A STUDY ON PORTFOLIO OPTIMIZATION STRATEGIES USING HYBRID DEEP LEARNING AND REINFORCEMENT LEARNING APPROACHES

Abstract

A critical part of investing is to optimise the portfolio management, enabling the maximisation of profits whilst reducing
risk. This study optimizes a portfolio using a hybrid deep learning model that incorporates Gated Recurrent Units (GRU) and
Bidirectional Long Short-Term Memory (BiLSTM) networks, since the complexity of financial markets may outstrip the
applicability of classical methods based on statistical models and historical data. The historical S&P 500 stock prices from October
2016 to 2023 were gathered and preprocessed using operations for missing value imputation, redundancy elimination, data
denoising, normalisation, and one-hot encoding. The most significant predictors identified by feature importance analysis were
momentum, liquidity, and volatility dynamics. The data were partitioned into training and test datasets, and the proposed hybrid
model was evaluated against LSTM, XGBoost, and Linear Regression models. The hybrid GRU-BiLSTM model outperforms
conventional approaches across assessment metrics such as R2 (95.80), RMSE (14.523), MAE (10.379), and MAPE (0.005). With
improved accuracy and strength in stock price forecasting and portfolio optimization, the results validate the hybrid model's
usefulness in both short-term and long-term time dependencies.