Multi-Agent Reinforcement Learning Based Financial Portfolio Optimization and Risk Prediction System

DOI

https://www.doi.org/10.59256/ijsreat.20260606015

Published In

Volume 6, Issue 6 (2026)

Published Online

November-December 2026

Pages

95-104

Article Type

Original Article

Check for updates

Portfolio management under uncertain, non-stationary market conditions is difficult for static allocation methods, which adapt poorly to changing risk. This paper presents a modular multi-agent framework that couples machine-learning-based risk prediction with reinforcement learning (RL) for risk-aware portfolio allocation. Four specialized agents (Market Analysis, Risk Prediction, RL, and Portfolio Optimization) exchange information through a shared coordination interface. The Market Analysis Agent derives technical indicators, including moving averages, RSI, and MACD, from historical OHLCV data. The Risk Prediction Agent uses a Random Forest regressor to estimate risk scores. The RL Agent, configured with Proximal Policy Optimization (PPO) and offering Deep Q-Network (DQN) as an alternative, produces BUY, SELL, and HOLD decisions. The Portfolio Optimization Agent assigns asset weights under diversification and rebalancing rules, and a Streamlit dashboard exposes the pipeline to users. A prototype demonstration on four U.S. equities (AAPL, TSLA, AMZN, GOOGL) with a notional investment of $100,000 produced risk scores between 0.25 and 0.65 and an RL-based allocation within one percentage point of equal weighting for every asset. Risk-model error metrics, RL training evidence, backtested return and drawdown statistics, and baseline comparisons are not yet available and are identified as required further evaluation. The work therefore documents an architecture and a working prototype and does not establish performance gains.