acceptodds
Under review as a conference paper at ICLR 2027

AllocFlow: End-to-End Portfolio Trading Action Generation from Raw Market Data

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. The primary value of imitation is to provide a coherent starting region for optimization, rather than a standalone performance target. In retrospective cryptocurrency evaluation after modeling trading fees, AllocFlow achieves an annualized arithmetic net return of and Calmar , establishing the value of our design.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
algorithm trading, quantitative trading, financial foundation models, flow matching, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HNFjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the gap between market forecasting and portfolio construction by directly generating joint signed allocations from raw multi-asset histories. It combines a frozen Kronos encoder with a cross-asset flow Transformer and uses imitation learning to initialize plausible portfolio proposals. It then refines these proposals with portfolio-level rewards to improve trading performance.

Critical strengths
  • The paper reports substantial performance gains after RL refinement across several evaluation metrics.
  • The paper provides relatively detailed descriptions of the training, inference, and evaluation procedures.
  • The proposed method is clear and easy to understand.
Critical weaknesses
  • Unclear motivation for the use of flow matching. Some statements in this manuscript are overstated.
  • Insufficiently justified method design, especially the RL formulation and optimization objective.
  • Insufficient empirical validation, including weak controls and limited market coverage.
Questions for authors
  1. The motivation for using flow matching is unclear. The paper does not establish why flow matching is needed for portfolio allocation. Continuous joint actions and stochastic candidate generation can also be achieved with simpler stochastic Transformer policies.
  2. The Stage 2 formulation is closer to a contextual bandit than dynamic portfolio RL. The action does not affect future market states, and current portfolio holdings are not included in the policy input.
  3. The RL training objective is misaligned with deployment. Training optimizes the Sharpe ratio of a fixed 24-hour portfolio with normalized gross exposure, while deployment generates hourly portfolios with variable exposure and transaction costs.
  4. The empirical baselines and controls are insufficient. The current comparisons do not isolate the contribution of flow matching. Stronger matched stochastic-policy baselines under the same RL objective and model capacity are needed.
  5. The use of policy-gradient-based Flow-GRPO is insufficiently justified. Since both the financial objective and the flow generation process are largely differentiable, the paper should justify why policy-gradient optimization is preferable to direct differentiable optimization.
  6. The experiments only consider cryptocurrency markets. Evaluation on broader equity universes, such as the Nasdaq-100 or S&P 500, is needed to support generalization.
6 comments

We thank the reviewer for the careful review and the valuable suggestions. We address the reviewer's concerns below.

Q1 and Q4. Motivation for flow matching and matched stochastic-policy baselines

The new controls show that flow matching is central to effective RL refinement in AllocFlow. Under the identical GRPO objective and matched model capacity, Flow-GRPO reaches a Calmar of 4.043 ±0.379, while the best stochastic Transformer policy reaches 1.92.

The one-pass Transformer in Table 2 is exactly the stochastic Transformer policy described by the reviewer. It is a bidirectional Transformer with about 1.0B parameters, matched to the flow head and reading the same tokens. It outputs a mean book μ and samples actions from N(μ, σ²I). Its stage 2 uses the same GRPO code as the flow head, with eight candidates per state, the same 24-bar gross-normalized Sharpe reward, the same clipping, KL and distillation terms, and 2000 steps, and differs from the flow head only in the policy kernel, a Gaussian instead of the flow SDE, which is what Section 5.1 of the submission refers to as different RL training kernels. We report it for exploration scales σ from 0.05 to 0.8.

Policy (matched capacity, same GRPO objective)Return % ↑Sharpe ↑Calmar ↑
Gaussian Transformer, IL start (Table 2)27.780.9490.878
Gaussian Transformer + GRPO, best σ (0.4)36.21.281.92
Flow Transformer + Flow-GRPO (ours)51.38 ±2.042.192 ±0.0864.043 ±0.379

The Gaussian policy is also sensitive to its exploration scale: the other four values of σ (0.05, 0.1, 0.2 and 0.8) all stay below a Calmar of 1 (−0.55, −0.62, 0.96 and −0.12; the σ = 0.2 run is the one in Table 2) and trade about three times the book per day. Starting from imitation policies of similar quality (Calmar 0.878 and 0.955), RL adds 3.1 to the Calmar of the flow policy, three times its gain for the Gaussian policy.

