From: Forecasting and trading cryptocurrencies with machine learning under changing market conditions
Article | Dependent variable | Frequency | Sample period | Models | Type (classification/regression) | Trading strategies (positions/trading costs) | Input set | Main findings |
---|---|---|---|---|---|---|---|---|
Madan et al. (2015) | Bitcoin prices in USD from Coinbase | 10-s, 10-min | 5Ā years since the inception of Bitcoin | Binomial logistic regressions (BLR) and random forest (RF) | Classification | ā | Prices and 16 blockchain features | 10-min data give a better sensitivity and specificity ratio than the 10-s data |
Kim et al. (2016) | Bitcoin, ethereum and ripple prices | Daily | Bitcoin: Dec-2013 to Feb-2016 Ethereum: Aug-2015 to Feb-2016 Ripple: Sept-2015 to Jan-2016 | Averaged one-dependence estimators (AODE) | Classification | Long/no trading costs | Trading information, and comments and replies posted in online communities | Comments and replies are good predictors of Bitcoin prices |
Żbikowski (2016) | Bitcoin prices in USD from Bitstamp | 15-min | Jan-2015 to Feb-2015 | Exponential moving average (EMA), box support vector machine (SVM) and volume weighted SVM (VW-SVM) | Classification | Long and short/trading costs of 0.2% | 10 technical analysis indicators | VW-SVM is the best model in terms of average return and maximum drawdown |
Jiang and Liang (2017) | Prices in USD of the 12 most traded cryptocurrencies at Poloniex | 30-min | Jun-2015 to Aug-2016 | Convolutional neural networks (CNN) with deep reinforcement learning | Regression | Long and short/trading costs of 0.25% | Returns | Mixed results between CNN portfolio and Online Newton Step and Passive Aggressive Mean Reversion portfolios |
Jang and Lee (2018) | Bitcoin price index in USD | Daily | Sep-2011 to Aug-2017 | Bayesian neural networks (BNN), linear regression and support vector regressions (SVM) | Regression | ā | 26 blockchain features, trading information, exchange rates and macroeconomic variables | The BNN is the best prediction model |
McNally et al. (2018) | Bitcoin prices in USD from CoinDesk | Daily | Aug-2013 to July-2016 | Bayesian recurrent neural (RNN) and long short term memory (LSTM) | Classification and Regression | ā | OHLC prices, difficulty, and hash rate of blockchain | The best time lengths are 100Ā days for the LSTM and 20Ā days for the RNN |
Nakano et al. (2018) | Bitcoin returns in USD from Poloniex | 15-min | July- 2016 to Jan-2018 | Artificial neural networks (ANN) | Classification | Long, and long and short/transaction costs of 0.025%,0.05% and 0.1% | Returns and 4 technical analysis indicators | Higher performance of the ANN strategy, except in the last month of data. Results are highly sensitive to the model specification and input data |
Vo and Yost-Bremm (2018) | Bitcoin prices in USD, CNY, JPY, EUR from 6 online exchanges | 1-min | Jan-2012 to Oct-2017 | Random forests (RF) and a deep learning model | Classification | Long and short/no trading costs | 5 technical analysis indicators | RF is the best model for a frequency of 15-min |
Alessandretti et al. (2019) | Price indexes of 1681 cryptocurrencies in USD | Daily | Nov-2015 to Apr-2018 | Ensemble of regression trees built by XGboost and long short term memory network | Regression | Long/transaction costs of 0,1%, 0,2%, 0,5% and 1% | Price, market capitalization, market share, rank, volume, and age | All strategies, produce a significant profit (expressed in bitcoin) even with transaction fees up to 0.2% |
Atsalakis et al. (2019) | Bitcoin ethereum, litecoin and ripple returns | Daily | Sep-2011 to Oct-2017 | PATSOSāa hybrid neuro-fuzzy model | Classification and regression | Long and short/no transaction costs | Returns and prices | PATSOS outperforms other competing methods and produces a return significantly higher than the Buy-and-Hold (B&H) strategy |
Catania et al. (2019) | Bitcoin, ethereum, litecoin and ripple returns in USD | Daily | Aug-2015 to Dec-2017 | Linear univariate and multivariate regression models, and selections and combinations of those models | Regression | ā | Returns and several exogenous financial variables | Statistically significant improvements in forecasting returns when using combinations of univariate models |
de Souza et al. (2019) | Bitcoin prices in USD | Daily | May-2012 to May-2017 | Artificial neural network (ANN) and support vector machine (SVM) | Classification | Long and short/5 USD | OHLC prices | SVM provides conservative returns on the risk adjusted basis, and ANN generates abnormal profits during short run bull trends |
