Demographic-Aware Deep Portfolio Selection for Sip Management Using a Hybrid Temporal Transformer–Policy Network

Resumen

Systematic Investment Plans (SIPs) need personalization that goes beyond generic risk scores. Conventional robo-advisors underuse demographic signals (age, income, occupation, location, goals) that strongly shape risk capacity and liquidity needs. How can we fuse investor demographics with market dynamics to select and rebalance mutual-fund SIP portfolios that better align with life-stage and risk appetite while remaining robust to regime shifts? We propose DATS-PS (Demographic-Aware Temporal Selector for Portfolio Selection), a new deep learning framework. A Demographic Encoder (tabular transformer) produces a latent “risk-capacity vector.” A Market Temporal Transformer ingests multi-factor fund time series (returns, drawdowns, macro proxies). A Risk-Aware Policy Head (distributional actor-critic) outputs SIP weights under budget/solvency constraints. Training is multi-objective: maximize risk-adjusted return (Sortino), minimize downside CVaR, penalize turnover, and enforce goal-attainment via target-tracking loss. A life-event simulator perturbs demographics (e.g., income shocks, relocation) to learn adaptive allocations. On historical mutual-fund data with synthetic demographic cohorts, DATS-PS improves 12-month goal-tracking accuracy and reduces 95% CVaR versus baselines, while lowering turnover. Portfolios stay closer to investor objectives across life-stage changes, yielding higher satisfaction proxies and improved resilience.