR-10% portfolios calculated for the individual trading fre-
quencies. In terms of return and risk-adjusted measures, we see that intraday portfolios outperform the
daily portfolios, showing us that there is indeed some value in trading on higher frequencies. Unexpectedly,
risk-wise the portfolios perform equally. The faster reactions of higher trading frequencies do not result de-
crease or increase in risk compared to the daily trading frequency. However, the higher trading frequencies
execute a substantial larger amount of transactions. As a result the BEC of the daily portfolio is much
rin

higher, making their performance more robust against small increases of transaction costs or the bid-ask
spread. Finally, the POR shows that the FDR-10% methodology finds much less outperforming rules for
the 10 min. frequency. A possible reason is that the larger amount of transactions and costs makes more
trading rules unprofitable.
ep

8 In order to obtain the individual p-values for every rule, we use the stationary bootstrap from Politis and Romano (1994)

with q = 0.1 and N = 1000.

9 Elliott and Timmermann (2008) show that giving a larger weight to the performing best trading rules can decrease the

influence of bad rules, especially when a large amount of rules are selected. Bajgrowicz and Scaillet (2012), however, do mention
that giving a larger weight to the best rules has an effect very similar to reducing the FDR+ target level. Weighting the
technical trading rules by their contribution to the total in-sample return of the portfolio did not change the OOS-performance
Pr

significantly.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 Return Risk Risk-Return Trades and costs
Portfolio Total + - S.D. Skew. Kurt. MDD VaR SR Calmar # Trades Av. TC BEC (Excess) POR
Benchmark 4.15 40.48 -36.33 0.77 -0.81 11.87 -1.80 -6.34% 0.66 0.28 - - - -
(a) General performance
Best
Pr
Mean 4.22 21.42 -17.21 0.52 0.48 9.76 -1.07 -4.02% 0.99 0.48 766.00 -0.14% 0.55% ( 0.01%) -
SR 1.96 11.20 -9.23 0.40 0.70 24.12 -0.97 -2.17% 0.59 0.24 462.00 -0.13% 0.43% ( -0.47%) -
Best 50
Mean 4.54 22.38 -17.84 0.47 0.91 9.01 -0.61 -3.47% 1.20 0.91 723.70 -0.15% 0.63% ( 0.05%) -
ep
SR 2.89 12.31 -9.42 0.31 1.42 19.08 -0.45 -1.93% 1.13 0.79 483.62 -0.14% 0.60% ( -0.26%) -
FDR-10%
Mean 4.28 26.18 -21.89 0.54 0.06 8.04 -1.15 -4.06% 0.98 0.46 577.81 -0.13% 0.74% ( 0.02%) 4.09%
SR 3.65 19.40 -15.74 0.43 0.73 11.64 -0.49 -2.93% 1.04 0.91 713.92 -0.14% 0.51% ( -0.07%) 3.29%
rin
(b) FDR-10% by frequency
Mean
Daily 3.46 24.49 -21.03 0.56 -1.73 33.96 -0.86 -4.23% 0.76 0.49 61.73 -0.14% 5.61% ( -1.10%) 6.18%
60 Min. 4.34 25.28 -20.93 0.53 0.23 10.50 -0.72 -4.07% 1.00 0.74 672.23 -0.15% 0.65% ( 0.03%) 5.79%
30 Min. 4.61 25.49 -20.88 0.54 0.36 9.51 -0.88 -3.94% 1.05 0.65 831.54 -0.14% 0.55% ( 0.06%) 3.41%
tn
10 Min. 4.76 25.97 -21.21 0.53 0.53 6.68 -0.84 -3.95% 1.11 0.69 866.45 -0.13% 0.55% ( 0.07%) 2.22%
SR
Daily 3.18 17.40 -14.22 0.42 0.06 13.96 -0.68 -2.73% 0.91 0.57 96.07 -0.15% 3.31% ( -1.01%) 4.81%
60 Min. 3.86 20.00 -16.13 0.45 0.78 15.63 -0.62 -3.19% 1.06 0.76 713.94 -0.14% 0.54% ( -0.04%) 5.94%
30 Min. 3.51 19.59 -16.08 0.44 1.06 11.77 -0.56 -3.01% 0.97 0.77 1021.58 -0.15% 0.34% ( -0.06%) 2.87%
ot
10 Min. 3.77 20.55 -16.78 0.48 0.73 10.25 -0.75 -3.32% 0.97 0.61 1128.40 -0.14% 0.33% ( -0.03%) 0.83%

6
(c) FDR-10% by technical trading rule
Mean
MA 3.78 23.08 -19.30 0.51 0.27 8.12 -1.25 -3.87% 0.91 0.37 424.70 -0.17% 0.89% ( -0.09%) 4.13%
pe
RSI 1.48 11.17 -9.69 0.35 -3.72 97.66 -0.86 -1.78% 0.51 0.21 48.95 -0.13% 3.03% ( -5.43%) 4.19%
S&R 3.39 19.42 -16.02 0.44 0.46 10.45 -0.69 -3.28% 0.94 0.60 322.65 -0.16% 1.05% ( -0.23%) 5.02%
Filter 3.30 20.44 -17.14 0.45 -0.08 12.08 -0.87 -3.50% 0.89 0.47 164.27 -0.12% 2.01% ( -0.51%) 5.37%
CB 3.70 23.92 -20.23 0.52 0.18 8.71 -0.87 -3.97% 0.88 0.52 198.31 -0.15% 1.86% ( -0.23%) 5.58%
OBV 4.80 27.94 -23.14 0.55 0.47 6.85 -0.75 -4.22% 1.07 0.79 734.40 -0.15% 0.65% ( 0.09%) 4.22%
er
SR
MA 3.46 21.36 -17.90 0.48 0.39 9.50 -0.59 -3.60% 0.88 0.71 577.59 -0.16% 0.60% ( -0.12%) 4.02%
RSI 1.11 10.26 -9.16 0.30 -0.46 22.11 -0.53 -1.92% 0.43 0.24 73.46 -0.12% 1.51% ( -4.14%) 1.99%
S&R
Filter
1.87
2.36
14.94
14.69
-13.07
-12.33
0.43
0.43
-3.10
-2.71
89.71
92.89
-1.12
-1.03
r
-2.64%
-2.36%
0.52
0.66
0.20
0.28
345.39
249.86
-0.14%
-0.11%
0.54% ( -0.66%)
0.95% ( -0.71%)
2.68%
3.61%
CB 2.55 15.07 -12.52 0.38 0.66 19.50 -0.55 -2.53% 0.81 0.56 237.78 -0.14% 1.07% ( -0.67%) 2.27%
OBV
ev
4.85 23.30 -18.45 0.48 1.05 9.43 -0.60 -3.38% 1.25 1.00 966.67 -0.15% 0.50% ( 0.07%) 6.29%

