Enhancing Decision-Making in Digital Asset Markets Through Hybrid Machine Learning and Deterministic Optimization

Abstract

The paper presents a new hybrid approach (LSTM-RF-SLSQP) that acts as an advanced decision-support tool for institutional crypto portfolio management. The methodology includes the use of LSTM neural networks to detect non-linear temporal patterns and RF algorithms to detect structural market noise. The ability to predict future prices based on LSTM-RF model has been extensively verified out-of-sample using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), and shown to have smaller prediction error than standalone algorithms on very volatile assets. Then, based on robust return and risk forecasts, SLSQP optimization algorithm allocates asset weights aiming at maximizing Sharpe ratio under specific institutional constraints, applying buy-and-hold approach with quarterly rebalance. The empirical study is performed for a portfolio of ten major cryptocurrencies (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, TRX, LINK, DOT) using data provided by KuCoin exchange from January 2024 to January 2026. The numerical experiments demonstrated the significant superiority of the suggested framework compared to all benchmarks. Namely, the LSTM-RF-SLSQP approach provides the impressive annual return of 107.40%, Sharpe ratio of 2.04, with maximum drawdown equal to -3.80%.