Personalised Multi-Market Investment Advisory Using Reinforcement Learning and Investor Risk Profiling

Abstract

Retail investors in emerging markets, particularly Pakistan, face a pronounced advisory gap: professional portfolio guidance is either inaccessible due to cost and account minimums, or unavailable in a form that covers domestic equities alongside international markets. This paper presents AIPRS (AI-Powered Portfolio Recommendation System), an investment advisory platform that brings three artificial intelligence methodologies together in a single dashboard. A K-Means clustering model segments investors into four behavioural risk profiles (Conservative, Moderate, Aggressive, and Very Aggressive) derived from a seven-dimensional profiling form, and maps each profile to a Modern Portfolio Theory (MPT)-based asset allocation. Two custom Proximal Policy Optimisation (PPO) actor-critic agents, trained from scratch in PyTorch over three million environment steps each, generate personalised BUY, HOLD, or SELL advisory signals for United States and Pakistan Stock Exchange (PSX) equities respectively, with the investor's risk cluster encoded directly in the agent's 24-dimensional state vector and a risk-adaptive confidence threshold applied at inference time. A FinBERT-based natural language processing pipeline scores financial news headlines as Likely Positive, Neutral, Likely Negative, or Inconclusive, providing qualitative market context alongside the quantitative signals, and a user-based collaborative filtering mechanism surfaces asset preferences trending among same-cluster peers. The system is deployed as a live, working dashboard covering both markets end to end. Experimental results are reported for both PPO agents against buy-and-hold and random-action baselines, together with classification performance (precision, recall, F1) on held-out price data and cluster-validity statistics for the investor segmentation model. To the authors' knowledge, few openly documented systems combine personalised RL-based advisory signal generation with investor risk clustering across both a developed and an emerging equity market in a single deployable product.