Ethereum Price Forecasting Using Bidirectional GRU Neural Networks Optimized with SGD

Abstract

Accurate forecasting of cryptocurrency prices is a challenging task due to their high volatility, nonlinear dynamics, and strong temporal dependencies. Ethereum, as one of the most prominent cryptocurrencies, exhibits complex price movements influenced by market speculation, external events, and rapidly changing investor behavior. Traditional statistical and shallow machine learning approaches often fail to capture these characteristics effectively, leading to suboptimal prediction performance. Motivated by these limitations, this study proposes an Ethereum price forecasting model based on a Bidirectional Gated Recurrent Unit (BiGRU) neural network optimized using Stochastic Gradient Descent (SGD). The proposed approach leverages bidirectional recurrent learning to capture both past and future contextual information during training, while the use of SGD aims to improve generalization performance and training stability. A structured methodology is applied, including data preprocessing, sliding-window sequence construction, model training, and systematic evaluation. The model is evaluated using standard regression metrics, including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and visual comparison between predicted and actual prices. Experimental results demonstrate that the proposed BiGRU–SGD model is able to closely track Ethereum price trends and capture temporal dependencies with satisfactory accuracy on unseen testing data. The findings indicate that combining bidirectional recurrent architectures with carefully configured optimization strategies provides a robust solution for cryptocurrency price forecasting. Future work may extend this framework by incorporating additional market indicators, sentiment data, or advanced architectures such as attention mechanisms and transformer-based models to further enhance predictive performance.