Paper: arXiv 2611.19937

Authors: Annan Wang, Yuchen Xie, Yu Cen, Leon Liangyu Chen, Qianyu Julie Zhu, Xin Li, Tianyi Wei, Weisi Lin

Abstract

Real-world portfolio management requires converting noisy market observations into coordinated, simultaneous allocations across multiple assets under shared capital constraints. Advances in forecasting models and alpha discovery help identify opportunities. However, traditional pipelines often decouple signal forecasting from portfolio construction, potentially misaligning learning with allocation and execution objectives. We present AllocFlow, an end-to-end flow-matching framework that directly generates joint portfolio allocations from raw multi-asset market histories. To capture dependencies among portfolio positions, AllocFlow employs a cross-asset flow Transformer conditioned on temporal market representations to generate continuous, signed position weights across all instruments jointly. To address the sample inefficiency and instability of learning continuous allocation policies from scratch, we establish a two-stage training scheme: strategy-supervised imitation shapes a structured distribution of candidate portfolios, followed by group-relative reinforcement learning (RL) that refines joint decisions using portfolio-level financial rewards and a frozen policy reference. On hourly cryptocurrency perpetual futures, RL refinement raises the annualized net return of the imitation policy from \(27.79\%\) to \(51.38\%\) and its Calmar ratio from \(0.955\) to \(4.043\), surpassing the teacher strategy on every metric, and AllocFlow reaches a Calmar ratio of \(4.90\) on six further months of unseen 2026 data without retraining.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces a novel end-to-end flow-matching framework for portfolio allocation, combining imitation learning and reinforcement learning. It demonstrates strong empirical results on cryptocurrency perpetual futures with robust backtesting and out-of-sample performance. The methodology is well-articulated, though some mathematical derivations could be more explicit in the excerpt.

Research Flowchart

  flowchart TD
    A["Research Goal: End-to-End Joint Portfolio Allocation from Raw Market Data"] --> B{"Key Methodology: AllocFlow Framework"};
    B --> C["Stage 1: Strategy-Supervised Imitation Learning"];
    C --> D["Stage 2: Group-Relative Reinforcement Learning (RL)"];
    B -- Computational Processes --> E["Cross-Asset Flow Transformer + Flow-Matching"];
    A -- Data/Inputs --> F["Raw Multi-Asset Market Histories (e.g., Hourly Crypto Perpetual Futures)"];
    D & F --> G["Key Findings: Significant Performance Improvement"];
    G -- Outcomes --> H["Annualized Net Return: 27.79% (Imitation) -> 51.38% (RL); Calmar Ratio: 0.955 (Imitation) -> 4.043 (RL); 4.90 Calmar on unseen 2026 data"];

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes AllocFlow, an end‑to‑end cross‑asset flow‑matching Transformer that generates joint continuous signed portfolio allocations directly from raw multi‑asset market histories.

Source sentence
We present AllocFlow, an end-to-end flow-matching framework that directly generates joint portfolio allocations from raw multi-asset market histories.

Assumes 6

  • Trading is performed on hourly cryptocurrency perpetual futures with a flat 5 bps fee per unit of absolute traded weight. execution
    Source sentence
    Orders are changes from current to target weights, including entries and exits when the eligible list changes, and incur 5 basis points (bps, 0.01%) per unit of absolute traded weight. This fee is the maximum of the base-tier taker fees of major perpetual-futures venues and the fees in prior crypto trading papers, and the median half-spread of the traded contracts is only 0.24 bps (Appendix C.3).
  • The eligible asset universe is reconstituted weekly as the 20 assets with highest cumulative trading volume over the prior 60 days. market
    Source sentence
    Every week, the eligible list is recomputed as the 20 assets with the highest cumulative trading volume in the preceding 60 days. This causal reconstitution tests a changing opportunity set rather than a list selected with knowledge of the full test period.
  • The Kronos encoder is frozen after pre‑training that ends in June 2024 and contains no information from the test period. model
    Source sentence
    Kronos pretraining ends in June 2024 [46], before the hourly test from November 2, 2024, to December 31, 2025; thus the frozen encoder has no test-period information in its weights.
  • Orders are executed at the target weight change each hour without additional market impact beyond the stated fee. execution
    Source sentence
    Orders are changes from current to target weights, including entries and exits when the eligible list changes, and incur 5 basis points (bps, 0.01%) per unit of absolute traded weight. Our backtest therefore uses the most conservative setting: it charges the base-tier taker fee on every change of the target book at every hourly rebalance, although a policy that trades toward a target book can execute with limit
Show 2 more
  • All price and volume data are publicly available and free to use. data
    Source sentence
    Evaluation design We use cryptocurrency perpetual futures (futures contracts without expiry that can be held long or short) because granular historical trading records are freely and publicly available over long periods.
  • Flow matching provides a suitable stochastic policy class for joint allocation of signed weights. model
    Source sentence
    Flow matching suits portfolio allocation for three reasons. First, the action is a joint distribution over complete books of signed weights, for example, when several correlated assets are candidates for the same short.

