acceptodds
Withdrawn from ICLR 2027

From Optimization to Generation: Learning Portfolio Distributions with Simplex-Space Flow Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Portfolio construction typically seeks a single optimal portfolio from a market state, while the uncertainty and diversity of feasible investment decisions remain largely unexplored. In this work, we formulate portfolio construction as a conditional generative modeling problem, where the objective is to learn a distribution over feasible portfolio allocations conditioned on the observed market state, rather than predicting a single deterministic portfolio. We introduce Simplex-Space Flow Matching (SS-FM), which learns continuous probability flows directly over the portfolio simplex and generates diverse allocations that reflect different portfolio preferences. SS-FM generates diverse feasible portfolio candidates, while reinforcement learning learns how to make decisions within this generated candidate space using realized portfolio-level feedback, enabling the model to move beyond imitation of predefined portfolio construction rules. To provide informative conditioning signals, we incorporate overnight market information into alpha prediction through gated residual fusion, capturing information that is not contained in intraday price and volume dynamics. Experiments on the CSI 500 and S&P 500 demonstrate that modeling portfolio construction as conditional generation provides an effective alternative to deterministic and conventional optimization-based approaches.

Primary area:
generative models
Keywords:
Generative Modeling,Portfolio Optimization,Simplex space Flow Matching,Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zD7dOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper designs a Simplex-Space Flow Matching (SS-FM) to learn probabilities of portfolio weights under market states, then uses PPO or GRPO to learn a policy to decide the concrete weights from the learned probability distribution.

The paper is poorly written, and the motivations, novelty, and logis are unclear. It is hard to see where the novelty is and why the proposed method can be better than the existing portfolio learning approaches. The experiments are compared with only a few standard policies but not state-of-the-art machine learning or deep/reinforcement learning approaches for portfolio selection and construction.

Critical strengths

None

Critical weaknesses
  • The novelty and motivations are unclear. Neither learning a portfolio probability distribution nor a policy to determine the exact portfolio is new, and it is unclear how new and effective it is to combine these two together.
  • The method description doesn't show clearly the novelty and contributions, and why the proposed method can be better than existing methods.
  • The experiments are simple and not convincing. No state-of-the-art machine/reinforcement learning approaches in FinTech is included in the comparison. The parameter setting looks ad hoc and not systemtic. There are only two datasets, which is not comprehensive enough.
Questions for authors
  • Where is the novelty of your proposed method? It looks like every component is existing.
  • After the portfolio weight distribution WtW_t has been learned, only hth_t is used to realise the actual portfolio. hth_t is also used in learning WtW_t. Therefore, the overall end-to-end process can be seen as learning an actual portfolio, with WtW_t as an intermediate output/layer. Please explain the differences.
  • In the first learning stage, the MVO, MS, and RP portfolios are used as teacher portfolios. In machine learning, students cannot be better than teachers. Do you aim to learn portfolio distributions similar to MVO, MS, and RP rather than better than them? Please explain more clearly why use these teacher portfolios.
  • How to set the number of assets K?
  • How do the fcrossf_{cross}, ftempf_{temp}, faucf_{auc} and fαf_{\alpha} operators in equations (7)-(9) work?
  • How is πϕ\pi_{\phi} in equation (18) represented?
  • Why are the the number of generated candidates is G = 32 and transaction cost is 3 basis points in the experiments?
  • Which time periods of the datasets were used? I saw the figures are from 2025-2026, which is very short. How about other years and periods?
Reviewer_HtGEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Problem. Long-only portfolio construction over a daily Top-30 universe, framed as learning a distribution over simplex-valued allocations conditioned on the market state instead of predicting one allocation.

Method.

  • A Transformer market encoder over ten days of intraday features, with cross-stock attention. For CSI 500, overnight and opening-auction features are added by gated residual fusion. The same encoder predicts next-day Open-to-Close return and ranks assets to form the Top-30 universe.
  • SS-FM. Conditional flow matching in centred log-ratio (CLR) coordinates of the simplex, trained to transport noise to teacher portfolios from MVO, Maximum Sharpe, and Risk Parity. At inference it generates G=32G=32 candidate allocations.
  • Decision layer. PPO either selects one candidate (categorical policy) or outputs weights directly from (ht,Wt)(h_t, W_t). GRPO initialised from SS-FM is also tested. Reward is a weighted sum of normalised net return, Sharpe, turnover, and cost.

Evaluation. Monthly walk-forward with a 12-month training window, one validation month, and one test month. Test period May 2025 to May 2026 on CSI 500 and June 2025 to May 2026 on S&P 500. Transaction cost 3 bps. Baselines are Gaussian, MLP, Diffusion, and Euclidean FM generators and the three optimisation teachers.

Claims.

  • SS-FM has the highest single-run net return among generators on CSI 500 and is the only generator with positive return on S&P 500.
  • PPO on top of frozen SS-FM raises CSI 500 net return from 60.8% to 105.1% and lowers turnover.
  • SS-FM has lower seed variance than Diffusion under Softmax aggregation.
Critical strengths

1. The problem framing is reasonable. Treating allocation as conditional generation over the simplex, with CLR coordinates to respect the compositional constraint, is a sensible way to represent allocation uncertainty, and separating proposal generation from execution choice is a clean decomposition.

