Have a personal or library account? Click to login
Robustness of Support Vector Machines in Algorithmic Trading on Cryptocurrency Market Cover

Robustness of Support Vector Machines in Algorithmic Trading on Cryptocurrency Market

Central European Economic Journal
Volume 5 (2018): Issue 52 (January 2018)
By: Robert Ślepaczuk and  Maryna Zenkova  
Open Access
|Aug 2019

Figures & Tables

Fig. 2.1

Left panel: non-linear data divided by an SVM with a polynomial kernel of degree 3. Right panel: the same non-linear data divided by an SVM with a radial kernelSource: James, G., Witten, D., Hastie, T., Tibshirani, R. An Introduction to Statistical Learning with Applications in R.
Left panel: non-linear data divided by an SVM with a polynomial kernel of degree 3. Right panel: the same non-linear data divided by an SVM with a radial kernelSource: James, G., Witten, D., Hastie, T., Tibshirani, R. An Introduction to Statistical Learning with Applications in R.

Fig. 3.1

Percentage of total market capitalization by dominanceSource: https://coinmarketcap.com/charts/Note: the chart presents the dominance of the cryptocurrencies with the largest market capitalization over the period from 28/04/2013 to 05/08/2018.
Percentage of total market capitalization by dominanceSource: https://coinmarketcap.com/charts/Note: the chart presents the dominance of the cryptocurrencies with the largest market capitalization over the period from 28/04/2013 to 05/08/2018.

Fig. 3.2

Distribution of volatility-adjusted returns with 3-month history from 01/10/2014 to 12/31/2014 with cut-off lines to identify the number (training size) of the assets with the highest (green) and the lowest (red) volatility-adjusted returns to use for SVM trainingSource: own work.Note: the distribution of 91-days volatility-adjusted returns where the positive tail sets are the most positive volatility-adjusted returns, and the negative tail sets are the most negative. The vertical dotted lines represent the cut-off. The + and – regions are the ones used for SVM training.
Distribution of volatility-adjusted returns with 3-month history from 01/10/2014 to 12/31/2014 with cut-off lines to identify the number (training size) of the assets with the highest (green) and the lowest (red) volatility-adjusted returns to use for SVM trainingSource: own work.Note: the distribution of 91-days volatility-adjusted returns where the positive tail sets are the most positive volatility-adjusted returns, and the negative tail sets are the most negative. The vertical dotted lines represent the cut-off. The + and – regions are the ones used for SVM training.

Fig. 3.3

Visualization of the data set splits and their proportionsNote: the figure represents in what proportion data is split to run the tuning of the meta parameters.
Visualization of the data set splits and their proportionsNote: the figure represents in what proportion data is split to run the tuning of the meta parameters.

Fig. 4.1

The equity line of the SVM strategy in comparison with the benchmark strategiesNote: the graph shows equity lines of the SVM strategy and four benchmark strategies over the period from 01/01/2015 to 01/08/2018. EqW with equity line drawn in black outperforms all the benchmark strategies and also SVM strategy.
The equity line of the SVM strategy in comparison with the benchmark strategiesNote: the graph shows equity lines of the SVM strategy and four benchmark strategies over the period from 01/01/2015 to 01/08/2018. EqW with equity line drawn in black outperforms all the benchmark strategies and also SVM strategy.

Fig 4.2

Drawdowns of the SVM strategy in comparison with the benchmark strategiesNote: the graph shows the drawdown lines for SVM strategy and four benchmark strategies over the period from 01/01/2015 to 01/08/2018. SVM strategy drawn in green reaches the ‘deepest’ drawdown line if compared to the other benchmark strategies.
Drawdowns of the SVM strategy in comparison with the benchmark strategiesNote: the graph shows the drawdown lines for SVM strategy and four benchmark strategies over the period from 01/01/2015 to 01/08/2018. SVM strategy drawn in green reaches the ‘deepest’ drawdown line if compared to the other benchmark strategies.