Flow matching fits portfolio allocation for three reasons. First, the action is a joint distribution over books of 20 signed weights, for example when choosing which of several correlated assets to short. Second, group-relative RL compares candidate books, so the candidates should be coherent alternatives. The flow samples complete alternative books from the learned distribution, while a diagonal Gaussian perturbs a single mean book coordinate by coordinate, which in the sweep above raises turnover to about three times the book per day. Third, the stochastic sampler of the flow provides the closed-form transition likelihoods that Flow-GRPO optimizes.

We have added the σ sweep to Table 2 and state this motivation directly in Section 3 of the revised manuscript.

Q2. Contextual-bandit formulation of stage 2

Stage 2 is deliberately formulated as a contextual bandit over target books. As in standard portfolio RL for liquid markets (e.g., Jiang et al., 2017), the policy is a price taker and its trades do not move future prices. Holdings therefore affect outcomes only through trading costs, which the ledger charges on every change of the target book. Under this formulation the reward measures the quality of a complete book without the credit-assignment noise of long, path-dependent rollouts. We state this formulation explicitly in Section 3.3 of the revised manuscript.

AllocFlow also extends directly to the closed-loop setting: with the current holdings as input, the closed-loop policy reaches a Calmar of 4.59 with a turnover of 1.38, compared with 4.043 and 1.976 for AllocFlow (table under Q3). We start from one of the Flow-GRPO policies (seed 42) and give its action head the current holdings as an additional input. For each asset, six holdings features (the position, its signed logarithm, whether the asset is held, and the gross and net exposure of the book) are embedded by a two-layer MLP and added to the action token of that asset. The last layer of this MLP is zero-initialized, so training starts exactly from the trained policy. Training runs for 2000 steps with a learning rate of 1e-5 and a batch size of 16.

Q3. The RL training objective and deployment

Putting deployment terms into the reward lowers the deployed performance. With gross exposure in the reward, the Calmar falls from 4.043 to 2.76, and with fees in the reward, it falls to 1.98.

The reward is a controlled comparison (Section 3.3): the eight candidate books of the same market state are held over the same 24-hour path at the same gross exposure, so their reward differences come only from their composition, not from their leverage. Every reported result is measured on the deployed ledger, where the policy generates a new book with its own gross exposure every hour and pays the fee on every rebalance. We compare our reward with the two aligned alternatives described by the reviewer and with the closed loop of Q2. The alternatives are single runs with the same seed as the policy of Q2, and AllocFlow is reported as in Table 1.

Training objectiveMDD % ↓Sharpe ↑Calmar ↑TO ↓
AllocFlow (Table 1)12.81 ±1.512.192 ±0.0864.043 ±0.3791.976
Sharpe scaled by gross exposure18.11.762.762.07
Sharpe after fees26.01.371.981.30
Closed loop of Q2 with holdings input, fine-tuned from AllocFlow11.92.154.591.38

Putting gross exposure into the reward raises the MDD to 18.1%. Charging fees in the reward lowers the turnover to 1.30, but the MDD rises to 26.0%. Combined with a no-trade band at execution, AllocFlow reaches a Sharpe of 2.46 and a Calmar of 5.75. We add this comparison to Section 5.5 of the revised manuscript.

Q5. Why policy-gradient Flow-GRPO rather than direct differentiable optimization

Policy-gradient optimization performs substantially better. From the same IL checkpoint and with the same reward, Flow-GRPO reaches a Calmar of 4.043 ±0.379, while direct differentiable optimization reaches at most 1.74.

For the differentiable alternative, we follow DRaFT and back-propagate the same gross-normalized 24-bar Sharpe through the deterministic 4-step ODE for 2000 steps. The loss combines the Sharpe reward with an L1 anchor to the IL book and a penalty on gross exposure, and the table lists the learning rate and the weights of the reward and the anchor in each setting.

SettingLearning rateReward weightAnchor weightCalmar ↑Sharpe ↑TO ↓
IL start (Table 1)---0.9550.9541.143
DRaFT 11e-50.11.01.141.071.17
DRaFT 21e-51.01.01.040.873.59
DRaFT 35e-50.11.01.741.411.19
DRaFT 41e-51.00.10.950.804.50
Flow-GRPO (ours, Table 1)---4.043 ±0.3792.192 ±0.0861.976

Policy gradients are preferable for three reasons. First, Flow-GRPO only needs reward values, so it also applies to rewards computed on books rounded to the execution grid, such as the deployed after-fee P&L, which provide no useful pathwise gradient. Second, each pathwise gradient follows a single realized 24-hour price path and pushes the book toward whatever that path rewarded. A larger reward weight therefore mainly increases turnover, to 3.59 and 4.50 per day in DRaFT 2 and 4. Group-relative scoring instead standardizes rewards within each group of candidates and clips the advantage, which makes it robust to heavy-tailed returns. Third, the stochastic sampler that generates diverse candidate books also provides the closed-form likelihoods that Flow-GRPO optimizes. We have added this comparison to Section 5 of the revised manuscript.

