LSTM vs. Traditional Models in Portfolio Optimization: A Systematic Review of Predictive Accuracy and Diversification Outcomes

AbstractThis study systematically reviews the literature at the intersection of Long Short-Term Memory (LSTM)-based prediction models and Modern Portfolio Theory (MPT)-driven portfolio optimization. It evaluates two core dimensions: (1) the predictive accuracy of LSTMs versus traditional statistical methods across diverse market regimes, and (2) the diversification benefits of integrating LSTM forecasts into MPT frameworks. The review quantifies the incremental value of LSTM-MPT integration, identifies methodological gaps like LSTM's sensitivity to small datasets, and offers insights for asset managers and policymakers. It advocates for adaptive strategies that leverage ML’s predictive power while retaining MPT’s foundational principles.Keywords: LSTM, Modern Portfolio Theory, predictive accuracy, diversification, financial forecasting, hybrid models.