Multi-Agent Reinforcement Learning Based Financial Portfolio Optimization and Risk Prediction System (International Journal Of Scientific Research In Engineering & Technology)

Summary

• The paper proposes a modular four-agent framework (Market Analysis, Risk Prediction, RL, and Portfolio Optimization) for risk-aware portfolio allocation.p.2
• Allocation constraints require weights summing to one and a 40% single-asset cap, with transaction-cost penalties.p.2
• The Risk Prediction Agent applies HOLD for risk scores ≤ 0.4 and SELL for scores ≥ 0.7.p.4
• PPO and DQN are each configured for 50,000 training timesteps.p.7
• On four U.S. equities, TSLA received the highest risk score (0.65) and AMZN the lowest (0.25).p.8
• The prototype produced risk scores of 0.25 to 0.65 and RL allocations of 24.46% to 25.84%, which the authors say do not establish improved returns.p.10

Key insights

1. Market analysis, risk prediction, RL decision-making and portfolio optimization are separated and connected through a shared coordination interface.p.2
2. A Random Forest risk model and RL agents (PPO primary, DQN alternative) generate BUY, SELL and HOLD decisions.p.2
3. The risk score is a normalized value in [0, 1] for each stock.p.4
4. The RL weights sum to 100% and all lie below the 40% cap for the $100,000 demo portfolio.p.9
5. A 75% directional-accuracy target and a 20% Sharpe-ratio improvement target are requirements, not results.p.8
6. The RL allocation does not follow the risk ranking: the highest-risk stock received the largest RL weight.p.9
7. All return, risk-adjusted performance and constraint-compliance claims are not yet validated experimentally.p.5