AllocFlow, a new AI framework, uses flow Transformers and reinforcement learning to optimize multi-asset portfolio trading with significant performance gains.
A new AI framework, AllocFlow, has been introduced to revolutionize portfolio management by generating joint portfolio allocations directly from raw multi-asset market data. The research, published on November 17, 2026, on arXiv (2611.19937) and supported by an open-source implementation on GitHub, demonstrates significant advancements in trading system design and performance.
AllocFlow addresses a key limitation in traditional trading pipelines, which often separate signal forecasting from portfolio construction, potentially misaligning objectives. Instead, AllocFlow employs a cross-asset flow Transformer to process raw hourly market data and generate continuous, signed position weights for all portfolio instruments simultaneously. This approach ensures coordinated and optimized allocation decisions under shared capital constraints.
The framework utilizes a two-stage training scheme to overcome inefficiencies in learning continuous allocation policies. The first stage, strategy-supervised imitation, establishes a structured distribution of candidate portfolios by mimicking a causal alpha strategy. The second stage, group-relative reinforcement learning (Flow-GRPO), refines these portfolios by sampling and scoring eight candidate books per market state over the next 24 hourly bars. This refinement process optimizes portfolio-level financial rewards while maintaining stability through a frozen reference policy.
On hourly cryptocurrency perpetual futures, AllocFlow demonstrated remarkable performance improvements. 043. 90 without retraining, surpassing the teacher strategy on all metrics.
Source link