Table 2: Out-of-sample performance (after transaction costs) for different in-or-out strategies. Four types of measures are shown for every strategy: return, risk,
ie
risk-return measures, and trades and costs. The return measures are the total (logarithmic) return, where a total return of 4.15 = 415% (or e4.15 − 1 = 6243.4%),
and the total return during positive (+) and negative (+) periods in the benchmark. The risk measures are the annualized (ann.) standard deviation, skewness,
kurtosis, maximum drawdown (MDD) and the 95% value-at-risk respectively. As risk-return measures we consider the ann. sharpe (SR) and Calmar ratios. As
trades and costs measures we show the amount of trades (#), the average cost of 1 transaction (Av. TC), and the break-even-transaction costs (BEC). The BEC
shows how high a %-point increase in transaction costs can be until the total return of the strategy is 0 (or equal to the benchmark between brackets). Lastly, POR
is the proportion of the trading rules, which were considered for the strategy, that are selected as statistically outperforming by the FDR-10% methodology. The
we
table shows the performance of the Benchmark, Best, Best 50, FDR-10% portfolios, and the FDR-10% portfolios calculated for the individual trading frequencies
and trading rule types, for each performance measure.
d

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 (a) General performance

Portfolio Benchmark Best Best 50 FDR10

Mean SR

d
4

we
Return

ie
2014 2016 2018 2020 2022 2014 2016 2018 2020 2022
Time

ev
(b) FDR-10% by frequency

Benchmark Daily 30 Min.

Portfolio
All 60 Min. 10 Min.

r
Mean SR

4
er
Return

2
pe
0

2014 2016 2018 2020 2022 2014 2016 2018 2020 2022
Time
ot

(c) FDR-10% by type

Benchmark MA Filter CB
Portfolio
All RSI SR OBV
tn

Mean SR

4
rin
Return

0
ep

2014 2016 2018 2020 2022 2014 2016 2018 2020 2022
Time

1
Pr

Figure 2: Out-of-sample cumulative return of our strategies against the benchmark over time, after transaction costs. Panel
(a) shows the cumulative returns (expressed in log returns) for both performance measures, for the benchmark portfolio, the
portfolio consisting of only the best performing TTR (best), the portfolio consisting of the best 50 performing TTRs, and
the FDR-10% portfolio. Panel (b) and (c) show the cumulative returns for the FDR-10% portfolios calculated for the unique
trading frequencies and technical trading rule types respectively.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Panel (c) shows the performance of the FDR-10% portfolios calculated for the individual types of technical
trading rules. Here we see that mainly the MA and the OBV rules outperform all other rules in terms of

we
absolute return measures and risk-return measures, in case of the OBV rules, even though their POR is not
much different than the other rules. These MA and OBV rules also have a larger amount of trades than the
other rules. The RSI rules are the worst performers.
Figure 2 show the cumulative returns of all these portfolios over time. Panel (a) show the cumulative
returns for the general performance portfolios. It shows clearly that our trading rules manage to limit losses
during the downturns in 2014-2015 and 2018-2019. The mean-return portfolios also can predict the bullish

ie
periods well, which results into a similar total return, with less risk compared to the benchmark. Panel (b)
shows the cumulative returns of the FDR-10% portfolios calculated for the individual trading frequencies.
Here we see that the portfolios trading on a daily frequency react slower to the bullish periods in 2016-2017
and 2020-2021, leading them to have a lower overall return. Panel (c) shows the cumulative return of the

ev
FDR-10% portfolios calculated for the individual types of technical trading rules. Most trading rules limit
the downside risk during the downturn. The outperformance of the OBV rules can be clearly seen. Lastly,
we note the bad performance of the RSI rules. As a trend reversal trading rule it is seemingly not able to
predict the bull market in 2017, and stays out of the market for this whole period (which results in a flat
cumulative return for the Mean RSI portfolio), which results in a underperformance in comparison to the

r
benchmark and other trading rules.

5. Discussion
er
Although we show that it is possible to outperform, in terms of absolute return or risk-adjusted perfor-
mance measures, a buy-and-hold strategy in bitcoin using simple technical trading rules, certain remarks
pe
about our results can be made.
First, because we do not allow shorting, our technical trading rules may not be able to profit overtly
from price decreases. Shorting is not possible on Bitstamp, but may have been possible on other bitcoin
exchanges during our sample. Usually shorting has extra costs in form of a borrow rate. For our results
we used a borrow rate of 0.02% for every 4 hours the position is opened10 . In table A.3 in Appendix A,
panel (a), we show the general performances of our portfolios with shorting allowed11 . We see that in
comparison with the in-or-out strategies all strategies have higher absolute returns, paired with an increase
ot

of the (downside) risk. As a result most portfolios also outperform the benchmark in terms of risk-return
measures, indicating that our technical trading rules indeed can profit from price decreases. However, the
extra borrow rate results in a increase of the transaction costs, tempering the increase of the absolute
returns, making the risk-return measures comparable to the portfolios of our in-or-out strategies. Panel (b)
tn

shows the subdivision by trading frequency. Here we see that intraday frequencies (mainly the 30 min.) still
perform significantly better than the daily frequency, showing that intraday portfolios profit from shorting.
The transaction costs show us that the daily portfolios also suffer from higher carrying costs for daily short
positions, reducing their overall return. panel (c) shows the subdivision by technical trading rule. as was
the case for the in-or-out strategy, the OBV rules outperform the other technical trading rules. Lastly, the
rin

risk measures of the general FDR-10% and Best 50 portfolios are generally lower than those of the other
portfolios, indicating that there may be some diversification benefits.
Second, in our approximation of realistic transaction costs, we use the lowest possible trading fee and
bid-ask spread. Taking into account that the transaction fee on Bitstamp is volume based12 , in reality there
should be a trade-off between the transaction fees and the bid-ask spread, where a low(high) transaction
volume leads to high(low) fees per transaction, but a low(high) bid-ask spread. Knowing that there can be
ep

a difference of up to 0.50%-points in transaction fees and of 1%-points on average between the 0-BTC and
100-BTC-deep bid-ask spread, this could have a significant impact on our result, especially after seeing the