Q6. Evaluation on equity universes

AllocFlow targets crypto perpetual futures. The market trades around the clock, allows long and short positions at the same fee, and has a universe of the 20 most-traded contracts that rotates every week. The hourly bars, the 24-bar horizon, the rotating universe and the cost model of AllocFlow are built for this structure. We thank the reviewer for this suggestion, and the revision states this scope explicitly in Section 1.

Wording of claims

We thank the reviewer for this comment. In the revision, every claim cites its supporting table or figure.

We hope these responses address the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score. Please feel free to let us know if there are any further questions.

Reviewer_K6jDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to generate a complete long–short target portfolio directly from multi-asset market histories. The authors develop a two-stage framework: Stage 1 trains a cross-asset flow Transformer, conditioned on frozen Kronos representations, to generate continuous joint allocations through strategy-supervised flow matching; Stage 2 retains a frozen copy of the imitation policy as a reference and further updates the same action head through group-relative RL.In an hourly backtest, the authors report 51.38% annualized arithmetic net return, 2.192 Sharpe, and 4.043 Calmar. Ablations examine the action head, teacher strategy, encoder, model depth, and reward choice.

Critical strengths

With the same encoder, action-head design, and fee model, the authors report that annualized net return rises from 27.79% to 51.38%, while maximum drawdown falls from 29.10% to 12.81%. Improving both measures after modeled fees is the clearest point in favor of acceptance, within the reported historical backtest.

Critical weaknesses

Every reported RL run starts from an imitation-trained policy; there is no RL run from a random action head under matched conditions (§5.2, Table 4). The results show that RL improves the IL policies, but cannot show whether IL improves on training RL directly. This matters because the two-stage training design is a central contribution.

Questions for authors

What exact rules are used to select and combine the AlphaGen and AlphaSAGE factors into AgenticAlpha? What is AgenticAlpha’s standalone result under the same evaluation ledger?

5 comments

We thank the reviewer for the detailed review and for highlighting the improvement of both return and drawdown after fees. We address the reviewer's concerns below.

W1. RL from a random action head under matched conditions

Without imitation learning, Flow-GRPO from a randomly initialized action head loses 75% per year, while the same RL stage after imitation reaches a Calmar of 4.043 ±0.379.

Following the reviewer's suggestion, we train Flow-GRPO from a randomly initialized action head with the same encoder, sampler, reward and number of RL steps as the reported run. The two reference terms are switched off because there is no imitation policy to anchor to.

Training (same encoder, head and reward)Return % ↑MDD % ↓Sharpe ↑Calmar ↑
Random head without training−81.194.4−2.11< 0
RL from random head, 4.8× steps21.1171.50.150.12
IL + Flow-GRPO (ours)51.38 ±2.0412.81 ±1.512.192 ±0.0864.043 ±0.379

With the same number of RL steps, the policy stays at the level of an untrained head and loses 75% per year. With 4.8 times more RL steps, which roughly matches the total compute of IL and RL, its gross exposure drifts to 4.5 times the capital and its drawdown exceeds the capital. With imitation learning, the same RL stage raises the Calmar from 0.955 to 4.043 ±0.379.

This follows from how group-relative RL learns. GRPO learns from the relative rewards of eight candidate books sampled for the same market state. When the head is random, the candidates are close to random portfolios and their reward differences are dominated by noise, so the group-relative advantage carries little information. Imitation learning provides a proposal distribution of coherent books in which the comparison becomes informative, and RL then improves the policy substantially. GRPO is used the same way in other domains: DeepSeekMath applies it to a supervised model and Flow-GRPO applies it to a pretrained image generator.

