Financial time series forecasting with long short-term memory (LSTM) : a comparative experiment between deep learning and econometrics

Abstract

Financial market characterized by inherent stochasticity and volatility has posed unique challenges for predictive modelling where the returns on securities are considered unpredictable. This thesis contributes to the evolving domain of financial time series forecasting by exploring the predictive application of Long Short-Term Memory (LSTM) networks - a deep learning-based forecasting technique renowned for its capability to capture long-term dependencies within sequence data. Given the limitations of traditional econometrics models like Autoregressive Integrated Moving Average (ARIMA) in handling non-linearity and non-stationarity, the study seeks to validate whether LSTM can enhance the predictive accuracy through a comparative analysis. The study conducts an empirical experiment to compare the performance of LSTM and ARIMA models in predicting 20-day future prices of the S&P 500 index using historical daily prices from January 2000 to January 2024. Walk-forward validation methodology with 12 data folds for training, validating and testing is employed to ensure the fairness and reliability in constructing the predictive models. In each fold, four optimal models of each model type were tuned by adjusting their hyperparameters to identify the best performing model for the test set. The findings reveal that LSTM models outperform ARIMA models in three performance metrics using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Specifically, the average deduction in prediction errors obtained by LSTM is between 82 – 84 percent in terms of MAE and RMSE, and 3.96% lower in terms of MAPE. Furthermore, it is discovered that the predictions of LSTM model are consistent with minimal variations across the data folds compared to greater variation across prediction outcomes from ARIMA models. The findings of this study indicate a potential of deep learning techniques in enhancing the forecasting capabilities and setting new benchmarks in predictive accuracy.