10 This is the borrow rate/margin fee currently used on the Kraken.com exchange.
Pr

11 With Dj,t−1 ∈ {1, 0, −1}.

12 A higher trading volume results in lower transaction fees and vice versa.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
BEC in table 2. Table A.4 in Appendix A shows the general performance of the in-or-out and shorting-
allowed portfolios recalculated for the 2 extreme cases. Namely, the case with the highest13 bid-ask spread

we
and the lowest transaction fees, and the case with the lowest bid-ask-spread and the highest transaction fees.
Looking at the average transaction costs, we see that they more than quintuple due to a higher bid-ask and
triple due higher fees. As a result all selected outperforming rules execute markedly less trades. Because of
the higher costs, our portfolios performed worse than in our ideal case, however most of the in-or-out and
some of the shorting-allowed portfolios still managed to outperform the benchmark risk-return wise. The
interpretation of the risk measures also has not changed. Surprisingly, the lower amount of trades, leads to

ie
a much higher BEC for all our portfolios, making their performance more robust to increases in transaction
costs. What is not shown in the table is that because of the higher transaction costs, the portfolios trading
on daily frequencies now perform comparable or better than the intraday portfolios.
Third, we know from theory that in order to avoid data mining when selecting outperforming technical

ev
trading rules, we should construct a large enough technical trading rule universe. However there is a trade-
off, as using too many technical trading rules can also decrease the power of our statistical tests. In our study
we used a substantial amount of 75,360 technical trading rules, which could have influenced our ability to
select truly outperforming rules. Table A.5 in Appendix A, panel (a), replicates the general results of table
2 using only 1/6th of the technical trading rules. This amounts to 12,340 technical trading rules analysed,

r
which is more comparable to the amount of technical trading rules used in other studies about technical
trading rule performance. However, looking at the table we see that using a smaller investment universe
er
produces very similar results.
Fourth, every month we select the best performing technical trading rules of the past year in order to
form our portfolios. An argument can be made, however, that a shorter forming period will increase profits,
as the effectiveness of specific technical trading rules is time-varying. panel (b) and (c) of table 2 shows the
general performance of portfolios formed with a 6- and 3-month forming period respectively. We see that
pe
most of these portfolios still outperform the benchmark risk-return wise, and generally underperform the
12-month forming period portfolios. The longer forming period probably results in a more stable prediction
for finding truly outperforming trading rules in the volatile BTC market.
Finally, although it is not possible on Bitstamp, margin trading could also be an alternative to the
shorting-allowed and in-or-out strategy. Our margin trading strategy is as follows: if a neutral signal is
given, go long in bitcoin. If a buy signal is given, borrow money at a fixed fee and double your position,
ot

on the other hand if a sell signal is given, liquidate the position and buy the risk-free asset14 . The results
are shown in panel (d) from table A.5 in Appendix A. We see that all portfolios largely outperform all
earlier discussed portfolios in terms of absolute returns. These large returns are however paired with an
increased risk. Interestingly, although the margin trading strategies still outperform the benchmark risk-
tn

return wise, the increased risk results in most of the in-or-out and shorting-allowed strategies having a better
risk-adjusted performance.

6. Conclusion
rin

In this paper we study if the use of simple technical trading rules can outperform buying and holding
Bitcoin. We find that our rules indeed can outperform the buy-and-hold strategy out-of-sample, especially
risk-return wise. More specifically we find that technical trading rules can be very successful in limiting
downside risk and large drawdowns. Additionally we find that combining multiple technical trading rules,
across rule type and trading frequency, has diversification benefits. Although trading on higher frequencies
usually increases returns, it makes the portfolio more vulnerable to small and unexpected increases in
ep

transaction costs. Trading on daily frequency is much more robust to these increases. We also find that
mainly the on-balance volume and the moving average rules performed better than the other technical

13 The 100-BTC-deep bid-ask spread is the deepest spread that was available to us.
Pr

14 With Dj,t−1 ∈ {2, 1, 0}. We set the margin trading fee at 0.02% for every 4 hours the position is opened, as was the case
with the borrow rate in the shorting-allowed portfolios.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
trading rules. Lastly, we find that allowing shorting and margin trading increases performance, however,
this is paired with a increase in (downside) risk and transaction costs.

ie we
r ev
er
pe
ot
tn
rin
ep
Pr

10

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 Appendix A. Extra tables

Return Risk Risk-Return Trades and costs

Portfolio Total + - S.D. Skew. Kurt. MDD VaR SR Calmar # Trades Av. TC BEC (Excess) POR
Pr
Benchmark 4.15 40.48 -36.33 0.77 -0.81 11.87 -1.80 -6.34% 0.66 0.28 - - - -
(a) General performance
Best
Mean
ep
4.52 6.97 -2.44 0.75 0.45 9.63 -1.23 -5.82% 0.74 0.45 1027.00 -0.22% 0.44% ( 0.04%) -
SR 1.94 6.32 -4.38 0.60 0.84 21.82 -1.19 -4.30% 0.39 0.20 1028.00 -0.16% 0.19% ( -0.21%) -
Best 50
Mean 5.31 10.41 -5.10 0.56 1.31 14.58 -0.56 -3.74% 1.18 1.17 1217.62 -0.22% 0.44% ( 0.10%) -
SR 3.96 7.01 -3.05 0.42 1.93 21.04 -0.44 -2.59% 1.16 1.10 1032.88 -0.19% 0.38% ( -0.02%) -
FDR-10%
rin
Mean 4.98 15.52 -10.54 0.59 0.32 6.56 -0.95 -4.36% 1.03 0.65 934.84 -0.20% 0.53% ( 0.09%) 2.85%
SR 5.04 12.83 -7.79 0.50 0.78 10.51 -0.65 -3.42% 1.24 0.95 1045.89 -0.17% 0.48% ( 0.09%) 2.22%
(b) FDR-10% by frequency
Mean
Daily 3.65 12.12 -8.46 0.61 -1.14 25.04 -1.20 -4.63% 0.73 0.37 85.58 -1.04% 4.27% ( -0.58%) 4.03%
tn
60 Min. 4.28 14.44 -10.15 0.61 -0.79 25.82 -0.87 -4.29% 0.86 0.61 959.64 -0.21% 0.45% ( 0.01%) 4.31%
30 Min. 5.72 15.53 -9.81 0.62 0.81 8.02 -0.85 -4.61% 1.14 0.83 1271.02 -0.17% 0.45% ( 0.12%) 2.59%
10 Min. 5.75 15.45 -9.70 0.63 0.94 10.78 -1.09 -4.42% 1.12 0.65 1188.27 -0.20% 0.48% ( 0.14%) 1.71%
SR
Daily 2.29 7.48 -5.20 0.36 0.10 13.21 -0.77 -2.35% 0.77 0.36 93.17 -0.60% 2.45% ( -1.99%) 3.44%
60 Min. 4.37 11.79 -7.42 0.54 -1.14 40.17 -0.99 -3.72% 0.99 0.54 952.44 -0.19% 0.46% ( 0.02%) 3.65%
ot
30 Min. 6.66 14.23 -7.57 0.57 1.15 12.51 -0.71 -3.99% 1.44 1.16 1238.04 -0.16% 0.54% ( 0.20%) 2.08%
10 Min. 4.02 12.70 -8.68 0.60 0.48 13.94 -1.12 -4.09% 0.82 0.44 1219.52 -0.21% 0.33% ( -0.01%) 0.87%

