Application of Temporal Fusion Transformer for Multivariate Forecasting of Indonesian Composite Stock Price Index

Abstract

Accurate stock market index forecasting remains challenging because of complex temporal dependencies, high volatility, and uncertainty arising from dynamic financial conditions. This study investigates the application of the Temporal Fusion Transformer (TFT) to multivariate, multi-horizon forecasting of the Indonesian Composite Stock Price Index (IHSG). The proposed framework integrates historical IHSG movements, global financial indicators, technical indicators, and calendar features to generate point and probabilistic forecasts. Daily financial data from January 2000 to December 2025 were analyzed using a 252-trading-day historical input window to forecast the subsequent 30 trading days. Point-forecasting performance was evaluated using Mean Absolute Percentage Error (MAPE), whereas probabilistic performance was assessed using Quantile Risk (q-Risk). The TFT produced a MAPE of 13.723% over the 30-trading-day forecasting horizon. Its probabilistic forecasts maintained consistent quantile ordering, although further calibration is required to improve the reliability of uncertainty estimates. Variable selection analysis identified historical IHSG movements as the most influential predictive information, followed by the Relative Strength Index (RSI), lagged crude oil prices, volatility, and global market indicators, including movements in the S&P 500. The study contributes to financial forecasting research by demonstrating an interpretable deep learning framework that integrates point prediction, uncertainty representation, and variable importance analysis within a unified forecasting approach. The findings indicate that TFT can provide not only multi-horizon index forecasts but also transparent insights into the information structure underlying IHSG movements, thereby supporting more interpretable financial forecasting applications.