Temporal Fusion and Hybrid ARIMA-LSTM Models for Robust Time Series Forecasting: A Case Study on Ethereum Price Dynamics

Abstract

Ethereum price forecasting remains a challenging task due to the highly volatile and nonlinear nature of digital asset markets. This study proposes a hybrid time-series forecasting model that integrates an Autoregressive Integrated Moving Average (ARIMA) model with a Long Short-Term Memory (LSTM) network to improve Ethereum price prediction accuracy. The ARIMA model captures linear dependencies and extracts statistical residuals, which are then incorporated as an additional input feature for the LSTM network to enhance its learning of complex temporal patterns. The model is trained using a dataset containing historical Ethereum price data, with MinMax normalization applied to the closing prices for improved stability and provides 0.398821 MAE. A comprehensive ablation study evaluates different model configurations, demonstrating that the ARIMA residuals significantly enhance predictive performance. The hybrid ARIMA-LSTM model achieves a Mean Squared Error (MSE) of 0.1846, outperforming standalone LSTM and ARIMA models. Further residual and error analysis confirm that the model effectively mitigates autocorrelation in forecasting errors while maintaining stable predictive performance.