A Comparison of Machine Learning and Deep Learning Models for Daily SET Index Forecasting: An End-to-End Prediction Pipeline

Abstract

This study aims to compare the forecasting performance of traditional statistical models, machine learning models, and deep learning models for daily prediction of the Stock Exchange of Thailand (SET) Index. Eight models were evaluated: naive persistence, ARIMA, Random Forest, XGBoost, and both regression and classification variants of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. Daily closing price data spanning January 4, 2010 to July 17, 2026 were used, yielding 3,945 trading days after constructing 19 technical indicator features. Data were split chronologically into training (2010-2021), validation (2022-2023), and test (2024-2026) sets. Results show that Random Forest achieved the highest directional accuracy at 51.8%, followed by XGBoost at 50.8%, while all four LSTM/GRU variants achieved directional accuracy between 48.2% and 49.8%, at or below the naive persistence benchmark (49.7%) and coin-flip chance. Only Random Forest and XGBoost numerically exceeded the 50% baseline; a binomial test indicated that this advantage was not statistically significant at the 0.05 level. In terms of price-level error (RMSE, MAPE), most models performed comparably to naive persistence, except ARIMA, which showed markedly higher error. These findings are consistent with the Efficient Market Hypothesis and indicate that model complexity does not guarantee superior forecasting accuracy. In practice, these results suggest that simple tree-ensemble models should be the first choice for SET Index directional forecasting before investing computational resources in more complex deep learning architectures. This study contributes an end-to-end forecasting pipeline that benchmarks diverse model families under identical data and evaluation conditions, offering value to both academic researchers and market practitioners in Thailand.