The same conclusion holds beyond GRPO. AllocFlow achieves four times the Sharpe of the best pure-RL method (MARS) with one quarter of its drawdown, and no pure-RL agent reaches the Sharpe of buying and holding BTC. We train pure-RL trading methods from scratch on our data: three recent methods, MARS (AAAI 2026), ReCAP (KDD 2026) and DeepAries (CIKM 2025), and eight classic RL algorithms. Each method uses its published architecture and hyper-parameters, adapted to our rotating universe, with the same data, split, fee and ledger as AllocFlow. Methods that are long-only by design are kept long-only. For each classic algorithm we train long-only and long-short variants (19 agents in total) and report the better one, which is long-only in every case.

MethodReturn % ↑MDD % ↓Sharpe ↑
MARS (AAAI 2026)25.952.30.50
DeepAries (CIKM 2025)−13.189.1−0.15
ReCAP (KDD 2026)−253.1*409.5−2.03
DDPG−19.483.9−0.23
PPO−35.9106.3−0.41
A2C−55.8129.2−0.63
SAC−59.3130.0−0.67
TD3−80.5151.1−0.91
MADDPG−6.845.3−0.10
MAPPO−48.3114.3−0.58
DQN−76.7129.4−1.06
BTC buy-and-hold23.837.70.55
IL + Flow-GRPO (ours)51.38 ±2.0412.81 ±1.512.192 ±0.086

*Returns are annualized arithmetic returns on fixed notional as in the paper, so very high turnover combined with the fee can exceed −100%.

Leading RL methods for crypto trading make a closely related design choice. EarnHFT (AAAI 2024), MacroHFT (KDD 2024) and FineFT (KDD 2026) all guide their agents with a hindsight-optimal oracle, either as a regularizer or through demonstrations, and the IL stage of AllocFlow plays the same role with a teacher strategy. We have added these results to Section 5.2 and Appendix C of the revised manuscript.

Q1. Construction of AgenticAlpha and its standalone result

AllocFlow surpasses its teacher on every metric. On the same ledger, AgenticAlpha alone reaches a Calmar of 1.733, while AllocFlow reaches 4.043 ±0.379 and lowers the drawdown from 19.57% to 12.81%.

AgenticAlpha is built only from training data. AlphaGen and AlphaSAGE each mine a set of alphas on the training period and combine them linearly. AgenticAlpha then combines the two methods at the portfolio level: the AlphaGen alphas and the AlphaSAGE alphas each manage half of the capital, and the two halves form one portfolio. Signals are converted into positions by the allocation pipeline of Appendix B.1: temporal smoothing over 24 hourly bars, volatility scaling with target 0.05, a gross allocation cap of 10 units, a per-name clip of ±2 units and a 0.1-unit execution grid. The table below lists the five alphas with the largest absolute weights from each method, and we add them to Appendix B.1. We will release all the original AlphaGen and AlphaSAGE alphas used by AgenticAlpha.

SourceAlpha expressionWeight
AlphaGenRef(Greater($vwap,-2.0),1d)+0.1494
AlphaGenWMA(Greater(Div(-5.0,Std(Greater($low,-0.5),48d)),-0.01),12d)−0.1476
AlphaGenLog($vwap)+0.0207
AlphaGenLess(2.0,Greater(Less(Greater($open,Mul($vwap,-2.0)),1.0),$open))−0.0049
AlphaGenSub($volume,Max(Greater(Div(5.0,$volume),0.01),1d))+0.0020
AlphaSAGEDiv(Pow(Less($open,5.0),$high),$volume)+0.0238
AlphaSAGETsMad(Add(-5.0,Mul(TsDiv($volume,12),10.0)),48)−0.0229
AlphaSAGETsRank(Sub(-30.0,$high),120)+0.0218
AlphaSAGETsStd(TsDiv(Rank(Mul(0.01,Mul(Abs($vwap),-0.5))),120),120)−0.0192
AlphaSAGESign(Mul(-1.0,TsDelta($vwap,336)))−0.0188

Window lengths are in hourly bars; the AlphaGen grammar writes them with the suffix d.

Policy (same ledger and 5 bps fee)Return % ↑MDD % ↓Sharpe ↑Calmar ↑
AgenticAlpha (teacher)33.9219.571.4721.733
AllocFlow after IL27.7929.100.9540.955
AllocFlow after IL + Flow-GRPO51.38 ±2.0412.81 ±1.512.192 ±0.0864.043 ±0.379

Imitation alone stays below the teacher: it is trained to match the teacher's books, not its return, and serves as the starting distribution for RL (Section 3.2). RL then lifts the policy well above the teacher, and this improvement does not require a profitable starting point (Table 4): with MACD (12h, 26h), whose imitation policy loses money, RL still reaches a Calmar of 1.856.