11
(c) FDR-10% by technical trading rule
Mean
MA 3.50 9.23 -5.73 0.63 0.51 6.97 -1.07 -4.76% 0.68 0.40 942.24 -0.29% 0.37% ( -0.07%) 2.63%
pe
RSI 0.02 2.73 -2.70 0.42 -3.64 103.38 -1.67 -2.34% -0.01 0.00 59.69 -1.55% 0.04% ( -6.90%) 3.50%
S&R 1.92 3.78 -1.86 0.58 0.96 13.44 -1.32 -4.20% 0.40 0.18 562.49 -0.29% 0.34% ( -0.40%) 3.28%
Filter 2.28 9.38 -7.10 0.44 0.63 14.96 -0.71 -2.85% 0.63 0.39 247.52 -0.34% 0.92% ( -0.75%) 4.63%
CB 2.72 10.52 -7.79 0.58 -0.23 17.88 -0.85 -4.21% 0.57 0.39 226.99 -0.49% 1.20% ( -0.63%) 3.85%
OBV 6.44 18.13 -11.69 0.63 0.35 8.41 -0.76 -4.58% 1.26 1.05 1078.24 -0.19% 0.60% ( 0.21%) 3.33%
SR
er
MA 4.11 7.19 -3.07 0.58 0.88 8.34 -1.20 -4.33% 0.87 0.42 943.03 -0.27% 0.44% ( 0.00%) 2.59%
RSI 0.19 2.25 -2.06 0.24 -0.17 42.14 -0.71 -1.04% 0.06 0.02 70.96 -0.68% 0.26% ( -5.58%) 2.71%
S&R 1.85 2.18 -0.33 0.39 5.90 130.64 -0.63 -1.95% 0.57 0.35 460.61 -0.21% 0.40% ( -0.50%) 2.38%
Filter
CB
OBV
3.50
1.06
6.16
8.12
5.12
15.50
-4.62
-4.05
-9.34
0.35
0.34
0.58
1.26
-0.33
0.74
21.91
15.96
10.86
-0.49
-0.43
-0.57
r
-1.89%
-2.25%
-4.03%
1.23
0.36
1.31
0.88
0.28
1.33
205.87
311.60
1336.38
ev -0.21%
-0.24%
-0.19%
1.70% ( -0.31%)
0.34% ( -0.99%)
0.46% ( 0.15%)
3.34%
2.40%
3.52%

Table A.3: Out-of-sample performance (after transaction costs) for the portfolios where shorting is allowed. Four types of measures are
shown for every strategy: return, risk, risk-return measures, and trades and costs. The return measures are the total (logarithmic) return,
where a total return of 4.15 = 415% (or e4.15 − 1 = 6243.4%), and the total return during positive (+) and negative (+) periods in the
ie
benchmark. The risk measures are the annualized (ann.) standard deviation, skewness, kurtosis, maximum drawdown (MDD) and the 95%
value-at-risk respectively. As risk-return measures we consider the ann. sharpe (SR) and Calmar ratios. As trades and costs measures we
show the amount of trades (#), the average cost of 1 transaction (Av. TC), and the break-even-transaction costs (BEC). The BEC shows how
high a %-point increase in transaction costs can be until the total return of the strategy is 0 (or equal to the benchmark between brackets).
Lastly, POR is the proportion of the trading rules, which were considered for the strategy, that are selected as statistically outperforming
we
by the FDR-10% methodology. The table shows the performance of the Benchmark, Best, Best 50, FDR-10% portfolios, and the FDR-10%
portfolios calculated for the individual trading frequencies and trading rule types, for each performance measure. d

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 Return Risk Risk-Return Trades and costs
Portfolio Total + - S.D. Skew. Kurt. MDD VaR SR Calmar # Trades Av. TC BEC (Excess) POR
Benchmark 4.15 40.48 -36.33 0.77 -0.81 11.87 -1.80 -6.34% 0.66 0.28 - - - -
(a) in-or-out, 100-BTC deep Bid-Ask, lowest fees
Pr
Best
Mean 3.31 23.85 -20.55 0.56 -0.45 9.28 -1.34 -4.60% 0.72 0.30 108.00 -0.98% 3.06% ( -0.78%) -
SR 1.87 12.47 -10.60 0.43 -0.06 28.63 -0.94 -2.41% 0.52 0.24 85.00 -0.71% 2.21% ( -2.67%) -
Best 50
Mean 3.47 22.70 -19.24 0.49 -0.02 8.59 -0.91 -3.98% 0.86 0.47 111.32 -0.85% 3.12% ( -0.61%) -
ep
SR 2.77 13.57 -10.80 0.33 0.61 16.69 -0.60 -2.09% 1.01 0.56 74.08 -0.78% 3.74% ( -1.86%) -
FDR-10%
Mean 3.81 25.49 -21.68 0.56 -1.65 33.78 -1.23 -4.09% 0.84 0.38 110.19 -0.91% 3.46% ( -0.30%) 2.12%
SR 3.08 17.94 -14.87 0.43 0.08 12.24 -0.68 -3.09% 0.87 0.55 108.10 -0.79% 2.85% ( -0.99%) 0.80%
(b) in-or-out, 0-BTC deep Bid-Ask, highest fees
rin
Best
Mean 3.33 23.37 -20.04 0.57 0.00 10.04 -1.28 -4.83% 0.72 0.32 273.00 -0.40% 1.22% ( -0.30%) -
SR 1.45 12.63 -11.17 0.42 -0.12 18.30 -0.78 -2.70% 0.41 0.22 208.00 -0.38% 0.70% ( -1.29%) -
Best 50
Mean 3.63 22.59 -18.96 0.48 0.34 8.23 -0.85 -3.78% 0.92 0.52 274.82 -0.39% 1.32% ( -0.19%) -
tn
SR 2.84 13.28 -10.44 0.32 0.93 17.40 -0.54 -1.95% 1.07 0.64 137.90 -0.39% 2.06% ( -0.95%) -
FDR-10%
Mean 4.22 27.64 -23.41 0.56 -0.25 8.46 -1.00 -4.58% 0.92 0.52 231.76 -0.35% 1.82% ( 0.03%) 2.76%
SR 2.41 19.53 -17.12 0.46 -0.04 11.84 -0.83 -3.42% 0.63 0.35 302.92 -0.38% 0.79% ( -0.57%) 1.59%
(c) shorting-allowed, 100-BTC deep Bid-Ask, lowest fees
ot
Best
Mean 2.71 11.29 -8.58 0.75 -0.14 5.13 -1.83 -6.31% 0.44 0.18 179.00 -1.15% 1.51% ( -0.80%) -

