Mean–Variance Portfolio Optimization Using Ensemble Learning-Based Cryptocurrency Price Prediction

Abstract The success of portfolio construction largely relies on accurately forecasting future asset performance. Advances in machine learning present significant opportunities to integrate prediction theory into portfolio selection. However, some studies indicate that relying on a single prediction model may lead to biased forecasts or suboptimal portfolio outcomes due to individual algorithms' inherent limitations and assumptions. This study proposes an ensemble learning approach for portfolio optimization by combining advanced predictive analytics with classical mean-variance (MV) models as a two-stage methodology. To do so, seven machine learning models, including LSTM, BLSTM, GRU, DMLP, RF, XGBoost, and SVR, are combined using ensemble approaches such as ridge regression (RR) and principal component regression (PCR). Next, the future insights are integrated into the MV framework to optimize asset allocations, aiming to balance risk and return effectively. An empirical analysis of top cryptocurrencies showed that ensemble learning improves prediction accuracy. Furthermore, the integration of forecasting results with the MV framework outperformed traditional approaches in terms of risk-adjusted returns and portfolio stability.