Komparasi XGBoost dan GRU+Attention untuk Prediksi IHSG Jangka Pendek Menggunakan Walk-Forward Analysis dan Metrik Finansial

Abstract

Predicting the Indonesia Composite Stock Price Index (IHSG) is challenging due to high market volatility. This study aims to compare XGBoost and GRU+Attention for short-term IHSG prediction, emphasizing realistic validation and practical profitability. Utilizing historical data from 2018 to 2026, the research employs the SEMMA framework and Walk Forward Analysis (70% training, 15% validation, 15% testing) to prevent data leakage. Unlike conventional studies relying solely on numerical errors, this research evaluates models using multidimensional metrics, including Directional Accuracy (DA), asymmetric financial metrics, and Profit and Loss (PnL) simulation. Results reveal a significant trade-off: while GRU+Attention achieved lower numerical errors (RMSE: 677.6), XGBoost demonstrated superior directional prediction with a DA of 55.33% and generated a cumulative profit of 1627.98 points. These findings conclude that directional accuracy and financial evaluation are more critical than pure numerical precision for trading strategies, establishing XGBoost as a more robust model for capturing market signals in emerging markets like the IHSG.

Keywords : XGBoost; GRU+Attention; IHSG; Walk-Forward Analysis; Directional Accuracy.