INTEGRASI MACHINE LEARNING DAN OPTIMASI PORTOFOLIO: TRADE-OFF ANTARA AKURASI PREDIKSI DAN KINERJA INVESTASI PADA PASAR SAHAM INDONESIA

Abstract

The surge of retail investors in the Indonesian stock market highlights a critical gap between market accessibility and financial literacy, leading to irrational decision- making. While machine learning (ML) models are increasingly adopted to navigate market volatility, high predictive accuracy is often misconstrued as a guarantee of optimal investment returns. This research investigates the empirical trade-off between predictive accuracy and economic performance by comparing a hybrid ML framework (LSTM, XGBoost, and Random Forest) with conventional methods (Markowitz Mean- Variance Optimization and Equal-Weight) in the Indonesian stock market. Using data from highly liquid stocks (LQ45 and IDX80) from 2019 to 2025, the results show that the single XGBoost model achieved the highest economic performance (Sharpe Ratio 0.96; CAGR 29.59%). Conversely, the hybrid framework yielded the lowest prediction error (RMSE 0.0418) but lower nominal returns, confirming a fundamental trade-off between predictive superiority and optimal asset allocation. Robustness analysis proves the integrated hybrid framework maintains positive risk-adjusted returns during highly volatile market crises. This study concludes that ML integration significantly improves investment decisions for novice investors, provided that predictive outputs are systematically synchronized with precise risk management strategies.