We hope these responses address the reviewer's concern about the two-stage design, and we would be grateful if the reviewer would consider raising the score. Please feel free to let us know if there are any further questions.

Reviewer_GjRhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper describes AllocFlow, an end-to-end generative model that maps raw multi-asset price and volume histories directly to portfolio weights. It combines a frozen financial encoder (Kronos) with a Diffusion Transformer flow head. The paper makes the following contribution:

  • A cross-asset flow transformer that generates a complete long-short portfolio.
  • The paper uses imitation learning to train the model to copy a teacher strategy so it starts with sensible positions.
  • The paper then uses reinforcement learning to further update the model towards better performance.
  • The paper validates the performance of the proposed method on crypto long-short trading.
Critical strengths

The idea of building a trading strategy using foundational model as feature representation and using IL/RL for portfolio construction is interesting. Nonetheless, i am not sure if this is super novel or surprising.

Critical weaknesses
  • there are a lot of jargons and out-of-context description in the paper, making it read like ai-generated and hard to parse. generally speaking, a term should be defined before it is used. Some jargons in the finance domain (albeit understandable by me ) should be clarified for general audience (bar, OHLCV). Kronos is a foundation model. etc.

  • The validation is a bit limited. only one year of out of sample period is tested. The evaluation setup also seems to optimistic (e.g. does not account for funding rate and bid-ask spread).

  • I am also not sure if this framework is indeed end-to-end in the conventional sense as the feature representation relies on a frozen foundation model, and the learning seems mostly focus on portfolio optimization.

Questions for authors

None.

4 comments

We thank the reviewer for the valuable comments and address each concern below.

W1. Terminology and definitions

We thank the reviewer for pointing this out. We revised the manuscript so that every term is defined before its first use, and added a terminology table to the appendix. The definitions are:

TermDefinition in the revision
BarMarket data aggregated over one fixed time interval. AllocFlow uses 1-hour bars.
OHLCVThe open, high, low and close prices and the traded volume within a bar.
KronosA foundation model for financial markets (Shi et al., 2025), pretrained to predict the next bar from tokenized OHLCV sequences. AllocFlow keeps it frozen and uses its hidden states as features.
Perpetual futureA futures contract without expiry that can be held long or short. Holders of long and short positions exchange periodic funding payments.
Long-short target bookThe signed weights over the 20 eligible assets that the policy outputs every hour. Positive weights are long positions and negative weights are short positions.
Gross exposureThe sum of absolute weights, ‖w‖₁.
Turnover (TO)The average daily sum of absolute weight changes.
Basis point (bps)0.01%. A 5 bps fee costs 0.05% of the traded value.
Sharpe ratioAnnualized mean of daily net returns divided by their standard deviation.
Maximum drawdown (MDD)The largest peak-to-trough loss of cumulative net return.
Calmar ratioAnnualized net return divided by MDD.
Teacher strategyThe strategy whose portfolios the action head imitates in stage 1.

W2. Out-of-sample period and evaluation setup

We extend the out-of-sample test to June 30, 2026. On unseen data from January to June 2026, AllocFlow reaches a Sharpe of 2.06 and a Calmar of 4.90 without retraining, a higher Calmar than on the original test (4.043).

Test length. Our original test runs from November 2024 to December 2025, which gives 14 months and 10,198 hourly out-of-sample decisions. It covers distinct market regimes, including the rally at the end of 2024, the sell-off in early 2025 and a long decline of altcoins. Following the reviewer's comments, we further evaluate the same trained policy, without retraining, from January 1 to June 30, 2026.

