Closed-Loop State-Conditioned Flow-Matching Portfolio Policy
Incorporate real-time portfolio state embeddings into the flow-matching velocity network to enable closed-loop, transaction-cost-aware rebalancing.
Source paper
Annan Wang, Yuchen Xie, Yu Cen, Leon Liangyu Chen, Qianyu Julie Zhu, Xin Li, Tianyi Wei, Weisi Lin
Generated by AI. The paper’s authors did not write these.
Incorporate real-time portfolio state embeddings into the flow-matching velocity network to enable closed-loop, transaction-cost-aware rebalancing.
Fine-tune frozen temporal market foundation models using low-rank adaptation during group-relative reinforcement learning to align market representations directly with portfolio objectives.
Introduce constrained factor projection layers into the flow-matching generation process to strictly enforce portfolio factor neutrality and risk constraints during action sampling.
AllocFlow introduces an end-to-end flow-matching framework that directly generates continuous, signed target portfolio weights for multi-asset trading from raw market histories. By coupling a frozen temporal market foundation encoder (Kronos) with a DiT-style cross-asset flow Transformer, AllocFlow bypasses intermediate alpha forecasts and portfolio optimization rules. The policy is trained using a two-stage recipe consisting of strategy-supervised imitation learning followed by Flow-GRPO reinforcement learning with a soft frozen reference policy.
The architecture extracts temporal market tokens per asset using a frozen Kronos foundation model encoder and maps them via an instrument-ID mapping to allocation tokens in a 32-layer Diffusion Transformer flow head. Continuous signed position weights are sampled using stochastic flow matching initialized from Gaussian noise, subject to a dollar-neutral mean-zero projection. Model training consists of Stage 1 strategy-supervised flow matching on synthetic teacher alpha allocations, followed by Stage 2 group-relative reinforcement learning (Flow-GRPO) evaluating 24-hour gross-normalized portfolio Sharpe ratios while regularized by an imitation reference model.