Machine learning, memory and efficiency in cryptocurrency markets

This paper empirically examines whether machine learning (ML) methods can capture long memory in the cryptocurrency markets. We design two tests to evaluate seven widely used ML regression algorithms and sequence-to-sequence (Seq2Seq) models to determine their ability to capture long-memory characteristics of financial data. Specifically, we assess their accuracy in estimating the fractional integration parameter d for both univariate and systemic memory. Additionally, we examine whether the predicted time series preserve the long-memory properties of the original cryptocurrency market data. Our findings reveal that most ML algorithms fail to handle long-memory series effectively, while models incorporating Long Short-Term Memory (LSTM) and Attention-LSTM components exhibit superior performance. Whilst comparing models using Mean Squared Error (MSE), we find that our tests identify models better for directional predictions. These results highlight the limitations of conventional ML mechanisms for long-range dependence and position Seq2Seq models as a promising alternative for addressing the complex movements of cryptocurrency time series. Our approach can be readily extended, offering both academics and practitioners a systematic procedure for evaluating arbitrary ML models, thereby yielding insights not only into their generalization performance but also into the interpretability of their capacity to model long-term dependence.