LSTM-Based Portfolio Optimization Across Multi-Market Equity Universes: A Comparative Study of LQ45, DJIA, SSE50, and a Global Portfolio

Abstract

This study integrates long short-term memory (LSTM) return forecasting with mean-variance optimization across Indonesia’s LQ45, the U.S. DJIA, and China’s SSE50 under a leakage-aware protocol (Jan–Feb 2026, 32 trading days). Forecast-free Minimum Variance proves more robust than LSTM-driven Maximum Sharpe, attaining the best Sharpe in three of four universes (LQ45: −0.001, SSE50: 0.015, Global: 0.042; DJIA: 0.145, drawdown −1.9%). A permutation test confirms LSTM forecasts are indistinguishable from random reorderings (p = 0.892). Apparent gains under naive pipelines stem from look-ahead leakage rather than forecasting skill; cross-market diversification (correlations 0.021 vs. 0.223 within-market) reduces risk as theory predicts. Minimum Variance-based guidance is offered for conservative and diversification-oriented investors.