Authors: Research Scholar Namrata Ramrao Pawar, Dr. Ganesh R. Teltumbade
Abstract: Precise stock prediction and efficient portfolio optimization are still problematic in practice because of the non-stationarity, volatility, and complexity of the financial time series. In this study, we present a Hybrid Artificial Intelligence Framework (HAIF), which is the combination of three different frameworks: (1) a Graph Neural Network (GNN) with an attention layer for modelling relationships between stock prices; (2) a Transformer with convolutional layers for predicting price movements in different periods ahead; and (3) a Deep Reinforcement Learning (DRL) model called Proximal Policy Optimization for managing transactions and balancing the portfolio in different conditions. Based on 5 years of daily S&P 500 time series from 2020 to 2025 with 50 constituent stocks, our model obtains Sharpe ratio = 1.84, annual return = 28.4%, and maximum drawdown = -11.2% while outperforming