Out-of-sample period (5 bps fee)FundingReturn % ↑MDD % ↓Sharpe ↑Calmar ↑
Nov 2024 to Dec 2025 (14 months, Table 1)without51.38 ±2.0412.81 ±1.512.192 ±0.0864.043 ±0.379
Nov 2024 to Dec 2025 (14 months)with50.22 ±2.2312.88 ±1.562.048 ±0.1013.932 ±0.386
Jan 2026 to Jun 2026 (new, 6 months)with62.5312.772.064.90

The out-of-sample period now totals 20 months, more than twice as long as the test windows reported by the most recent RL methods for crypto trading:

MethodMarket and barTest window
AlphaGAT (IJCAI 2025)crypto portfolio of 18 assets7 months (Sep 2023 to Apr 2024)
ArchetypeTrader (AAAI 2026)single crypto asset, 10 min8 months (Jan 2024 to Sep 2024)
AllocFlow (ours)crypto perpetuals portfolio, 1 h20 months (Nov 2024 to Jun 2026)

Funding. Following the reviewer's suggestion, we also settle funding for every contract on its own 4-hour or 8-hour schedule, using the funding history published by the exchange; the 2026 period is evaluated with funding. Because long and short positions largely offset each other's funding, every metric stays within its variation across seeds: the mean ± standard deviation intervals with and without funding overlap for return, MDD, Sharpe and Calmar (first table above). Funding therefore has a negligible effect on the results.

Fee. The 5 bps fee is the maximum over the actual base-tier fees of major perpetual-futures venues, centralized and decentralized, and the fees used in prior crypto trading papers, as listed below.

Perpetual-futures venueTypeTaker, base tierMaker, base tier
Binance USDⓈ-MCEX5.0 bps2.0 bps
OKXCEX5.0 bps2.0 bps
HyperliquidDEX4.5 bps1.5 bps
Coinbase InternationalCEX4.0 bps2.0 bps
Published crypto trading paperMarketFee per trade
Borrageiro et al. (IEEE Access 2022)BitMEX perpetual futures, 5 min5 bps
EarnHFT (AAAI 2024)Binance, second-level2 bps
MacroHFT (KDD 2024)Binance, minute-level2 bps
ArchetypeTrader (AAAI 2026)crypto, long-short, 10 min2 bps
FineFT (KDD 2026)Binance perpetual futures, 5 min2 bps
AllocFlow (ours)crypto perpetual futures, 1 h5 bps

Higher volume tiers, fee discounts (10% when paying in BNB on Binance, up to 40% for staking on Hyperliquid) and maker orders only lower the venue rates. 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 orders at a maker fee of at most 2 bps. All methods in the paper are evaluated with the same fee.

Bid-ask spread. AllocFlow trades high-volume USDⓈ-M perpetual contracts, re-selected every week by their trading volume over the past 60 days. These contracts have deep order books and are typically quoted at a spread of one tick. Half a tick corresponds to 0.24 bps of price for the median traded asset-hour, less than 5% of the 5 bps fee charged on every unit of traded weight (0.44 bps when weighted by traded volume and 0.006 bps for BTC).

W3. End-to-end design

AllocFlow learns the entire trading decision end to end, from raw OHLCV histories to the signed portfolio, in contrast to a forecast-then-optimize pipeline. Kronos serves solely as a representation encoder: its hidden states condition the DiT, which directly generates portfolio actions, and its next-bar forecasts are never used as inputs to a separate trading strategy. Kronos is designed for forecasting rather than trading: traded directly, its own forecasts lose 44.5% per year after fees zero-shot and 34.6% per year after LoRA adaptation (Table 3). AllocFlow therefore takes only its representation, and every trading decision is learned by the DiT. The action head receives no return forecast, risk model or alpha factor. From the generic features of a model pretrained to predict the next bar, it decides which of the 20 changing assets to hold long or short, how large each position should be and when a trade is worth its fee. Four results show that this decision is learned by the head:

  • On the same frozen features that lose money when Kronos trades its own forecasts, the AllocFlow head earns 51.38% per year (Table 1).
  • On identical features, the flow head reaches a Calmar of 4.04, compared with 1.31 for an MLP head and 0.96 for a one-pass Transformer (Table 2).
  • With a randomly initialized head and no imitation, the same RL loses 75% per year (see our response to Reviewer K6jD), so the features alone do not determine profitable trades.
  • AllocFlow surpasses its own teacher on every metric, with a Calmar of 4.043 compared with 1.733 (see our response to Reviewer K6jD), so the head does more than reproduce the teacher's portfolio construction.

Keeping a pretrained encoder frozen is common practice in systems described as end-to-end. LLaVA (NeurIPS 2023, Section 4.2) describes its second stage as end-to-end fine-tuning while keeping the vision encoder frozen. CLIPort (CoRL 2021, abstract and Section 3.1) describes an end-to-end framework that uses internal features from a frozen pretrained CLIP encoder for action prediction. GR00T N1 (NVIDIA, 2025), which combines a pretrained vision-language backbone with a diffusion transformer action module, an architecture similar to that of AllocFlow, is trained end-to-end while the language component of its backbone stays frozen. In portfolio learning, end-to-end also refers to learning allocations directly under a portfolio objective instead of predicting returns and then optimizing (Zhang, Zohren and Roberts, Deep Learning for Portfolio Optimization, JFDS 2020), and AllocFlow is end-to-end in this sense as well.

