Forecasting Hourly Cryptocurrency Returns Using Recurrent Neural Networks: A Comparative Study of GRU, LSTM and Classical Models

This study evaluates short-horizon forecasting of hourly cryptocurrency returns using two recurrent neural architectures—GRU and LSTM—estimated in more than 500 configurations across Bitcoin, Ethereum, Binance Coin and Litecoin. We adopt a unified protocol with intersection evaluation windows to ensure identical data coverage across models, and we compare magnitude-based errors (RMSE, MAE, MASE, sMAPE) with direction-based performance (Directional Accuracy, DA). Classical benchmarks (ARIMA/ETS, GARCH and a Random Walk random-walk) are estimated under the same one-step-ahead design. Empirically, GRU networks consistently achieve lower errors and higher DA than LSTM and traditional models. Best GRU configurations reach DA ≈ 0.65–0.72 depending on the asset, while requiring smaller amplitude recalibration. The results indicate that parsimonious recurrent gating is well-suited to the high-volatility, short-memory structure of cryptocurrency returns. Methodologically, the paper replicates and extends a previously published currency-market framework to a more turbulent domain, reinforcing the external validity of the findings.