A Multi-Agent Deep Learning And Reinforcement Learning Framework For Stock Price Prediction And Buy–Sell Decision Support In Volatile Equity Markets

Authors: GLN Sravan Kumar, Dr. Kasireddy Sandeep Reddy, K. Santoshini, B. Ramesh

Abstract: This study proposes a multi-agent deep learning framework, termed Neural-Arbitrage, designed to enhance stock price prediction and buy–sell decision support in volatile equity markets. The framework integrates convolutional neural networks, recurrent neural architectures, and deep reinforcement learning to model non-linear price dynamics, volatility clustering, and regime shifts. Using historical secondary market data for Apple Inc. (AAPL) and Tesla Inc. (TSLA), representing contrasting volatility profiles, the proposed model is empirically evaluated against traditional econometric models (ARIMA) and single-agent neural models (LSTM). Model performance is assessed using mean squared error, directional accuracy, and Sharpe ratio. Results indicate that the proposed multi-agent architecture achieves superior predictive accuracy and significantly improves risk-adjusted returns, particularly during high-volatility periods. The findings demonstrate the effectiveness of cooperative learning and reinforcement-driven decision optimization in dynamic financial environments, contributing to the growing literature on intelligent algorithmic trading systems.