We hope these responses address the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score. Please feel free to let us know if there are any further questions.

Thank you again for your valuable comments on our paper. We would like to kindly follow up on our responses, which address your concerns as follows:

  • Terminology: We define every term at its first use and added a terminology table to the appendix.
  • Out-of-sample period: We extended the test to 20 months, more than twice the test windows of the most recent RL methods for crypto trading. On the new period from January to June 2026, AllocFlow reaches a Calmar of 4.90 without retraining.
  • Funding and bid-ask spread: Settling funding keeps every metric within its variation across seeds, and the median half-spread is only 0.24 bps, against a 5 bps fee that is the maximum of the base-tier fees of major venues and the fees used in prior work.
  • End-to-end design: Kronos serves only as a representation encoder. Its own forecasts lose 44.5% per year when traded directly, while the action head learns the full trading decision and surpasses its teacher on every metric.

We would be very grateful if you could let us know whether these responses address your concerns, and if so, we kindly hope you would consider updating your score. We are also happy to clarify anything that remains unclear.

Comments on the paper

1 comment

We sincerely thank the AC, and the reviewers. As the discussion period comes to a close, we summarize the motivation, novelty and evidence of AllocFlow, followed by how the discussion resolved each concern.

Motivation. Progress in learning-based trading has largely focused on better forecasts and alpha factors. Yet a better forecast does not by itself make a better trade: its value depends on how it is turned into positions, combined with other opportunities and rebalanced over time. Pipelines that optimize signals separately from portfolio construction therefore leave a gap between what the model learns and what a trading account needs. A portfolio decision is inherently joint, as the size and direction of each position must be chosen in relation to all the others.

Novelty. AllocFlow closes this gap by making the trading decision itself the object of learning. A cross-asset flow Transformer generates the complete signed portfolio in one decision, directly from raw market histories, with no intermediate forecast, alpha score or optimizer, while a frozen financial foundation model only supplies the market representation. Training follows a two-stage recipe: imitation of a strategy provides coherent candidate portfolios, and group-relative RL with a frozen reference then refines complete portfolios under portfolio-level rewards. To our knowledge, AllocFlow is the first method that generates the complete multi-asset portfolio as the action of a flow policy and refines it with RL, bringing flow-based generative modeling, which has advanced image generation and robot control, to portfolio decision-making.

Evidence. After fees, RL raises the Calmar ratio of the imitation policy from 0.955 to 4.043, more than twice that of its teacher, and the policy keeps its performance on six further months of unseen 2026 data without retraining. K6jD regarded this joint gain in return and drawdown as the clearest point in favor of acceptance, and HNFj noted the substantial gains from RL. Matched controls show that each component is necessary, because a Gaussian policy, direct differentiable optimization or the removal of imitation each substantially degrades performance.

How the concerns were resolved

  1. K6jD raised the rating. K6jD confirmed that the new experiments resolve the concerns about the imitation stage and the teacher, and the remaining question about overfitting in model selection is answered by the unseen 2026 period, which was never used for training or model selection.

  2. HNFj's design questions are settled by the matched controls above, and the follow-up asks only for equities. The baselines HNFj cites never carry a model across markets, because each re-mines its factors and sets its own trading rule per market, and Table 1 applies the same protocol to crypto. AllocFlow learns the trading rule itself, so equities require retraining under different conditions such as restricted short selling on CSI, which is a separate study rather than a flaw. Crypto perpetual futures are also the testbed of recent trading methods at AAAI, KDD and NeurIPS.

  3. Every concern of GjRh is addressed in the revision. GjRh rated these concerns as minor or fixable. The revision adds a terminology table, extends the test to 20 months, shows that settling funding leaves every metric within its variation across seeds and that the median half-spread is far below the fee, and defines end-to-end explicitly. GjRh has not disputed any of these answers.

We hope you will take these points into account in the final decision. If you find the idea of learning the complete portfolio decision with a generative policy promising, we would be sincerely grateful for your support in recommending AllocFlow for acceptance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.