acceptodds
Withdrawn from ICLR 2027

Regularization-Aligned Learning for Sparse Portfolio Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse portfolio optimization seeks high risk-adjusted return from a small set of assets. Optimizing portfolios using estimated returns and risk can amplify estimation error and impair out-of-sample performance. We refer to systematic modifications intended to mitigate this instability broadly as regularization. Learned return forecasts provide one form of regularization, and decision-focused learning incorporates portfolio performance into model training. Yet the existing decision-focused sparse tangency method evaluates dense soft portfolios during training and selects coordinates that mix assets, so its learning objective need not reflect the intended sparse asset decisions. Tangency decisions alone also leave forecast scale unidentified, although it affects signal aggregation and uncertainty adjustment. In this work, we introduce Regularization-Aligned Sparse Tangent Portfolio Optimization (RA-STPO) to make the form, scale, and use of regularization explicit. It learns bounded corrections to a frozen prediction anchor by averaging losses on resampled sparse return paths, then uses calibration, measured uncertainty, and past relative performance to control their deployment. In the main comparison across five markets, RA-STPO achieves the highest average Sharpe ratio and terminal wealth. Our code is available at https://anonymous.4open.science/r/RA-STPO-A210.

Primary area:
optimization
Keywords:
decision-focused learning, sparse portfolio optimization, robust optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GehaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors studied long-only, fully invested portfolio optimization under the assumption that there is no short selling or borrowing. In their study they identified two key shortcomings of the DF-STPO model; first, the DF-STPO model is trained using dense soft portfolios; second, the DF-STPO model operates in the Cholesky-transformed coordinate space rather than the asset space. The authors developed an alternative approach called RA-STPO. RA-STPO learns to apply bounds to each asset-specific correction around a fixed prediction-based anchor. During the training phase, RA-STPO generates random samples of sparse asset subsets. For each generated sample, RA-STPO solves a non-negative quadratic allocation problem.

Finally, RA-STPO optimizes a loss function based upon the Sharpe Ratio. At test time, RA-STPO calibrates and ensembles the corrected signals, accounts for uncertainty, and applies a causally-updated performance weighting factor to combine the corrected portfolio with the original anchor portfolio. The authors provided guarantees regarding approximation, scale-invariance, differentiability, and robustness. The authors evaluated RA-STPO against four other portfolio optimization techniques by applying them to data from five different equity markets. Their results indicated that RA-STPO had the highest average Sharpe Ratio and terminal wealth among all the techniques tested. The authors performed additional analyses including ablation studies and sensitivity analysis; however, their evaluation used the current index constituents and did not account for transaction costs.

Critical strengths

The RA-STPO model has the ability to evaluate real-time (and) true sparse, long-only, unit-budget allocations during training using a different method than simply doing a dense relaxation and then making a "hard" choice in what we will call evaluation. One important thing to note is that the sampled asset sets are dependent on the anchor being fixed and not dependent on the correction parameters. As such, the lack of a selection-gradient term can be seen as justified for the specified objective and not as an unrecognized approximation to a selection rule that depends on parameters.

Critical weaknesses

Appendix D.3 has indicated that PFL and DF-STPO checkpoints were chosen based on the average validation rank information coefficient, whereas the RA-STPO correction checkpoints were selected based on the average validation Sharpe ratio. These selections have a significant impact on how you compare results. If your main evaluation metric is the Sharpe ratio, then this selection process will affect both how you choose models and which learning approach you use.

In addition, Appendix D.5 has fixed DF-STPO's loss mixture and soft-selection sharpness at α=0.5\alpha=0.5 and β=10\beta=10, without evaluating how they respond to validation. Even if we match the number of networks used in each case, these differences remain.

Although the no-correction ablation study is useful, it does not demonstrate that the bounded correction and resampled sparse-path objective performs better than a less complicated anchor-based decision-focused fine-tuning control, given that all three share the same allocator, loss, and checkpoint criteria. Similarly, the J=1J=1 sensitivity study continues to randomly draw a new face for each epoch; therefore, it does not allow us to determine the benefits of resampling as opposed to a deterministic anchor-face control. Both of these points are important because the primary contribution of this work lies in developing this specific learning construction, rather than simply demonstrating that using decision-focused training can enhance a prediction-focused model.

