Tuesday, 17 November 2026 · 7 min read
AllocFlow generates joint portfolio allocations from raw market data with improved returns
AllocFlow uses a frozen market encoder and a flow Transformer to jointly generate continuous portfolio weights, trained first by imitation of a strategy and then refined by group-relative reinforcement learning, doubling returns and improving risk metrics on crypto futures.
Based on the paper AllocFlow: End-to-End Portfolio Trading Action Generation from Raw Market Data by Annan Wang, Yuchen Xie, Yu Cen, Leon Liangyu Chen and 4 others
Challenges of joint portfolio allocation from raw market data
Portfolio management requires converting noisy, multi-asset market data into coordinated allocations that respect capital constraints and risk considerations. Traditional pipelines separate forecasting signals from portfolio construction, which can misalign learning objectives and lead to suboptimal allocations.
A portfolio's value depends not only on individual asset forecasts but also on how positions relate to each other, including direction, size, and exposure. Coordinated long and short positions can exploit relative opportunities, so decisions must be made jointly rather than independently.
Learning continuous, joint portfolio allocations directly from raw market data is difficult due to the large continuous action space, noisy outcomes, and the need to balance exploration with trading costs. Reinforcement learning can optimize financial outcomes but struggles with sample inefficiency and instability without a good initialization.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short