12
SR -0.94 3.82 -4.76 0.53 -0.26 13.46 -2.24 -4.40% -0.24 -0.06 193.00 -0.55% -0.49% ( -2.64%) -
Best 50
Mean 1.77 10.34 -8.57 0.51 0.47 7.74 -0.95 -3.94% 0.41 0.22 266.78 -0.97% 0.66% ( -0.89%) -
SR 1.79 8.00 -6.21 0.37 0.91 14.70 -0.57 -2.67% 0.58 0.38 157.88 -0.85% 1.13% ( -1.49%) -
pe
FDR-10%
Mean 3.56 16.74 -13.18 0.58 0.09 6.41 -0.97 -4.69% 0.75 0.45 237.26 -0.84% 1.50% ( -0.25%) 1.51%
SR 1.45 10.68 -9.24 0.43 0.04 9.65 -1.00 -3.31% 0.39 0.17 235.18 -0.59% 0.62% ( -1.15%) 0.95%
(d) shorting-allowed, 0-BTC deep Bid-Ask, highest fees
Best
er
Mean 2.27 5.60 -3.33 0.76 -0.10 5.65 -1.88 -6.10% 0.36 0.15 440.00 -0.54% 0.52% ( -0.43%) -
SR -0.98 4.60 -5.58 0.56 -0.19 11.85 -2.10 -4.51% -0.23 -0.06 340.00 -0.37% -0.29% ( -1.51%) -
Best 50
Mean
SR
3.12
2.75
9.87
7.11
-6.74
-4.36
0.52
0.38
0.70
1.44
7.56
16.55
-0.71
-0.42
r
-3.81%
-2.61%
0.73
0.89
0.54
0.79
631.74
458.12
-0.47%
-0.41%
0.49% ( -0.16%)
0.60% ( -0.31%)
-
-
FDR-10%
Mean 5.66 17.95 -12.29 0.59 0.30 6.92 -0.88 -4.51% 1.18 0.80 420.22 -0.47% 1.35% ( 0.36%) 1.91%
ev
SR 2.73 13.90 -11.17 0.48 -0.62 17.15 -0.86 -3.33% 0.70 0.38 454.61 -0.37% 0.60% ( -0.31%) 1.37%

Table A.4: Out-of-sample performance (after transaction costs) for the in-or-out and shorting-allowed strategies calculated with either a 100-BTC
ie
deep bid-ask spread or the highest possible trading fees on the platform. Four types of measures are shown for every strategy: return, risk,
risk-return measures, and trades and costs. The return measures are the total (logarithmic) return, where a total return of 4.15 = 415% (or
e4.15 − 1 = 6243.4%), and the total return during positive (+) and negative (+) periods in the benchmark. The risk measures are the annualized
(ann.) standard deviation, skewness, kurtosis, maximum drawdown (MDD) and the 95% value-at-risk respectively. As risk-return measures
we consider the ann. sharpe (SR) and Calmar ratios. As trades and costs measures we show the amount of trades (#), the average cost of 1
we
transaction (Av. TC), and the break-even-transaction costs (BEC). The BEC shows how high a %-point increase in transaction costs can be until
the total return of the strategy is 0 (or equal to the benchmark between brackets). Lastly, POR is the proportion of the trading rules, which were
considered for the strategy, that are selected as statistically outperforming by the FDR-10% methodology. The table shows the performance of
the Benchmark, Best, Best 50, FDR-10% portfolios.
d

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 Return Risk Risk-Return Trades and costs
Portfolio Total + - S.D. Skew. Kurt. MDD VaR SR Calmar # Trades Av. TC BEC (Excess) POR
Benchmark 4.15 40.48 -36.33 0.77 -0.81 11.87 -1.80 -6.34% 0.66 0.28 - - - -
(a) in-or-out, 1/6th investment universe
Pr
Best
Mean 3.95 20.70 -16.74 0.51 0.54 9.08 -1.10 -3.88% 0.96 0.44 582.00 -0.14% 0.68% ( -0.03%) -
SR 2.01 8.89 -6.88 0.35 1.46 32.62 -0.63 -1.70% 0.69 0.38 474.00 -0.13% 0.42% ( -0.45%) -
Best 50
Mean 4.37 23.32 -18.95 0.48 0.71 7.99 -0.75 -3.58% 1.12 0.72 637.54 -0.16% 0.69% ( 0.04%) -
ep
SR 3.36 15.36 -12.00 0.36 0.90 14.66 -0.44 -2.26% 1.15 0.94 528.02 -0.14% 0.64% ( -0.15%) -
FDR-10%
Mean 4.45 24.73 -20.28 0.52 0.28 7.29 -0.75 -3.97% 1.05 0.73 596.20 -0.14% 0.75% ( 0.05%) 4.36%
SR 3.90 19.95 -16.05 0.44 1.12 12.63 -0.65 -3.04% 1.08 0.74 764.85 -0.14% 0.51% ( -0.03%) 3.53%
(b) in-or-out, 6-month forming period
rin
Best
Mean 4.64 19.36 -14.71 0.50 -0.27 16.01 -0.76 -3.36% 1.15 0.75 930.00 -0.13% 0.50% ( 0.05%) -
SR 2.57 10.02 -7.45 0.37 1.29 31.44 -1.24 -1.99% 0.84 0.25 691.00 -0.13% 0.37% ( -0.23%) -
Best 50
Mean 3.68 20.58 -16.91 0.42 0.91 9.01 -0.76 -3.09% 1.06 0.59 925.30 -0.13% 0.40% ( -0.05%) -
tn
SR 2.29 11.35 -9.06 0.29 1.60 19.34 -0.43 -1.86% 0.96 0.64 605.56 -0.13% 0.38% ( -0.31%) -
FDR-10%
Mean 2.72 24.72 -21.99 0.53 -0.33 8.89 -0.92 -3.87% 0.62 0.36 699.49 -0.13% 0.39% ( -0.20%) 3.59%
SR 4.09 19.31 -15.22 0.41 1.07 10.95 -0.54 -2.84% 1.23 0.93 788.20 -0.13% 0.52% ( -0.01%) 2.72%
(c) in-or-out, 3-month forming period
ot
Best
Mean 3.25 19.00 -15.75 0.51 -0.28 13.93 -0.91 -3.85% 0.78 0.44 1410.00 -0.11% 0.23% (-0.06%) -

