Application of machine learning to filter inefficient trading signals generated by mechanical approach indicators

DOI: 10.31673/2412-9070.2025.051067

Abstract

The subject of this study is the application of machine learning algorithms to filter out ineffective trading signals generated by indicators of the mechanistic approach to the analysis of cryptocurrency markets. The purpose of the work is to develop and test ML models capable of increasing the profitability of trading strategies by filtering false signals that arise when using mechanistic indicators. The research tasks include: 1) formalization of mechanistic indicators (mechanistic moving average, MAS Buy, MAS Sell); 2) use of the triple barriers method to classify signals into effective and ineffective; 3) application and comparison of 5 machine learning algorithms (Summary Classifier, Catch22, Rocket, TimeCNN, Stacking); 4) evaluation of the effectiveness of models using the ROC metrics AUC, Precision, Recall, Average Precision and Sharpe Ratio; 5) ranking models using a multi-criteria approach to decision-making. The results obtained showed that machine learning models outperform the basic Dummy model in most cases, especially for long positions, where higher Sharpe Ratio and total return values were recorded. The best results were demonstrated by the Catch22, Rocket, and Stacking classifiers. On the other hand, short positions turned out to be less effective, which is associated with the growing trend of cryptocurrencies in 2019–2025. It was also found that variations in windows (16, 32, 64, 128) significantly affect the results, confirming the importance of parameter optimization. Thus, the work demonstrates the feasibility of integrating machine learning algorithms into a mechanistic approach to improve the quality of trading signals and the profitability of strategies. The results can be used to improve algorithmic trading systems and develop new approaches to the application of ML in the direction of quantitative finances.

Keywords: mechanistic approach; cryptocurrency; retrospective testing; machine learning; algorithm; filtering; classifier; adaptive learning; metric; target variable.