A research team has published a three-phase deep learning system that tries to make AI-driven portfolio management actually adapt to individual investors — taxes and all.
The system, described in a preprint, chains three stages. First, it trains an encoder on a broad asset corpus using a T5-based time series model called Chronos, producing a representation that works on any publicly traded asset without retraining — just a 50-point metadata vector per ticker. Second, it fine-tunes a "Mixture of Experts" actor-critic model that can simultaneously pursue six investment goals: short-term alpha, short-term gains, long-term gains, capital preservation, tax-loss harvesting, and long-term-gains-only strategies. A routing layer blends the relevant expert heads depending on the active goal and market regime, which the authors say reduces the gradient conflicts that plague multi-objective training. Third, a 76-parameter LoRA module adapts the whole thing to a specific user at inference time by reading their actual brokerage transaction history — no questionnaire required.
The significance is less about any single technique and more about the combination. Most prior financial RL research locks the system to a fixed list of tickers and a single objective; when a user's goals or holdings change, the model has to be retrained. The authors claim all three of those limitations are addressed here. Whether that holds outside a research setting — with real tax rules, real latency constraints, and real brokerage data pipelines — is a different question.
The system would face stiff regulatory headwinds in practice: personalized investment advice from an automated model sits squarely in SEC registered-investment-adviser territory, and the gap between a promising arXiv preprint and a compliant product is rarely small.