Questions for authors

I'm interested in why you've selected DF-STPO using validation rank information coefficient instead of using validation sharpe ratio. Additionally, I'd like to know if your results from the primary comparison would be different if your checkpoints were chosen in a similar fashion to how you've selected the checkpoints of RA-STPO based on portfolio-level validation. Lastly, can you explain to me how you decided to select α\alpha and β\beta and if there was a validation search involved? Your answer will greatly assist me in interpreting the improvements that your method has shown compared to the most similar decision-focused baseline.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Geha (1/2)

We thank the reviewer for the careful and detailed reading, and especially for recognizing that the parameter-independent proposal law justifies omitting the selection-gradient term. The questions about fair comparison were exactly the right ones, and we ran every requested control. Entries are five-market mean annualized Sharpe ratios (SR). All additional models were trained with the paper's code and configuration.

W1, Q1 (asymmetric checkpoint criteria). Thank you for identifying this asymmetry. In the paper, the forecasting baselines (PFL and DF-STPO) were selected by validation rank IC, a criterion that does not depend on the allocator. The correction was selected by the validation SR of its allocated portfolio. We agree that this makes the comparison harder to interpret, so we crossed the criteria in both directions.

MethodCheckpoint criterionMean SR
RA-STPOvalidation SR (paper)2.357
RA-STPOvalidation rank IC2.350
PFL-32: OSCAR / mSSRM-PGA / ASMP-AFBA / greedyvalidation rank IC (paper)2.077 / 1.852 / 1.900 / 2.054
PFL-32: same allocatorsvalidation SR of greedy allocation1.741 / 1.822 / 1.705 / 1.901
DF-STPO-32validation rank IC (paper)1.826
DF-STPO-32validation SR of its own evaluation rule1.537

Selected by rank IC, RA-STPO stays above every baseline (2.35). Selecting the baselines by validation SR lowers their test SR rather than raising it. The main comparison is therefore not driven by the asymmetric criteria. These results also show that one year of validation SR is a noisy selection criterion, which we discuss under Q2.

W2, Q2 (DF-STPO's α and β). Thank you for this question. In the paper, α = 0.5 followed the fixed default of the DF-STPO authors, and β = 10 was set without a validation search. The DF-STPO authors state in their public discussion that they used β = 5 in all of their experiments, so we searched α ∈ {0.1, 0.5, 0.9, 1.0} × β ∈ {5, 10, 20}. Each configuration was trained with four seeds, and each member was checkpointed by validation SR. Configurations were then ranked by mean validation SR over the five markets. Test SR is shown for every configuration.

αβValidation SRTest SR
1.0201.750 (selected)1.732
0.9201.7431.756
0.510 (paper setting, validation-SR checkpoints)1.6981.524
0.551.6641.537
0.9101.6031.624
0.1101.5941.418
0.951.4731.602
0.151.4131.506
1.0101.3401.602
0.510 (paper setting, rank-IC checkpoints)1.2711.792
0.1201.2201.921
1.051.2001.502
0.5201.1681.682

Retrained with 32 members, the selected configuration (α = 1.0, β = 20) reaches a test SR of 1.857, or 1.870 when the same forecasts are allocated by greedy. Both are close to the default configuration (1.83) and well below RA-STPO (2.36), so tuning does not change the comparison. The best test value in the grid (1.92, with four members) could not have been identified from validation.

More importantly, the rank correlation between validation and test SR across these 13 configurations is −0.03 (p = 0.92). With a one-year validation window, hyperparameter and checkpoint selection is close to random. This applies equally to our own checkpoint selection, and we regard it as a genuine weakness of our evaluation protocol.

AuthorsOpenReview ↗

Response to Reviewer Geha (2/2)

W3, W4 (simpler anchor-based control; deterministic anchor-face control). Thank you for these two suggestions; they turned out to be the most informative experiments for us. All controls share the frozen anchors, the sparse allocator, the batch-SR loss, the checkpoint rule, and the deployment, and each uses 16 seeds. With all switches at their defaults, the control code reproduces the paper's RA-STPO members exactly.

