A Hybrid Multi-Agent Framework with Reinforcement Learning for Personalized Financial Planning and Asset Allocation

Personal financial planning is a sequential decision problem shaped by income variability, expenditurepatterns, liquidity constraints and long-horizon goals. Conventional tools are rule-based and weaklyadaptive. This study presents a hybrid multi-agent framework that combines structured financialanalysis, reinforcement learning for strategy selection, and natural language explanation. The systemintegrates specialized agents for risk assessment, goal feasibility evaluation, heuristic asset allocation,and DQN-based strategy selection. A compact five-dimensional state representation encodes the riskscore, goal feasibility, equity allocation, savings rate, and financial runway. Trained on 3,000 episodesof synthetic data (seed=42), the final 100-episode mean reward improved by 13.1% over the trainingmean. Evaluation of 100 fixed scenarios showed that the learned policy outperformed a heuristicbaseline by 70.3% in cumulative reward, 49.1% in terminal balance, and 30 percentage points in goalon-track rate. These findings constitute simulation-based evidence that a compact RL policy cancomplement rule-based financial analysis; however, generalization to real-world settings requiresfurther validation.