Data

Backtest sample: 2020-01 – 2025-12 · cryptocurrency perpetual futures · crypto · intraday

  • Cryptocurrency perpetual futures (hourly bars) · public exchanges (e.g., Binance, OKX, Hyperliquid) market-prices public
    weekly top‑20 assets by 60‑day volume
    Source sentence
    Evaluation design We use cryptocurrency perpetual futures (futures contracts without expiry that can be held long or short) because granular historical trading records are freely and publicly available over long periods. Policy training and test samples are strictly separated in time.

Costs

Modeled Costs are part of the reported headline results.

A flat 5 bps fee per unit of absolute traded weight is applied to every change of the target book; median half‑spread is 0.24 bps.

Source sentence
Orders are changes from current to target weights, including entries and exits when the eligible list changes, and incur 5 basis points (bps, 0.01%) per unit of absolute traded weight. This fee is the maximum of the base-tier taker fees of major perpetual-futures venues and the fees in prior crypto trading papers, and the median half-spread of the traded contracts is only 0.24 bps (Appendix C.3).

Finds 5

Reported performance (AllocFlow): Sharpe 2.19 · annual return 51.4%

  • Group‑relative RL refinement raises the annualized net return of the imitation policy from 27.79 % to 51.38 % and its Calmar ratio from 0.955 to 4.043, surpassing the teacher strategy on every metric. performance
    Source sentence
    On hourly cryptocurrency perpetual futures, RL refinement raises the annualized net return of the imitation policy from $27.79\%$ to $51.38\%$ and its Calmar ratio from $0.955$ to $4.043$, surpassing the teacher strategy on every metric, and AllocFlow reaches a Calmar ratio of $4.90$ on six further months of unseen 2026 data without retraining.
  • On six months of unseen 2026 data, AllocFlow achieves a Calmar ratio of 4.90 without retraining. performance
    Source sentence
    Out-of-sample period Return ↑ MDD ↓ Sharpe ↑ Calmar ↑ Nov 2024–Dec 2025 (Table 1) 51.38 ±2.04 12.81 ±1.51 2.192 ±0.086 4.043 ±0.379 Jan–Jun 2026 (new) 62.53 12.77 2.06 4.90 C.3.
  • A parameter‑matched Gaussian policy trained with the same objective attains at most a Calmar of 1.92, less than half of the flow policy’s 4.043. performance
    Source sentence
    Across exploration scales σ (Table 2b), it reaches a Calmar of at most 1.92, versus 4.043 for the flow policy, and stays below 1 in the other four settings.
  • Replacing the frozen Kronos encoder with a frozen iTransformer causes severe performance loss (negative returns and Calmar), showing the importance of good market features. performance
    Source sentence
    Replacing frozen Kronos with a frozen iTransformer while keeping the flow head loses heavily after both IL and RL on the same ledger (Table 5), and reward refinement does not recover the Kronos-conditioned performance. With the losing Kronos-direct deployment, this shows that good market features become valuable only when a suitable head learns to allocate from them.
  • Reward ablation shows that training with a Sharpe‑based reward yields the highest Calmar (4.043) while other reward choices either reduce Calmar or produce negative returns. performance figure from elsewhere in the paper
    Source sentence
    Cumulative return gives the highest annualized return, but the Sharpe-trained policy has the smallest MDD and highest Calmar while trading more than the return-focused variants, so its downside improvement is not simply the result of trading less; the risk-adjusted-return variant loses money, so nominal risk adjustment does not by itself ensure a useful training signal. Training reward Return MDD Sharpe Calmar TO IL

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-17 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (66)

