Paper review · November 18, 2026
AllocFlow: a Flow Matching model that generates portfolio trading decisions directly from raw market data
AllocFlow: End-to-End Portfolio Trading Action Generation from Raw Market Data
The authors initialized a Flow Matching policy by imitating a strategy, then refined it with group-relative reinforcement learning, producing portfolio allocations for crypto perpetual futures that beat the Teacher strategy on every metric.
Problem
Conventional quant pipelines optimize signal prediction (alpha) and portfolio construction separately, so the training objective can diverge from the actual allocation and execution goals. Learning a continuous multi-asset allocation policy from scratch with reinforcement learning alone is sample-inefficient and unstable.
Method
With the pretrained financial foundation model Kronos kept frozen, per-asset time-series Representations are extracted as tokens. Conditioned on these, a DiT-style cross-asset Flow Transformer generates signed target weights for all assets at once. The weights are constrained to sum to 0 (dollar-neutral). In stage 1, the model imitates the decisions of a strategy combining AlphaGen and AlphaSAGE alphas (AgenticAlpha) using Flow Matching and auxiliary losses. In stage 2, the policy is refined with Flow-GRPO, which samples 8 candidate portfolios per state and compares them by the Sharpe ratio obtained when each is held for 24 hours along the same market path. The frozen imitation-learning policy serves as an anchor so updates do not drift too far.
Results
- In an hourly crypto perpetual futures test (14 months), the RL stage raised annualized net return from 27.79% for the imitation-learning policy to 51.38%, Sharpe from 0.954 to 2.192, and Calmar from 0.955 to 4.043.
- After RL, Calmar (4.043) was more than twice that of the Teacher strategy (1.733), and the authors reported beating the Teacher on every metric.
- Without retraining, it achieved a Calmar of 4.90 on an additional 6 months of 2026 data.
- Under the same RL objective, a Gaussian policy with a matched parameter count reached a Calmar of at most 1.92, and directly optimizing the same reward by differentiation reached at most 1.74. Starting RL from random initialization without imitation learning produced an annual loss of 75%.
Limitations
The evaluation is limited to hourly data for the 20 crypto perpetual futures with the highest trading volume. Fees were assumed to be 5bp, and the setup assumes a price taker whose trades do not affect prices. The authors note that dollar neutrality does not guarantee beta neutrality or a risk cap, and that the evaluated decoder has no cap on gross exposure. The provided text contains no validation on other asset classes.
Why it matters
The paper shows an approach that generates target portfolios directly from market data instead of separating prediction from portfolio construction. It transfers Flow policies and GRPO-style reinforcement learning, previously used in image and robotics work, to financial allocation, and is a useful reference for quant practitioners.
Key terms
- Flow Matching
- A generative modeling technique that learns a velocity field moving noise toward target data. Here it is used to generate the weight vector for all assets at once.
- Flow-GRPO
- A reinforcement learning method that samples several candidates for the same situation and updates the Flow policy using their relative rewards within the group.
- Calmar ratio
- Annualized net return divided by maximum drawdown (MDD). It shows return relative to the risk of losses.
- Perpetual futures
- A crypto derivative that lets traders hold long and short positions indefinitely, with no expiry date.
Original
Authors: Annan Wang, Yuchen Xie, Yu Cen, Leon Liangyu Chen, Qianyu Julie Zhu, Xin Li, Tianyi Wei, Weisi Lin · Posted 2026-11-15 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.