Fig. 4.3

Equity lines of the SVM strategy with changing reallocation period RE: 1 week (base case), 1 month and 3 daysNote: the graph shows the equity lines of the SVM strategy with changing reallocation period RE over the period from 01/01/2015 to 01/08/2018. The length of the reallocation period significantly impacts the portfolio performance.
Equity lines of the SVM strategy with changing reallocation period RE: 1 week (base case), 1 month and 3 daysNote: the graph shows the equity lines of the SVM strategy with changing reallocation period RE over the period from 01/01/2015 to 01/08/2018. The length of the reallocation period significantly impacts the portfolio performance.

Fig. 4.4

Equity lines of the SVM strategy with changing number of assets N in the portfolio: 25, 20, 15, 10, 5 and VARNote: the graph shows the equity lines of the SVM strategy with changing number of assets N in the portfolio over the period from 01/01/2015 to 01/08/2018. The worst performance is noticed when only 5 coins are kept in the portfolio during a reallocation period. The lower the number of coins in the portfolio, the higher is the portfolio turnover.
Equity lines of the SVM strategy with changing number of assets N in the portfolio: 25, 20, 15, 10, 5 and VARNote: the graph shows the equity lines of the SVM strategy with changing number of assets N in the portfolio over the period from 01/01/2015 to 01/08/2018. The worst performance is noticed when only 5 coins are kept in the portfolio during a reallocation period. The lower the number of coins in the portfolio, the higher is the portfolio turnover.

Fig. 4.5

Equity lines of the SVM strategy with various transaction costs (%TC equalled 2%, 1% and 0.5%)Note: the graph shows the equity lines of the SVM strategy with changing transaction costs %TC in the portfolio over the period from 01/01/2015 to 01/08/2018. Performance of the portfolios heavily depends on the magnitude of transaction costs, which can be obviously seen from the behaviour of the equity lines.
Equity lines of the SVM strategy with various transaction costs (%TC equalled 2%, 1% and 0.5%)Note: the graph shows the equity lines of the SVM strategy with changing transaction costs %TC in the portfolio over the period from 01/01/2015 to 01/08/2018. Performance of the portfolios heavily depends on the magnitude of transaction costs, which can be obviously seen from the behaviour of the equity lines.

Fig. 4.6

Equity lines of the SVM strategy with changing length of the training set TS: 25%, 50%, and 100%Note: the graph shows the equity lines of the SVM strategy with changing length of training set %TS in the portfolio over the period from 01/01/2015 to 01/08/2018. As lines are evolving very close to each other, one may conclude that the change of the parameter %TS does not exercise a significant impact on the portfolio statistics.
Equity lines of the SVM strategy with changing length of the training set TS: 25%, 50%, and 100%Note: the graph shows the equity lines of the SVM strategy with changing length of training set %TS in the portfolio over the period from 01/01/2015 to 01/08/2018. As lines are evolving very close to each other, one may conclude that the change of the parameter %TS does not exercise a significant impact on the portfolio statistics.

Technical indicators used for the creation of feature set to train SVMs

FeatureFull nameParameters
MOM, n daysMomentum for close prices, n daysn = 10 days
ΔV, n daysVolume change n daysn = 10 days
RSIRelative Strength Indexn = 10 days
FIForce IndexN/A
Williams %RWilliams Percent Rangen = 10 days
PSARParabolic stop and reversal systemAcceleration factor by default set to 2% increasing by 2% with a maximum of 20%

Descriptive statistics for 10 largest and 10 smallest cryptocurrencies by MarketCap in TOP100 as of date 01-08-2018

