New Foundation Model for Tax-Aware Personalized Portfolio Management

A new research paper details a three-phase deep reinforcement learning system designed for personalized, tax-aware portfolio management. This innovative approach addresses common limitations in financial reinforcement learning, such as reliance on specific tickers, monolithic objectives, and static user models. The first phase involves pretraining a cross-asset encoder using self-supervised learning on a multi-asset corpus, augmented by a time series foundation model called Chronos. This encoder can generalize to any publicly traded asset without retraining, using a 50-dimensional metadata vector. The second phase fine-tunes a Mixture of Experts (MoE) portfolio actor-critic with PPO, optimizing for six distinct investment goals simultaneously, including tax-loss harvesting. An intent router blends expert advice based on the active objective and market conditions. Finally, a lightweight personalization layer is added in phase three. This layer adapts to individual users at inference time by fine-tuning a LoRA module on real brokerage transaction history, inferring investment objectives from actual trading behavior rather than questionnaires. A natural language parser converts free-form goals into structured parameters.