Comparative Analysis of Statistical and Machine Learning Models for Forecasting Cryptocurrency Prices and Directional Movements

This work compares statistical and deep learning methods for forecasting cryptocurrency prices and directional movements. Daily price series were analyzed from 11/09/2017 to 12/31/2024 for four crypto assets (Bitcoin, Ethereum, Litecoin, and Ripple). The statistical methods used were ARIMA and ETS, and the deep learning methods were LSTM, RNN, and GRU. The results indicate that ARIMA and ETS achieved the lowest MAPE values in most asset-horizon combinations, while recurrent models showed stronger directional accuracy, especially at the 7- and 14-day horizons. In the Diebold-Mariano tests applicable to ARIMA, the null hypothesis of equal predictive accuracy relative to the random walk was not rejected in any of the evaluable comparisons. The results therefore reinforce the importance of jointly considering error magnitude, movement direction, and statistical significance when comparing forecasting models.