Transformer Model for Long-Term Stock Market Volatility Prediction

Abstract

Predicting long-term stock market volatility in emerging stock markets is a significant challenge due to the data’s inherent non-linear and complex nature. This study aims to design and evaluate a Transformer-based deep learning architecture to enhance the accuracy of long-term volatility prediction, with a case study on the Indonesia Stock Exchange Composite Index (IDX Composite). The proposed method integrates multi-modal time-series data, including historical market data (IDX Composite, blue-chip stocks), key macroeconomic indicators (interest rates, inflation), and cross-aset data (USD/IDR). The Transformer architecture, leveraging its self-attention mechanism, is employed to capture complex long-range temporal dependencies within the data. The model was trained to predict future realized volatility over a 21-day horizon. Its performance was rigorously evaluated against popular baseline architectures, namely Long Short-Term Memory (LSTM) and GARCH(1,1), using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) metrics. The evaluation results on the test set demonstrate that the proposed Transformer model consistently outperforms both LSTM and GARCH models, indicated by lower RMSE and MAE values. This finding highlights the superior capability of the Transformer architecture in modeling complex non-linear relationships and long-term dependencies from multi-modal financial data. In conclusion, the Transformer architecture proves to be a robust and effective framework for long-term volatility prediction, offering a promising solution for risk analysis and strategic investment decision-making in emerging markets.