AllocFlow: end-to-end flow matching turns raw market data into portfolio allocations
Annan Wang · hf · 2026-11-17
A paper from NTU, Stanford, MIT, and independent researchers presents AllocFlow, an end-to-end flow-matching framework that generates joint portfolio allocations directly from raw multi-asset market histories.
- Motivation: traditional pipelines decouple signal forecasting from portfolio construction, potentially misaligning learning with allocation and execution objectives
- Architecture: a cross-asset flow Transformer conditioned on temporal market representations jointly generates continuous, signed position weights across all instruments
- Training: two-stage imitation learning followed by group-relative reinforcement learning, addressing sample inefficiency and instability of learning continuous allocation policies from scratch
- Results: improved risk-adjusted returns on cryptocurrency perpetual futures
Lead authors Annan Wang and Yuchen Xie; project page at allocflow.github.io.
More from Research
- Reactorfield's first AI-for-Science cohort: agent swarms boost neural simulations from 12 to hundreds per week — MaxUnfried · 2026-11-17
- Latent feedback cuts Transformer training compute by up to 46% with 4 lines of code — IFM · 2026-11-17
- PaperCritic: RL-trained AI reviewers beat frontier models at judging paper soundness — CarnegieMellonU · 2026-11-17
- LoopFill: training-free speculative decoding speeds up looped Transformers — UIUC-CS · 2026-11-17
- SPILLWAY: deterministic defense stops all templated prompt injections on tool-using agents — saintmachines · 2026-11-17
- SJTU's Calpha folds proteins from a general-purpose LLM's internal residue states — SJTU · 2026-11-17