13
SR 1.54 11.92 -10.38 0.44 -3.37 85.10 -1.98 -2.73% 0.42 0.09 1028.00 -0.11% 0.15% (-0.25%) -
Best 50
Mean 3.84 20.72 -16.88 0.43 0.77 7.55 -0.79 -3.22% 1.10 0.59 1018.38 -0.13% 0.38% (-0.03%) -
pe
SR 2.10 11.90 -9.80 0.29 0.70 14.91 -0.44 -1.97% 0.86 0.57 767.68 -0.12% 0.27% (-0.27%) -
FDR-10%
Mean 4.32 25.87 -21.55 0.54 -0.08 9.73 -0.90 -4.03% 0.98 0.59 693.51 -0.12% 0.62% ( 0.02%) 5.92%
SR 3.30 17.53 -14.23 0.40 0.77 14.60 -0.58 -2.57% 1.02 0.69 841.34 -0.13% 0.39% (-0.10%) 3.38%
(d) margin trading strategy
er
Best
Mean 8.94 47.62 -38.68 1.09 0.17 9.66 -2.04 -8.28% 1.01 0.54 1075.00 -0.33% 0.83% ( 0.45%) -
SR 5.21 39.42 -34.20 1.03 0.43 15.13 -1.97 -7.58% 0.62 0.33 810.00 -0.33% 0.64% ( 0.13%) -
Best 50
Mean
SR
8.36
7.17
48.13
43.53
-39.76
-36.36
0.98
0.92
0.91
1.03
8.23
10.10
-1.42
-1.14
r
-7.10%
-6.62%
1.05
0.97
0.73
0.78
1289.02
1080.10
-0.31%
-0.30%
0.65% ( 0.33%)
0.66% ( 0.28%)
-
-
FDR-10%
Mean 6.50 46.07 -39.57 0.89 0.69 6.90 -1.56 -6.83% 0.90 0.52 796.76 -0.34% 0.82% ( 0.30%) 7.76%
ev
SR 5.79 42.08 -36.29 0.82 0.54 7.31 -1.20 -6.30% 0.87 0.60 750.27 -0.28% 0.77% ( 0.22%) 1.95%

Table A.5: Out-of-sample performance (after transaction costs) for the in-or-out strategy (a) calculated for a investment universe consisting of
ie
only 12,340 trading rules (1/6th of the original amount), (b,c) where the outpeforming rules of the past (b) 6-months/( (c) quarter) are selected,
(d) and for the margin trading strategy. Four types of measures are shown for every strategy: return, risk, risk-return measures, and trades
and costs. The return measures are the total (logarithmic) return, where a total return of 4.15 = 415% (or e4.15 − 1 = 6243.4%), and the
total return during positive (+) and negative (+) periods in the benchmark. The risk measures are the annualized (ann.) standard deviation,
skewness, kurtosis, maximum drawdown (MDD) and the 95% value-at-risk respectively. As risk-return measures we consider the ann. sharpe
we
(SR) and Calmar ratios. As trades and costs measures we show the amount of trades (#), the average cost of 1 transaction (Av. TC), and the
break-even-transaction costs (BEC). The BEC shows how high a %-point increase in transaction costs can be until the total return of the strategy
is 0 (or equal to the benchmark between brackets). Lastly, POR is the proportion of the trading rules, which were considered for the strategy,
that are selected as statistically outperforming by the FDR-10% methodology. The table shows the performance of the Benchmark, Best, Best
50, FDR-10% portfolios.
d

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Appendix B. Technical Trading Rules

we
We compare a total of 18,840 technical trading rules divided over 6 commonly used classes and do this
for 4 sampling frequencies, summing to 75,360 technical trading rules in total. In the following sections
we will shortly15 describe how we implement each of these classes of technical trading rules. Table B.6 in
section Appendix B.7 shows all the used parameters for every calculated technical trading rule.

Appendix B.1. Moving averages

ie
A simple16 moving average (MA) with length j at time t is calculated as the arithmetic mean of the j
last prices up until time t. Mathematically we define it as follows:
j−1

ev
1X
MAt (j) = Pt−i (B.1)
j i=0

In the overlapping MA trading strategy, a long (short) position is taken at time t if the short MA is higher
(lower) or equal than the long MA at time t. This gives following trading signals for our MA strategies:

r
(
1, for MAt (p) ≥ MAt (q)
DMA,t+1 = (B.2)
−1, for MAt (p) < MAt (q)
er
with p < q.

Appendix B.2. Relative strength indicator

The relative strength indicator (RSI) is a type of oscillator that takes values between 0 and 100. Traders
pe
use these values in order to determine oversold and overbought periods for a particular security. We calculate
the RSI as follows:
100
RSIt (h) = 100 − (B.3)
1 + RSt (h)
Ut (h)
with RS = Dt (h)
ot

Where Ut (h) and Dt (h) are the average number of periods with a positive or negative return in the last h
periods, respectively. The RSI at time t indicates an overbought(oversold) period if it is v units higher(lower)
than 5017 . A buy(sell) signal is given when the oscillator moves out of the oversold(overbought) region. This
tn

results in following trade signals:


 for RSIt−1 ≤ 50 − v

 1,
and RSIt > 50 − v




DRSI,t+1 = for RSIt−1 ≥ 50 + v (B.4)
-1,
rin


and RSIt < 50 + v





0, otherwise

15 Fora more extensive description we refer the reader to Sullivan et al. (1999) and Hsu et al. (2016) inter alia.
16 Other types of MAs exist. Some of the more commonly used are: exponential (EMA), weighted (WMA) and filtered MAs.
ep

17 Wilder (1978) suggests a lower bound value of 30 and an upper bound value of 70, corresponding with a v of 20. Usually,

traders apply a v of 20 and an h of 14 days.