ControlCorrectionTraining facesCompleteCorrected only
RA-STPO (paper)bounded, m + σ⊙tanh(g)resampled around the anchor, J = 82.3572.260
Deterministic anchor faceboundedfixed anchor face (J = 1, no perturbation)2.3502.269
Current faceboundedface of the current corrected signal2.2872.202
Unbounded correctionlinear, m + σ⊙gresampled, J = 82.2122.118
Decision-focused fine-tuning of the PFL networknetwork itselfcurrent face2.0841.939
Same, with resampled facesnetwork itselfresampled, J = 82.1172.047
No correction (anchor only)––2.046–

Most of the gain comes from parametrizing the update as a bounded correction around a frozen anchor. This adds 0.24–0.27 over decision-focused fine-tuning of the predictor itself, and about 0.15 of it comes from the bound itself (bounded vs. unbounded correction). Resampled sparse paths add essentially nothing over the deterministic anchor-face control (+0.01). No market shows a significant difference (HAC p-values: ESTX 0.89, FTSE 0.29, KOSPI 0.82, NIKKEI 0.88, S&P 0.63). The reviewer's concern was therefore well founded: the paper's attribution of the gain to resampling is not supported by these controls.

Survivorship and transaction costs (noted in the summary). Thank you for noting both issues. Restricting the S&P 500 panel to the constituents at the start of the test period removes 27 later entrants (e.g., SMCI, KKR, APO). This lowers every method's SR (no retraining).

MethodCurrent constituents (465)Point-in-time constituents (438)Change
Historical / OSCAR2.4442.088−0.36
Historical / mSSRM-PGA1.4420.781−0.66
PFL / OSCAR1.8071.694−0.11
DF-STPO1.6281.532−0.10
RA-STPO2.3922.109−0.28

RA-STPO remains the highest, but only 0.02 above Historical/OSCAR. Stocks delisted during the test period are still missing from our data, so this removes the bias only partially. With proportional costs, the advantage of daily rebalancing disappears at 10 bps (RA-STPO 0.76 vs. 1/N 1.36; daily one-way turnover 65%). We acknowledge both as fundamental limitations of our evaluation.

Thank you again for the thoughtful review. Your questions on controls and selection criteria changed how we attribute our own results.

Thank you for the additional experiments. I have decided to maintain my original score.

Reviewer_6rQGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes RA-STPO, a decision-focused learning method for sparse portfolio optimization. The method first trains a model to predict asset returns and then learns bounded corrections to improve portfolio performance. It uses greedy asset selection, uncertainty adjustment, and adaptive blending to construct sparse portfolios. Experiments on five equity markets show improvements in average Sharpe ratio and terminal wealth over the compared methods.

Critical strengths

The paper provides useful theoretical analysis of the proposed method. In particular, it establishes an approximation guarantee and an upper bound for greedy asset selection, as well as sparse feasibility and exact local gradients for the training procedure. These results provide theoretical support for the algorithmic design.

Critical weaknesses

Sparse Sharpe ratio maximization has been extensively studied, with existing optimization algorithms and theoretical guarantees. The proposed method mainly focuses on learning bounded corrections to predicted returns and combining them with uncertainty adjustment and adaptive allocation. Although these components are useful, the overall methodological novelty appears limited. The paper does not sufficiently demonstrate the advantages over directly applying existing sparse portfolio optimizers to the corrected return predictions.

Questions for authors

1.Once the corrected return predictions are available, why not directly use existing sparse Sharpe ratio optimization algorithms at deployment? Could the authors compare these algorithms with the proposed greedy selection method using the same corrected return signals? 2.Table 10 shows that using the corrected mean alone achieves a higher Sharpe ratio in four of the five markets. What is the main benefit of the additional adaptive blending procedure? Could the authors provide more detailed ablation results to justify its necessity? 3.RA-STPO requires solving multiple quadratic programs during training and additional optimization problems during deployment. Could the authors further evaluate the computational efficiency and scalability of the proposed method, particularly for larger sparsity levels (k)? 4.Since the portfolios are rebalanced daily, could the authors evaluate performance under realistic transaction costs? Would the reported improvements remain significant after accounting for these costs?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6rQG (1/2)

We thank the reviewer for the careful review and for recognizing the theoretical analysis. The questions were precise and directly testable, so we answer each with additional experiments. Entries are five-market mean annualized Sharpe ratios (SR) with k = 10 unless stated otherwise.

