Forecasting Bitcoin Prices with Deep Learning Models

Abstract

This study compares the forecasting performance of four deep learning architectures—GRU, LSTM, RNN, and CNN—for one-step-ahead Bitcoin price prediction. A grid search determined the optimal configuration, which was applied uniformly across models to ensure fair evaluation. Using daily BTC closing prices from January 2018 to July 2025, it is found that the GRU model achieved the lowest forecasting errors (MSE, RMSE, MAE, MAPE) and the highest R², with LSTM performing closely behind. Visual analyses confirmed that GRU and LSTM maintained stronger alignment with actual prices during volatile periods. To assess economic value, model forecasts were integrated into a rule-based trading strategy under realistic market frictions, including a 0.10% transaction cost and a 0.10% trading threshold, with both short-selling-enabled and long-only variants tested. The GRU strategy with short-selling generated the highest terminal wealth (approximately 24% higher than the Buy-and-Hold benchmark) and superior risk-adjusted returns, measured by CAGR, Maximum Drawdown, and Sharpe Ratio. The findings demonstrate that careful hyperparameter optimization, coupled with an architecture capable of capturing complex temporal dependencies, can significantly improve both predictive accuracy and trading profitability in cryptocurrency markets. These results provide practical implications for designing AI-driven trading systems.