Bitcoin price prediction using Autoencoder-based GRU and LSTM models

Abstract

The high volatility of Bitcoin prices presents a major challenge in forecasting, as predictive models must capture both short-term fluctuations and long-term trends. Despite extensive studies on deep learning for cryptocurrency prediction, there remains a lack of systematic comparative analysis of autoencoder-based recurrent architectures, particularly AE-LSTM and AE-GRU, across both univariate and multivariate input settings, with statistical validation. This study analyzes Bitcoin price data from 2018 to 2025 to evaluate and compare the performance of two hybrid deep learning models, Autoencoder-based Gated Recurrent Unit (AE-GRU) and Autoencoder-based Long Short-Term Memory (AE-LSTM), in Bitcoin price prediction. The experiments explore the effects of dropout, learning rate, and epochs, using both univariate and multivariate inputs (Open, High, Low, Close). Results show that AE-GRU consistently outperforms AE-LSTM across all configurations, achieving up to 16.5% higher MAPE-based accuracy. The best performance was achieved by the multivariate AE-GRU, dropout = 0.1, learning rate = 0.001, epoch = 100, with RMSE 1667.125, MAE 1145.718, MAPE 2.33%, and R² 0.995017. Moreover, AE-GRU demonstrates faster training efficiency, requiring 185–195 ms/step, while AE-LSTM takes 208–215 ms/step under the same conditions. AE-GRU's superior accuracy and efficiency are attributed to its simplified gating structure and the feature compression capability of the autoencoder, which enhances learning stability and generalization. Overall, the AE-GRU model offers robust predictive performance and computational efficiency. It is a reliable framework for real-time cryptocurrency forecasting and a promising foundation for advanced deep learning architectures in financial time-series analysis.