The largest 10 cryptocurrencies in TOP100 as of 01-08-2018
Name%ARC%ASD%MDDIR1IR2Date of startVolume, mUSDMarketCap, USD
bitcoin11875.869.71.62.601-10-2014188843839225862
ethereum437.7145.284.3315.720-08-201532317124399552
ripple226.7164.687.11.43.601-10-201449913468236361
bitcoin-cash83.5198.584.40.40.405-08-20176996672807179
eos51625387.921214-07-2017785220519698
stellar228.6178.882.61.33.501-10-20143014634665748
litecoin111.1119.279.10.91.301-10-2014803709789644
cardano698.6263.889.32.620.714-10-2017322624893338
iota48.4188.482.90.30.126-06-201730592460207729
tether-5.445.749.9-0.1015-03-20151402233258238
The smallest 10 cryptocurrencies in TOP100 as of 01-08-2018
Name%ARC%ASD%MDDIR1IR2Date of startVolume, mUSDMarketCap, USD
loom-network649.7217.480324.321-04-20182.898040413
gas365.8265.489.61.45.609-08-20172.691875052
tenx-97.624799.2-0.4-0.410-07-20178.191154578
nxt34.2158.195.60.20.101-10-20142.890165499
cybermiles-44.8202.287.4-0.2-0.104-05-20187.188375828
nuls5083.5382.279.113.3854.822-03-20184.188067102
byteball160236.891.20.71.209-01-20170.5686950232
bibox-token53.2265.787.90.20.108-06-201867.583456610
odem89471229.640.7389.785663801-08-20180.13582906522
electroneum-92.524395.3-0.4-0.415-11-20170.55281946739

Descriptive statistics for SVM strategy (sensitivity analysis)_ Descriptive statistics for the benchmark strategies have been placed above for convenient comparison_

Benchmark Strategies
Name%ARC%ASD%MDDIR1IR2%MT
S&P B&H13.615.514.20.90.8
BTC B&H147.476.869.71.94.16.3
EqW425.896.281.74.423.110.8
McW141.974.973.11.93.76.3
SVM173.6103.183.11.73.5143.7
ParametersSVM Strategy
NPosition%TSRE%TC%ARC%ASD%MDDIR1IR2%MT
25long only503d119.4108.790.60.20.0115.3
25long only501w1173.6103.183.11.73.5143.7
25long only501m1224.2101.586.02.25.8148.8
5long only501w1-21.8142.295.1-0.20.0189.3
10long only501w189.3131.785.00.70.7176.8
15long only501w1207.2115.782.01.84.5166.2
20long only501w1215.9110.082.32.05.1154.3
25long only501w1173.6103.183.11.73.5143.7
VARlong only501w1326.492.657.63.520.0105.6
25long only1001w1177.9103.385.11.73.6144.3
25long only501w1173.6103.183.11.73.5143.7
25long only251w1210.6103.685.52.05.0160.5
25long only501w0,5368.8110.276.53.316.1155.4
25long only501w1173.6103.183.11.73,5143.7
25long only501w229.6110.988,10.30,1154.9
Best performance of SVM strategy with a selected set of parameters
NPosition%TSRE%TC%ARC%ASD%MDDIR1IR2%MT
VARlong only501m1392.4388.9753.454.4132.38105.9

Descriptive statistics of the SVM strategy compared with the benchmark strategies

NRE%TCV%ARC%ASD%MDDIR1IR2%MT
S&P B&H----13.615.514.20.90.8-
BTC B&H----147.476.869.71.94.1-
EqW1001 w1100425.896.281.74.423.110.8
McW1001 w1100141.974.973.11.93.76.3
SVM251 w1100173.6103.183.11.73.5143.7
DOI: https://doi.org/10.1515/ceej-2018-0022 | Journal eISSN: 2543-6821 | Journal ISSN: 2544-9001
Journal RSS Feed
Language: English
Page range: 186 - 205
Published on: Aug 7, 2019
Published by: Faculty of Economic Sciences, University of Warsaw
In partnership with: Paradigm Publishing Services
Publication frequency: 1 issue per year
Keywords:
Machine learning,
support vector machines,
investment algorithm,
algorithmic trading,
strategy,
optimization,
cross-validation,
overfitting,
cryptocurrency market,
technical analysis,
meta parameters
Related subjects:
Business and economics,
Political economics,
Economic theory, systems and structures,
Microeconomics,
Macroeconomics,
Economic policy

© 2019 Robert Ślepaczuk, Maryna Zenkova, published by Faculty of Economic Sciences, University of Warsaw
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Previous articleVolume 5 (2018): Issue 52 (January 2018)Next article