Comparative Analysis of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Transformer Models in Predicting Stock Prices

Abstract

This research explores the application of LSTM, GRU and Transformer models for predicting stock prices, aiming to enhance accuracy in financial forecasting. Stock price prediction is crucial for investment decision-making, yet challenging due to market volatility and complex patterns. The objectives are to evaluate the performance of LSTM, GRU and Transformer models using key metrics such as test loss, MAE, and MSE, and to compare their predictive capabilities. The LSTM model demonstrates robust performance with low test loss and MAE, indicating precise predictions and effective pattern recognition in financial data. In contrast, the Transformer model also shows promising results with relatively low test loss and MAE, albeit with larger errors in MSE and MAE metrics. Both models highlight the potential for accurate stock price prediction, suggesting avenues for future research to optimize model performance and reliability in financial forecasting applications. The experimental results show that GRU Outperformed LSTM and Transformer with an MSE of 0.0008, MAE of 0.0023, and high test accuracy of 0.9833.