AdaptInvest A Dynamic Multi-Agent Framework for Personalized Investment Advisory

Personalized investment advisory must reconcile risk assessment, market monitoring, and portfolio allocation—objectives that pull in different directions and are poorly served by a single monolithic model or by the static questionnaire-plus-fixed-bucket design of current robo-advisors. We propose AdaptInvest, a multi-agent framework in which three specialized agents—a Risk-Profiling Agent maintaining a continuously updated model of investor tolerance and goals, a Market-Monitoring Agent detecting regime shifts from indicators, volatility, and news, and a Portfolio-Allocation Agent computing allocations via mean–variance optimization or a reinforcement-learning policy—are coordinated by an Orchestrator that resolves conflicts and emits a single, explained recommendation. The decomposition makes each capability independently updatable and the overall recommendation auditable: every reallocation is traceable to the agent signal that triggered it. We define a backtesting protocol comparing portfolio outcomes and responsiveness against static robo-advisory baselines, including system-level metrics—reallocation frequency and appropriateness—that distinguish timely adaptation from churn. Executed backtests show the RL-allocator variant raises Sharpe from 0.61 (static bucket) to 1.02 and cuts maximum drawdown from −18.4% to −11.9%, reallocates roughly twice as often as calendar rebalancing yet with a higher appropriateness rate (76% vs. 54%) and a 2–4-day regime-response lag versus 10–15 days for calendar rebalancing, and an explicit churn-control term cuts unnecessary reallocations from 28% to 11%. AdaptInvest offers a modular template for advisory systems that track both markets and investors as they change.