Han et al. (2019) | Bitcoin returns in USD | Daily | April-2013 to Mar-2018 | NARX Neural Network | Regression | ā | Returns | NARX is effective in predicting the tendency but not the jumps |
Huang et al. (2019) | Bitcoin returns in USD | Daily | Jan-2012 to Dec-2017 | Trees | Classification | Long and short/no trading costs | 124 technical indicators computed from the OHLC prices | Lower volatility, higher win-to-loss ratio and information ratio than those of every simple cut-off strategy or the B&H strategy |
Ji et al. (2019b) | Bitcoin returns in USD from Bitstamp | Daily | Nov.-2011 to Dec.-2018 | Deep Neural Network (DNN), Long Short Term Memory (LSTM), Convolutional Neural Network (CNN), Deep Residual Network (ResNet), combination of CNNs and RNNs (CRNN) and their combinations | Classification and regression | Long/no transaction costs | Prices and 17 blockchain features | Performances of the prediction models were comparable, LSTM is the best prediction model, DNN models are the best classification models, classification models were more effective for trading |
Lahmiri and Bekiros (2019) | Bitcoin, digital cash and ripple prices in USD | Daily | Bitcoin: July-2010 to Oct-2018 Digital Cash: Feb-2010 to Oct-2018 Ripple: Jan-2015 to Oct-2018 | Long Short Term Memory (LSTM) and Generalized Regression Neural Networks (GRNN) | Regression | ā | Prices | Predictability of LSTM is significantly higher than of GRNN |
Mallqui and Fernandes (2019) | Bitcoin prices in USD | Daily | Apr-2013 to Apr-2017 | Artificial neural networks (ANN), support vector machine (SVM) and ensembles | Classification and Regression | ā | OHLC prices, Blockchain information and several exogenous financial variables | Ensemble of recurrent neural networks and a Tree classifier is the best classification model, while SVM is the best regression model |
Shintate and Pichl (2019) | Bitcoin returns in CNY and USD from OkCoin | 1-min | Jun-2013 to Mar-2017 | Random sampling method (RSM) | Classification | Long and short/No transaction costs | OHLC prices | The proposed RSM outperforms several alternatives, but the profit rates do not exceed those of the B&H strategy |
Smuts (2019) | Bitcoin and ethereum prices in USD | 1-h | Dec-2017 to Jun-2018 | Long short term memory recurrent neural network (LSTM) | Classification | ā | Prices, volumes, Google trends, and Telegram chat groups dedicated to bitcoin and ethereum trading | Telegram data is a better predictor of bitcoin, while GoThe ensemble, by unweighted average of the four trading signals from the four models, after resampling the data, gives the best results.ogle Trends is a better predictor of ethereum, especially in one-week period |
Borges and Neves (2020) | Prices from Binance 100 cryptocurrencies pairs with the most traded volume in USD | 1-min | For each pair since beginning of trading at Binance until oct-2018 | Logistic regression, random forest, support vector machine, and gradient tree boosting and an ensemble of these models | Classification | Long/transaction costs of 0.1% | Returns, resampled returns, and 11 technical indicators | Ā |
Chen et al. (2020b) | Bitcoin price index and trading prices from Binance in USD | 5-min and daily | July-2017 to Jan-2018 for 5-min and Feb-2017, to Feb-2019 for daily | Logistic Regression (LR), Linear Discriminant Analysis (LDA), Random Forest (RF), XGBoost (XGB), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) | Classification | ā | 5-min: OHLC prices and trading volume. Daily: 4 Blockchain features, 8 marketing and trading variables, Google trend search volume index, Baidu media search volume, and gold spot price | For 5-min data machine learning models achieved better accuracy than LR and LDA, with LSTM achieving the best result (67% accuracy). For daily data, LR and LDA are better, with an average accuracy of 65% |
Chu et al. (2020) | Bitcoin, ethereum, dash, litecoin, MaidSafeCoin, monero and ripple from CryptoCompare in USD | Hourly | Feb-2017 to Aug-2017 | Exponential Moving Averages (EMA) for time series and cross-sectional portfolios | Classification and Regression | Long and short/No transaction costs | Trading prices | Momentum trading does not beat the passive trading strategies |
Sun et al. (2020) | 42 cryptocurrencies | Daily | Jan-2018 to Jun-2018 | LightGBM, SVM support vector machines (SVM) and Random Forests (RF) | Classification | ā | Trading data and macroeconomic variables | LightGBM outperforms SVM and RF, and the accuracy is higher for 2Ā weeks predictions |