Application, comparison and performance optimization of lstm, RNN, and gru in stock price prediction

Abstract

Stock price prediction is one of the most important tasks in financial analysis; it plays a vital role in investor decision-making, portfolio optimization, and risk management. Stock markets are volatile and non-linear by design making them hard to be accurately estimated using traditional statistical techniques. To address these shortcomings, deep learning methods have emerged as powerful alternatives, providing advanced approaches that can capture complex temporal and sequential dynamics within time series information. The choice of the framework is critical, with many models including Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Recurrent Neural Networks (RNN). The goal of this research is to study the prediction capabilities of such architectures by assessing their performance in forecasting eBay Inc. stock price. In addition, this study investigates the influence of the optimizer employed on the overall performance of the model, emphasizing recent optimization algorithms such as NAdam and AdamW that aim to speed up convergence and facilitate better generalization. It also compares these optimizers with the more commonly adopted Adam algorithm to evaluate their efficacy in improving the performance of your model. At this point, the empirical findings of this paper offer meaningful perspectives. Training on it and taking the R² value of the GRU model with Adam optimizer 0.9677 illustrates that this method best fits the capture of the stock price complex dynamics. In contrast, the RNN model had relatively less predictive performance than GRU overall in this domain but demonstrated a significant improvement with implementing the NAdam optimizer. In particular, the RNN’s R² score improved from 0.9146 (Adam) to 0.9224 (NAdam), illustrating how important it is to pay attention to the optimizer when working to make that final step in improving model performance. Through this work, we present a better understanding of the unexpected performance range offered by GRUs for analyzing financial time series data, and we also validate the potential of custom optimization strategies towards boosting modeling results, steering us toward more accurate and reliable out-of-sample stock price prediction solutions