Comparative Study of Statistical and Machine Learning Models for Cryptocurrency Return Prediction

Abstract

This study compares the performance of traditional statistical models and modern machine learning methods in predicting Bitcoin (BTC) returns. Using data from January 2022 to December 2024, both daily and weekly datasets were constructed with a comprehensive set of features, including technical indicators, on-chain metrics, macroeconomic variables, and behavioral signals. Linear regression models, ensemble learning methods (Random Forest, XGBoost), deep learning (LSTM), and GARCH-based time-series models were evaluated for both numeric return prediction and directional classification. Results show that while all models struggle to forecast exact return values, classification models achieve better performance, particularly when using weekly data. Among all tested models, the XGBoost classifier achieved the highest directional accuracy of 67.8% on weekly data, outperforming both statistical and deep learning baselines. These findings highlight the value of frequency selection and nonlinear modeling in improving cryptocurrency return predictability.