acceptodds
Under review as a conference paper at ICLR 2027

CASA-RL: Exact Set Policies for Sparse Portfolio Control

Abstract

Sparse portfolio reinforcement learning couples unordered asset selection with continuous allocation, while financial priors complicate attribution of learned gains. We introduce CASA-RL, a framework for ardinality-ware parse llocation built around three requirements. First, an exactly normalized fixed-cardinality distribution assigns probability directly to the selected asset set. Second, a support-conditioned logistic-normal allocation yields feasible stock/cash weights and an explicit joint likelihood for PPO. Third, matched initialization, support-prior ablations, and checkpoint comparisons separate prior-based performance from learned changes. In a 45-run study on reconstructed historical S&P 500 constituents, the validation-selected configurations of CASA-RL and two ordered-policy implementations achieve approximately % mean cumulative net return on a retrospective test interval. The selected trained policies match their initialized counterparts' stock supports on every test day. Removing the explicit support prior exposes substantial policy changes and checkpoint-dependent method rankings. The resulting framework aligns feasible portfolio actions, likelihood-based learning, and empirical attribution.

Primary area:
reinforcement learning
Keywords:
reinforcement learning / quantitative computing

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_qjNROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines an exactly normalized distribution over fixed-size asset sets with a conditional logistic-normal allocation policy. The joint likelihood supports PPO updates for sparse portfolios. Experiments compare this policy with ordered alternatives and examine how initialization, financial priors, and checkpoint selection affect the reported returns.

Critical strengths
  1. The action representation is coherent. The set distribution enforces cardinality, while the conditional allocation assigns positive weights to the selected stocks and cash. Sections 3.2–3.4 explain how both decisions enter one evaluable likelihood, avoiding a separate sparsification step after sampling.
  2. The initialization controls are informative. Section 5 shows that almost all of the selected policies' return is already present before training, and that removing the support prior changes both policy behavior and sensitivity to checkpoint selection. Reporting these results makes the contribution of learning easier to assess.
Critical weaknesses
  1. The benefit of the exact set policy remains unestablished. Its main distributional components are acknowledged prior constructions, so evidence that their combination improves learning is particularly important. In Table 1, CASA-RL gains only about 0.0022 percentage points over initialization; without the support prior, its advantage reverses at the terminal checkpoint. These results support the attribution analysis, but provide limited evidence for the value of the proposed policy representation.
  2. The evaluation provides limited evidence of generalization. Section 5.1 states that the 174-day test interval was inspected during earlier development, and Section 5.4 notes unequal earlier tuning budgets. The five seeds all use the same historical path. The disclosure is helpful, but the reported results cannot establish that the observed method differences persist outside the development setting.
Questions for authors
  1. With priors and checkpoint selection held fixed, what evidence distinguishes the learning behavior of the exact set policy from the ordered policy? Learning curves or gradient-variance measurements would help interpret the checkpoint-dependent results.
  2. Which design choices were informed by the retrospective test interval, and do the method differences persist on an untouched interval after those choices are fixed?
Reviewer_NcDkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies sparse portfolio optimization. The key challenge is that portfolio actions combine discrete, unordered stock selection with continuous allocation weights.

Critical strengths

None

Critical weaknesses

The submission presents some interesting ideas for addressing the methodological challenges identified by the authors, including fixed-cardinality constraints, discrete optimization, and incorporating financial priors into reinforcement learning.

However, from a practical trading perspective, the proposed solutions appear rather naive compared with established portfolio construction techniques. For example, sparsity can be encouraged through position penalties or thresholding, while portfolio allocation can be handled using more conventional optimization methods that explicitly account for risk factors, transaction costs, and other trading constraints. The paper does not sufficiently demonstrate why the proposed RL framework offers advantages over these classical approaches. Intuitively, I could not have a good reason on why the proposed solution can really better solve the fundamental question of portfolio optimization.

Furthermore, the experimental setup raises significant concerns. The number of independent training and evaluation samples is limited, the test period is relatively short, and the portfolio contains only a few stocks. These limitations make it difficult to assess the robustness and practical relevance of the reported results, particularly given that most of the performance appears to come from the financial prior rather than RL training.

Questions for authors

See the previous question.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.