Pr

14

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Appendix B.3. Support and resistance trading rules

we
Asset prices have certain levels, e.g. round numbers, around which orders tend to be clustered. As a
result, crossing these values gives a strong indication of price momentum. Usually traders determine these
levels by connecting past minima (support level) or past maxima (resistance level) graphically. As this is a
quantitative study we cannot easily do qualitative analysis with graphs in order to determine these support
and resistance levels. Consequently we use simple quantitative trading rule were we buy (sell) if the price
moves above (below) a previous maximum (minimum).
Mathematically, if we define maxt (j) and mint (j) as the maximum (resistance level) and minimum

ie
(support level) of the last j prices up until time t, then the trading signals given by the support and
resistance trading rule are:


ev
1,
 for Pt > maxt−1 (j)
DSR,t+1 = −1, for Pt < mint−1 (j) (B.5)

0, otherwise


Appendix B.4. Filter rules

r
Filter rules are another type of momentum indicator. A buy (sell) signal is given when the price moves
x% (y%) above its most recent low (high). Using maxt (j) and mint (j), which we defined in Appendix B.3,
we have:


er
for Pt > (1 + x)mint−1 (j)

 1,
and Pt−1 <= mint−1 (j)




DF,t+1 = for Pt < (1 − y)maxt−1 (j) (B.6)
pe
 -1,
and Pt−1 >= maxt−1 (j)





0, otherwise

We also use another type of filter rule where a buy (sell) signal is given if the price moves at least x% (y%)
above (below) its most recent low (high) and stays there for d(x) (d(y)) days. The position is neutralized
when the price moves y% (x%) below (above) its subsequent most recent high (low) and stays there for d(y)
ot

(d(x)) days. y% is less than x%, d(y) is less than or equal to d(x).

Appendix B.5. Channel Breakouts

We define a c%-trading channel as a period where the most recent maximum price is within c% of the most
tn

recent minimum price. In other words, a c%-trading channel exists at time t if (1 − c)maxt (j) <= mint (j).
When a trading channel exists, we apply support and resistance trading rules on the trading channel (see
Appendix B.3 for the specific trading signals).

Appendix B.6. On-Balance Volume Averages

rin

Research has shown that trading volume also contains valuable information. Technical traders pay close
attention to trading volume in order to make trading decisions. Therefore, we find it prudent to also include
a volume-based technical trading rule, namely the on-balance volume averages. We define the OBVt , the
on-balance volume at time t, as a running counter of the total volume traded up until time t. Namely:
(
ep

OBVt−1 + Vt , for rt >= 0

OBVt = (B.7)
OBVt−1 − Vt , for rt < 0
Where Vt is the total volume traded between t − 1 and t. In order to make trading decisions, we then
apply moving average strategies to our OBV time series instead of our price time series, as described in
Appendix B.1.
Pr

15

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Appendix B.7. Parameters

we
TTR Parameters Description Values
MA p Length of short MA 1,2,6,12,18,24,30,48,96,144,168
# 7,920 q Length of long MA 2,6,12,18,24,30,48,96,144,168,192
x Fixed percentage band filter 0,0.05,0.1,0.5,1,5
d Time delay filter 0,2,3,4,5
k Holding period 6,12,24,∞a

ie
RSI h Lookback period of the RSI 2,6,12, 14,18,24,30,48,96,144,168,192
# 720 v 50 ± v overbought/sold threshold 10,15,20,25
d Time delay filter 1,2,5
k Holding period 1,6,12,24,∞a

ev
SR j Lookback period 2,6,12,18,24,30,48,96,168
# 1,890 x Fixed percentage band filter 0.05,0.1,0.5,1,2.5,5,10
d Time delay filter 0,1,2,3,4,5
k Holding period 1,6,12,24,∞a
Filter j Lookback period 1,2,6,12,24
# 1,575 + 1,260 x Filter for buy signals 0.05,0.1,0.5,1,5,10,20

r
y Filter for sell signals 0.05,0.1,0.5,1,5,10,20
d(x) Time delay filter for buy signals 0,1,2,3,4,5
d(y) Time delay filter for sell signals
er 0,1,2,3,4
d Uniform time delay filter 0,1,2,3,4,5
k Holding period 6,12,18,20,24,∞a
CB j Lookback period 6,12,18,24,36,72,120,168
# 3,000 c c% trading channel 0.5,1,5,10, 15
x Fixed percentage band filter 0.05,0.1,0.5,1,5
pe
d Time delay filter 0,1,2
k Holding period 1,6,12,24,∞a
OBV p Length of short MA 2,6,12,18,24,30,48,96,144,168
# 2,475 q Length of long MA 2,6,12,18,24,30,48,96,144,168,192
x Fixed percentage band filter 0,0.01,0.05
d Time delay filter 0,2,3,4,5
k Holding period 6,12,∞a
ot

Fixed percentage band filter: filter around the signal threshold in order to reduce noisy trades.
Time delay filter: Number of periods that a signal must be valid before a transaction is executed.
Note: Holding period: Number of days a position is held, ignoring all other signals during that time.
a
∞ meaning that the position will be held until a different trading signal is given.
tn

Table B.6: Parameters used in the Technical Trading Rules.

The parameterization of our technical trading rules are based on earlier research by Sullivan et al. (1999)
and Hsu et al. (2016). We largely use the same parameters as in Hsu et al. (2016), however we adjust the
time lag parameters in order to make more sense at the 1 hour sample frequency (e.g. a MA length of
rin

25 is changed to a length of 24). In addition to the parameters described earlier, we also apply 3 common
parameters for every technical trading rule. In order to reduce the number of erroneous trades during volatile
periods, we apply a fixed percentage band filter, x, around the trade signal thresholds, and a time delay
filter, d, which represents the number of periods a trade signal should be valid before a trade is executed for
every technical trading rule. In addition we also define a holding period, k, which represents the number of
periods a trade position is held before returning to a neutral position. In total we compute 18,840 technical
ep

trading rules per sampling frequency, resulting in a total of 75,360 technical trading rules. Their individual
parametrizations can be seen in table B.6.

