Comparative Analysis of Econometric Models' and Machine Learning Algorithms' Effectiveness in the Cryptocurrency Market

This study is devoted to a comparative analysis of the effectiveness of econometric methods and machine learning algorithms in the cryptocurrency market. Such a study is important because this market is poorly studied, is very different from the stock or fiat markets and has unique features inherent only in it. The data used candles of digital currency Bitcoin against the US dollar. Data aggregation included technical analysis tools used as independent variables in models. These indicators were built with different versions of the step, and then the most important ones were selected based on the indicators of the importance features in the XGBoost and CatBoost algorithms. A meta-variable was also formed, representing the probability of price growth at time t, derived from the predictions of the MLP model. There Were written our own metrics for assessing the quality of regression models and custom loss functions that are sensitive to the wrong sign in predicting price changes. Logit, MLP, XGBoost, CatBoost, LSTM were used as classification models; as regression - XGBoost, CatBoost, ARMA, LSTM, hybrid ARMA-LSTM. The results indicate that the classification algorithms are much inferior to the regression ones. At the same time, the hybrid model ARMA-LSTM turned out to be the best of the regression models themselves, which managed to catch the price change on the test sample with an accuracy of 62%.