2. The walk-forward protocol is stated carefully. Training, validation, and test months are strictly separated, normalisation and teacher construction use training data only, and multi-seed sensitivity across candidate budgets is reported rather than hidden.

Critical weaknesses

1. The headline results are single runs that disagree with the paper's own multi-seed tables, and the most relevant prior work and baselines are missing.

Setting (CSI 500, G=32)SS-FMDiffusionFM
Table 2, single run, Pure60.8%31.9%49.0%
Table 5, 20 seeds, Pred-Utility41.2% ± 18.3%73.2% ± 29.0%25.7% ± 10.3%
Table 5, 20 seeds, Softmax50.5% ± 2.7%53.0% ± 3.5%51.8% ± 1.1%

The 60.8% headline is more than one standard deviation above its own 20-seed mean, and under both multi-seed protocols Diffusion outperforms SS-FM on CSI 500. The claim that SS-FM is the best generator on CSI 500 holds only for the single run.

The closest prior work, Diffolio (Jeon, Lee and Kang, WSDM 2026), which also generates a distribution over portfolio weights instead of a single allocation and adds risk guidance, is not cited. STABLE (Soun et al., ICLR 2026) is not cited either. The Diffolio that is cited (Cho et al., 2026) is a return-forecasting model and is not run as a baseline. Equal Weight is listed as a baseline in Section 4 but appears in no table. No market index return is reported. No RL baseline that learns weights directly from the state is run, although Jiang et al. (2017) and Winkel et al. (2024) are cited and the authors acknowledge the gap in Limitations. The best result in Table 4 comes from the Candidate-to-Weight PPO, which outputs weights directly without selecting from the generated set, so the generated candidates are not what drives the top number.

2. The contribution of the allocation method cannot be separated from the alpha model and the universe construction, and the evaluation is too short and too cheap.

The universe is the daily Top-30 by predicted next-day return from a shared alpha model. Stock selection is therefore done before any allocation method runs, and it is likely the dominant return source, yet alpha IC is never reported and no Equal Weight result on the same universe is given. The paper does not state as of which date the CSI 500 and S&P 500 constituent lists were taken, leaving survivorship bias unaddressed. The test period is one year (about 245 trading days), the training window is one year, and the S&P 500 RL result rests on three seeds. The transaction cost of 3 bps is unrealistic for CSI 500 constituents, where stamp duty and commissions put round-trip costs above 10 bps, and SS-FM's one-way daily turnover of 0.649 makes the result highly cost-sensitive. The overnight and auction fusion is listed as a contribution but never ablated.

3. The novelty is limited and the code is not released.

CLR coordinates are the standard compositional transform, flow matching is applied unchanged, the targets are imitations of classical optimisers, and the decision layer is standard PPO or GRPO over a restricted candidate set. No repository is provided, so none of the results can be reproduced.

Questions for authors

Q1. Reconcile single-run and multi-seed results. Why does the headline Table 2 report a single run when Table 5 shows a 20-seed mean of 41.2% ± 18.3% for the same setting? Please report means and intervals in the main tables and revise the CSI 500 claim given that Diffusion is higher under both multi-seed protocols.

Q2. Isolate the allocation method from stock selection. Report the alpha model's IC and the return of Equal Weight on the same daily Top-30 universe. Report the CSI 500 and S&P 500 index returns over the test period.

Q3. Universe construction. As of which date were the CSI 500 and S&P 500 constituent lists taken? If the final-period list was used, re-run with point-in-time constituents.

Q4. Missing prior work and baselines. Diffolio (Jeon, Lee and Kang, WSDM 2026) directly generates portfolio-weight distributions with diffusion and is the closest prior work to this paper's framing. Please position SS-FM against it and run it as a baseline, together with STABLE (Soun et al., ICLR 2026) and at least one RL policy that learns weights directly from the state (e.g. Jiang et al., 2017, or Winkel et al., 2024).

Q5. Role of the candidate set. The best method (SS-FM + PPO, 105.1%) outputs weights directly. What is the result of the same Candidate-to-Weight PPO when the SS-FM candidates are removed from its input? This is the ablation that would show whether generation matters.

Q6. Simplex ablation. Under 20 seeds, Euclidean FM and Diffusion match or beat SS-FM on CSI 500. Please state precisely under which protocol the CLR parameterisation helps and why.

Q7. Transaction costs and horizon. Report results at 10 and 20 bps, and extend the evaluation to at least two further non-overlapping years.

Q8. Auction fusion. Provide a with/without ablation of the overnight and auction features, since this is listed as a contribution.

Q9. Code. Please release the code and the recorded artifacts listed in Appendix D.3.

Reviewer_wLFHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether generating multiple candidate portfolios for a given market state improves portfolio allocation. It trains a flow-matching model on allocations from mean–variance, maximum Sharpe, and risk-parity optimizers, then uses reinforcement learning to choose the final allocation under a reward that accounts for returns, risk, and trading costs.

Critical strengths

None

Critical weaknesses

The main weakness is that the paper does not explain why portfolio construction requires a learned generator. A standard approach is to define a utility function that accounts for expected return, risk, turnover, and trading costs, then tune the weights of those terms to reflect the investor’s preferences and optimize it directly. The paper motivates its generator by appealing to diverse investor preferences, yet its final policy is trained on a single fixed reward. A comparison against a direct optimizer targeting the same trade-offs would help establish whether the added complexity improves the resulting allocations.

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.