Behavior-Aware Machine Learning Models for Real-Time Personalized Financial Advisory and Portfolio Optimization

Abstract

Traditional robo-advisory systems are based on static risk profiling and rule-based allocation strategies that do not change according to changing market conditions and changing preferences of the investors, resulting in sub-optimal risk-adjusted performance. This study proposes a behavior-aware personalized portfolio optimization framework that dynamically adapts the asset allocation to the investor's risk tolerance. The approach uses a multisource dataset of 110 equities on the NASDAQ and includes market stress indicators from Yahoo Finance data such as volatility and drawdown. A Proximal Policy Optimization (PPO) agent with the Hierarchical Risk Parity (HRP) weights is implemented and a behavior-aware reward function is developed to maximize the Sharpe metric while penalizing the drawdown, turnover and violating any risk. The model is tested with a walk-forward backtesting protocol with monthly rebalancing and transaction cost modelling to make sure that performance is realistically tested. Experimental results may be observed that the sharpe and sortino ratios of the proposed method are higher, the maximum drawdown is lower, and the turnover is lower than those of Mean-Variance, Risk Parity, HRP, and standard PPO baseline in the results. In addition, the volatility analysis tolerance provides an effective personalization for conservatives, moderate and aggressive investors.