Appendix C. FDR+ estimators

Pr

In this section we shortly replicate the methodology explained in Appendix B, C, D and E of Bajgrowicz
and Scaillet (2012). In order to estimate the amount of positive false discoveries, FDR+ , in our set of l
16

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
technical trading rules, we assume that the trading rules with no abnormal performance have uniformly
distributed p-values over [0,1]. The FDR+ is then defined as:

we
+ 1
ˆ + (γ) = F̂ = 2 π̂0 lγ
F DR (C.1)
R̂+ #{pk ≤ γ, φk > 0; k = 1, . . . , l}
Where π̂0 is the estimated proportion of rules with no abnormal performance, γ (∈ [0, 1]) is a cut-off
point for the p-values with abnormal performance, pk and φk are the p-value and performance of trading

ie
rule k, respectively. π̂0 is estimated as follows:

#{pk > λ; k = 1, . . . , l}
π̂0 (λ) = (C.2)
l(1 − λ)

ev
Where λ ∈ [0, 1] is a tuning parameter which determines the minimum value after which all other are
supposed to belong to non-abnormal trading rules. The selection of γ can be automated using the method
of Storey (2002). In order to obtain a 10%-FDR+ portfolio, we sort our trading rules with a positive
performance from smallest to largest p-value. We add the rule with the smallest p-value and compute the
FDR+ if the 10% level is not obtained, we add the rule with the next highest p-value until we reach a FDR+

r
of 10%.

Acknowledgements
er
The computational resources (Stevin Supercomputer Infrastructure) and services used in this work were
provided by the VSC (Flemish Supercomputer Center), funded by Ghent University, FWO and the Flemish
Government – department EWI.
pe
References
Bajgrowicz, P. and Scaillet, O. (2012). Technical trading revisited: False discoveries, persistence tests, and transaction costs.
Journal of Financial Economics, 106(3):473–491.
Bariviera, A. F., Basgall, M. J., Hasperué, W., and Naiouf, M. (2017). Some stylized facts of the Bitcoin market. Physica A,
484:82–90.
ot

Barras, L., Scaillet, O., and Wermers, R. (2010). False discoveries in mutual fund performance: Measuring luck in estimated
alphas. Journal of Finance, 65(1):179–216.
Benjamini, Y. and Hochberg, Y. (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple
Testing. Journal of the Royal Statistical Society: Series B (Methodological), 57(1):289–300.
Bitstamp (2012). PRESS RELEASE: Bitstamp introduces volume discount program. https://blog.bitstamp.net/post/
tn

press-release-bitstamp-introduces-volume-discount (accessed 4 October 2021).

Bitstamp (2015a). Bitstamp is open for business - better than ever! https://blog.bitstamp.net/post/
bitstamp-is-open-for-business-better-than-ever (accessed 4 October 2021).
Bitstamp (2015b). Reduced fee schedule. https://blog.bitstamp.net/post/reducedfeeschedule (accessed 4 October
2021).
Bitstamp (2019). Upcoming changes to fee schedule. https://blog.bitstamp.net/post/
upcoming-changes-to-fee-schedule (accessed 4 October 2021).
rin

Brownlees, C. T. and Gallo, G. M. (2006). Financial econometric analysis at ultra-high frequency: Data handling concerns.
Computational Statistics and Data Analysis, 51(4):2232–2245.
Cheah, E.-T. and Fry, J. (2015). Speculative bubbles in Bitcoin markets? An empirical investigation into the fundamental
value of Bitcoin. Economics Letters, 130:32–36.
Corbet, S., Eraslan, V., Lucey, B., and Sensoy, A. (2019). The effectiveness of technical trading rules in cryptocurrency markets.
Finance Research Letters, 31:32–37.
ep

[dataset] Bitcoincharts.com (2020). Complete historical trade data. http://api.bitcoincharts.com/v1/csv/ (accessed

26 November 2020).
[dataset] Bitcoinity.org (2020). Bid-ask spread. https://data.bitcoinity.org/ (accessed 26 November 2020).
[dataset] OECD (2020). ”Short-term interest rates” (indicator). (accessed 26 November 2020).
Detzel, A., Liu, H., Strauss, J., Zhou, G., and Zhu, Y. (2021). Learning and predictability via technical analysis: Evidence
from bitcoin and stocks with hard-to-value fundamentals. Financial Management, 50(1):107–137.
Elliott, G. and Timmermann, A. (2008). Economic forecasting. Journal of Economic Literature, 46(1):3–56.
Pr

Gerritsen, D. F., Bouri, E., Ramezanifar, E., and Roubaud, D. (2019). The profitability of technical trading rules in the Bitcoin
market. Finance Research Letters.

17

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552
 d
Grobys, K., Ahmed, S., and Sapkota, N. (2020). Technical trading rules in the cryptocurrency market. Finance Research
Letters, 32:101396.

we
Hsu, P. H., Taylor, M. P., and Wang, Z. (2016). Technical trading: Is it still beating the foreign exchange market? Journal of
International Economics, 102:188–208.
Hudson, R. and Urquhart, A. (2021). Technical trading and cryptocurrencies. Annals of Operations Research, 297:191–220.
Kristoufek, L. (2018). On Bitcoin markets (in)efficiency and its evolution. Physica A: Statistical Mechanics and its Applications,
503:257–262.
Nadarajah, S. and Chu, J. (2017). On the inefficiency of Bitcoin. Economics Letters, 150:6–9.
Nakamoto, S. (2008). Bitcoin: A Peer-to-Peer Electronic cash system.
Politis, D. N. and Romano, J. P. (1994). The Stationary Bootstrap. Journal of the American Statistical Association,

ie
89(428):1303–1313.
Romano, J. P., Shaikh, A. M., and Wolf, M. (2008). Formalized Data Snooping Based on Generalized Error Rates. Econometric
Theory, 24(2):404–447.
Sensoy, A. (2019). The inefficiency of Bitcoin revisited: A high-frequency analysis with alternative currencies. Finance Research

ev
Letters, 28:68–73.
Storey, J. D. (2002). A direct approach to false discovery rates. Journal of the Royal Statistical Society. Series B: Statistical
Methodology, 64(3):479–498.
Sullivan, R., Timmermann, A., and White, H. (1999). Data-Snooping, Technical Trading Rule Performance, and the Bootstrap.
The Journal of Finance, 54(5):1647–1691.
Urquhart, A. (2016). The inefficiency of Bitcoin. Economics Letters, 148:80–82.
Wilder, J. W. J. (1978). New concepts in technical trading systems. NC: Hunter Publishing Company, Greensboro.

r
er
pe
ot
tn
rin
ep
Pr

18

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4401552