Comparing Traditional Time Series Models and Deep Learning Approaches in Stock Price Forecasting

Dept of Mathematics, University College London, London, United Kindom

Abstract

This study conducts a comparative analysis between traditional statistical models and state-of-the-art deep learning approaches in the context of stock price forecasting. Using a decade of daily trading data from Apple Inc. (AAPL), spanning 2014 to 2024, this paper evaluates models on two forecasting tasks: predicting the closing price and predicting the next- day log-returns. The experiments demonstrate that the ARIMA model remains a strong baseline, particularly in controlling absolute prediction errors and maintaining stable performance across different market regimes. In contrast, recurrent neural networks such as GRU exhibit superior performance across almost all evaluation metrics, suggesting that they are better able to capture temporal dependencies in financial time series. However, deeper and more complex hybrid architectures, including stacked and convolutional variants, show a clear tendency to overfit the training data and fail to generalize to unseen periods, especially during volatile market phases. These findings underline the inherent difficulties of applying deep learning in financial forecasting, where challenges such as high noise levels, structural breaks, and low signal-to-noise ratios persist. Future work should therefore emphasize stronger regularization strategies, models explicitly designed to handle non-stationarity, and advanced methods of feature engineering and signal extraction to achieve more reliable predictive performance.