  1. Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading (2026) — link
  2. Decision-focused Sparse Tangent Portfolio Optimization (2026) — link
  3. The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting (2026) — link
  4. Online Portfolio Selection with ML Predictions (2025) — link
  5. FinCon : A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (2024) — link
  6. Learning to Trade via Direct Reinforcement (2001) — link
  7. A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem (2017) — link
  8. Online Portfolio Management via Deep Reinforcement Learning with High-Frequency Data (2023) — link
and 58 more
  1. Soft Imitation Reinforcement Learning with Value Decomposition for Portfolio Management (2024) — link
  2. AlphaZeroBeta : Deep Reinforcement Learning for Market-Neutral Portfolios (2026) — link
  3. Smart ``Predict, then Optimize’’ (2022) — link
  4. Deep Hedging (2018) — link
  5. Kronos : A Foundation Model for the Language of Financial Markets — → 2508.02739
  6. iTransformer : Inverted Transformers Are Effective for Time Series Forecasting (2023) — link
  7. Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
  8. AlphaSAGE : Structure-Aware Alpha Mining via GFlowNets for Robust Exploration — → 2509.25055
  9. AlphaAgent : LLM -Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay (2025) — link
  10. R&D-Agent-Quant : A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization — → 2505.15155
  11. STABLE : Shift-Tolerant Allocation via Black-Litterman Using Conditional Diffusion Estimates (2026) — link
  12. Solving Dynamic Portfolio Selection Problems via Score-Based Diffusion Models — → 2507.09916
  13. Decision-Weighted Flow Matching for Contextual Stochastic Optimization (2026) — link
  14. Flow Matching for Generative Modeling (2022) — link
  15. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (2023) — link
  16. \ensuremath \pi_0 : A Vision-Language-Action Flow Model for General Robot Control (2024) — link
  17. Diffusion Policy Policy Optimization (2025) — link
  18. Flow-GRPO : Training Flow Matching Models via Online RL (2025) — link
  19. ReinFlow : Fine-Tuning Flow Matching Policy with Online Reinforcement Learning (2025) — link
  20. Robust Performance Hypothesis Testing with the Sharpe Ratio (2008) — link
  21. The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality (2014) — link
  22. Portfolio Selection (1952) — link
  23. Empirical Asset Pricing via Machine Learning (2020) — link
  24. 101 Formulaic Alphas (2016) — link
  25. Deep Reinforcement Learning for Trading (2020) — link
  26. Learning Stock Trading Policies via Barycenter-Based Adversarial Inverse Reinforcement Learning (2026) — link
  27. Realistic Market Impact Modeling for Reinforcement Learning Trading Environments (2026) — link
  28. Scalable Diffusion Models with Transformers (2022) — link
  29. Proximal Policy Optimization Algorithms (2017) — link
  30. DeepSeekMath : Pushing the Limits of Mathematical Reasoning in Open Language Models (2024) — link
  31. Dynamic Trading with Predictable Returns and Transaction Costs (2013) — link
  32. OpenClaw-RL : Train Any Agent Simply by Talking (2026) — link
  33. Deep Direct Reinforcement Learning for Financial Signal Representation and Trading (2017) — link
  34. One Fits All: Power General Time Series Analysis by Pretrained LM (2023) — link
  35. Time-LLM : Time Series Forecasting by Reprogramming Large Language Models (2024) — link
  36. FinRL : A Deep Reinforcement Learning Library for Automated Stock Trading in Quantitative Finance (2020) — link
  37. Deep Reinforcement Learning for Portfolio Selection (2024) — link
  38. Visual Instruction Tuning (2023)
  39. CLIPort : What and Where Pathways for Robotic Manipulation (2021)
  40. GR00T N1 : An Open Foundation Model for Generalist Humanoid Robots (2025) — link
  41. Deep Learning for Portfolio Optimization (2020)
  42. MARS : A Meta-Adaptive Reinforcement Learning Framework for Risk-Aware Multi-Agent Portfolio Management (2026)
  43. DeepAries : Adaptive Rebalancing Interval Selection for Enhanced Portfolio Selection — → 2510.14985
  44. Regime-Adaptive Continual Learning for Portfolio Management (2026)
  45. EarnHFT : Efficient Hierarchical Reinforcement Learning for High Frequency Trading — → 2309.12891
  46. MacroHFT : Memory Augmented Context-aware Reinforcement Learning On High Frequency Trading — → 2406.14537
  47. FineFT : Efficient and Risk-Aware Ensemble Reinforcement Learning for Futures Trading (2026)
  48. ArchetypeTrader : Reinforcement Learning for Selecting and Refining Learnable Strategic Archetypes in Quantitative Trading (2026)
  49. AlphaGAT : A Two-Stage Learning Approach for Adaptive Portfolio Selection (2025)
  50. The Recurrent Reinforcement Learning Crypto Agent (2022)
  51. Directly Fine-Tuning Diffusion Models on Differentiable Rewards (2024)
  52. Continuous Control with Deep Reinforcement Learning (2016)
  53. Asynchronous Methods for Deep Reinforcement Learning (2016)
  54. Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor (2018)
  55. Addressing Function Approximation Error in Actor-Critic Methods (2018)
  56. Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments (2017)
  57. The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games (2022)
  58. Human-Level Control through Deep Reinforcement Learning (2015)

8 of 66 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.