A Three-Phase Foundation Model for Tax-Aware Personalized Portfolio Management
06:00 · July 1, 2026 · arXiv cs.AI RSS
We present a three-phase deep reinforcement learning system for personalized portfolio management that addresses three limitations shared by all prior financial RL work: 1) ticker lock-in, 2) monolithic objectives , and 3) static user models. Phase 1 pretrains a ticker-identity-free cross asset encoder via self-supervised learning on a multi-asset corpus, augmented by a frozen parallel branch using Chronos, a T5-based time series foundation model, fused via a learned gating mechanism. To our knowledge, this is the first application of a time series foundation model to portfolio management RL. The encoder generalizes to any publicly traded asset via a 50-dimensional observable metadata vector that requires no retraining for new tickers. Phase 2 fine-tunes a MoE (Mixture of Experts) portfolio actor critic with PPO under an objective-conditioned reward that simultaneously serves six distinct investment goals sampled per episode: short-term alpha, short-term gain, long-term gain, capital preservation, tax-loss harvesting, and long-term-gains-only. A MoE architecture assigns each objective to a specialized expert head (momentum, growth, defensive, tax-aware), and a learned intent router blends experts based on the active objective and current market regime, which eliminates cross-objective gradient conflict. Phase 3 adds a lightweight personalization layer further adapted at inference time to each individual via a 76-parameter LoRA module fine-tuned on real brokerage transaction history, inferring investment objectives from revealed trading behavior rather than questionnaires. A natural language intent parser converts free-form goals directly into structured investment objective parameters.
Summary
The article presents a three-phase deep reinforcement learning architecture for tax-aware, personalized portfolio management that explicitly targets three longstanding constraints in financial RL: models locked to fixed ticker universes, reward functions that enforce a single monolithic objective such as Sharpe ratio, and user profiles that remain static after initial questionnaire elicitation.
Phase 1 pre-trains a cross-asset encoder through self-supervised learning on a multi-asset corpus. A parallel frozen branch incorporates Chronos, a T5-based time-series foundation model pretrained on more than 100 billion observations, with the two streams combined by a learned gating layer. The resulting encoder represents any publicly traded security through a 50-dimensional metadata vector containing sector, fundamentals, analyst consensus, options signals, earnings dates, insider activity and institutional ownership, thereby removing any dependence on learned ticker embeddings and enabling zero-shot generalization to new assets.
Phase 2 fine-tunes a Mixture-of-Experts actor-critic policy with proximal policy optimization. An objective-conditioned reward samples one of six investment goals per episode—short-term alpha, short-term gain, long-term gain, capital preservation, tax-loss harvesting or long-term-gains-only—while four specialized expert heads (momentum, growth, defensive, tax-aware) and a learned intent router allocate computation according to both the active goal and prevailing market regime. This design isolates gradient updates across conflicting objectives and avoids the need for separate policies per goal.
Phase 3 introduces a lightweight personalization layer that adapts at inference time via a 76-parameter LoRA module fine-tuned on an individual’s real brokerage transaction history. A natural-language intent parser maps free-form statements such as “buy a house in three years” into the structured objective parameters used by the router, allowing preferences to be inferred from observed behavior rather than declared once and frozen. The complete system is exposed through a FastAPI service that integrates live brokerage feeds and presents inferred preferences for user confirmation before any adaptation occurs.
Why it matters
This research is highly relevant for Dutch AI researchers and fintech enterprises looking to deploy advanced, personalized financial AI systems. The integration of foundation models, MoE, and LoRA offers cutting-edge methodologies that can be adapted by the strong Dutch financial sector to improve algorithmic trading and wealth management.
