Stock market trend prediction using LSTM neural networks and financial time series

Abstract

This Final Degree Project examines the application of Long Short-Term Memory (LSTM) neural networks to the forecasting of financial market volatility using historical time series data. Traditional econometric models, such as ARIMA, GARCH, EWMA, and HAR, have been widely employed in volatility forecasting but are often limited by linear assumptions and restricted capacity to capture nonlinear dynamics and long-term dependencies inherent in financial markets. In contrast, deep learning models offer a flexible, datadriven framework capable of learning complex temporal patterns. The empirical analysis focuses on forecasting the logarithm of five-day realized volatility of the SPDR S&P 500 ETF Trust (SPY) over the period from 2010 to 2025. The model incorporates multiple explanatory variables, including daily log-returns, trading volume, the Parkinson intraday volatility estimator, and the VIX index. A rigorous preprocessing procedure is implemented, involving normalization, feature engineering, and sequence construction using a 60-day rolling window. Model evaluation is conducted using a chronological train–validation–test split to prevent look-ahead bias. The forecasting performance of the LSTM model is compared against established benchmark models, including persistence, EWMA, and HAR specifications. Forecast accuracy is assessed using root mean squared error (RMSE) and Diebold–Mariano statistical tests. The results indicate that, while the LSTM model effectively captures volatility clustering and regime shifts, it does not achieve statistically significant improvements over the persistence benchmark. In contrast, the EWMA and HAR models demonstrate superior and statistically significant forecasting performance. The findings suggest that, despite their theoretical advantages, deep learning models do not necessarily outperform simpler econometric approaches in short-term volatility forecasting. This work highlights the importance of robust benchmarking and contributes to a critical understanding of the practical limitations and potential of machine learning techniques in financial time series analysis.