Benchmarking Deep Learning Models for Bitcoin Price Forecasting

Abstract

The rapid appreciation and pronounced volatility of Bitcoin (BTC-USD) poses significant challenges for accurate price forecasting using financial time series models. This paper presents a systematic benchmarking study of six deep learning architectures, including Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Vanilla Recurrent Neural Network (RNN), one dimensional Convolutional Neural Network (CNN), Temporal Convolutional Network (TCN), and Transformer Encoder, for daily BTC USD price prediction. Experiments are conducted using 2,283 observations spanning January 2020 to March 2026 under a standardized training pipeline that includes a 60 day lookback window, Huber loss, Adam optimization, and a strict chronological split of 75% training, 10% validation, and 15% testing to mitigate data leakage. Empirical results indicate that the GRU consistently achieves superior out of sample performance, with a test coefficient of determination (R²) of 0.9138, root mean squared error (RMSE) of USD 4,911 and mean absolute percentage error (MAPE) of 3.84%. In contrast, Transformer, LSTM, TCN, CNN, and Vanilla RNN models exhibit substantially weaker generalization performance. The results suggest that the GRU’s gating mechanism, which enables selective retention and attenuation of historical information, is particularly effective in capturing regime dependent dynamics in highly volatile cryptocurrency markets. A 30 day out of sample forecast for April 2026 is further generated for all models, with the GRU projecting Bitcoin prices in the USD 66,000–67,000 range. The proposed benchmark provides reproducible experimental settings, quantitative performance comparisons, and architectural insights to support informed model selection for cryptocurrency price forecasting.