Weakness and Q1 (novelty; existing sparse Sharpe optimizers applied to the corrected signals). Thank you for this central question. We applied existing sparse optimizers to exactly the same signals.

AllocatorPFL, 32 membersRA-STPO, corrected mean onlyRA-STPO, full deployment
Greedy (NNOMP), used in the paper2.0542.2602.357
OSCAR2.0771.8802.230
mSSRM-PGA1.8521.9272.223
ASMP-AFBA1.9002.0412.326
PGD + k-sparse simplex projection (Kyrillidis et al., 2013)2.0572.1302.261

With RA-STPO's full deployment, every existing optimizer (2.22–2.36) stays above the best prediction-focused combination (2.08) and DF-STPO (1.83). The improvement therefore comes mainly from the learned signal. With the corrected mean alone, greedy is higher than the other optimizers (2.26 vs. 1.88–2.13). We also agree with the reviewer's assessment that the methodological novelty is limited relative to the sparse Sharpe ratio literature. The allocator is nonnegative OMP, and the remaining components are a signal correction plus deployment heuristics.

Q2 (benefit of adaptive blending). Thank you for reading Table 10 so closely. The observation is correct: the corrected mean alone has a higher SR in four of the five markets. Below are per-market results and further ablations.

VariantMeanESTXFTSEKOSPINIKKEIS&P
Complete (paper)2.3572.8372.4022.6111.5432.392
Corrected mean only (λ = 1)2.2602.8691.6642.6701.5922.505
Anchor only (λ = 0)2.0462.6512.5831.9231.1531.921
λ fixed at the validation prior2.2352.8831.8252.5091.5822.376
λ fixed at 0.52.2012.8252.1212.4411.4432.176
Without calibration2.3382.7612.4822.7381.4562.251
Without uncertainty adjustment2.3392.8642.4452.4481.4222.514
Without both2.2742.7422.4872.6201.1642.357

Blending acts as an online safeguard rather than as an improvement in every market. In FTSE 100, the correction underperformed its anchor (1.66 vs. 2.58). There, the credibility weight fell from 0.76 to 0.005, and the complete policy recovered an SR of 2.40. In the other markets, the weight ended between 0.83 and 0.99, and the cost relative to the corrected mean was 0.03–0.11. The worst shortfall relative to the better of the two books is 0.18 for the complete policy, compared with 0.92 for the corrected mean alone.

In fairness, a HAC test finds the small losses in KOSPI (−0.06) and S&P 500 (−0.11) significant (p ≈ 0.04), because the two return paths are highly correlated (0.98). The FTSE gain (+0.74) is significant at p = 0.02. Fixed or prior-only weights are worse than the causal update. Calibration and uncertainty adjustment each contribute only about 0.02 on average, and 0.08 together.

AuthorsOpenReview ↗

Response to Reviewer 6rQG (2/2)

Q3 (computational efficiency and scalability). Thank you for this practical question. The table shows the median allocation latency per date for the S&P 500 universe (N = 465) on one NVIDIA A10 GPU, the hardware used in the paper, in milliseconds.

Allocatork = 5k = 10k = 20k = 50
Greedy (NNOMP)7.413.225.962.5
OSCAR4.54.95.126.6
mSSRM-PGA84.980.6109.885.5
ASMP-AFBA1029.11230.71241.41225.9
PGD + k-sparse simplex projection91.3127.6164.9237.1

The greedy cost grows roughly linearly in k. Training one corrector takes 19 s, 73 s, and 277 s for k = 5, 10, and 20 (medians over 16 members; 1.3, 2.8, and 11.4 s per epoch), so training grows faster than linearly in k. Resampling J = 8 faces per epoch costs about 2.4× a deterministic face (2.8 s vs. 1.2 s per epoch at k = 10). This matters because the accuracy benefit of resampling turned out to be negligible (+0.01 SR) in our controls.

Q4 (transaction costs). Thank you for raising this point; it is the most important limitation of our evaluation. We charged proportional costs on drifted holdings under the paper's daily rebalancing.

Method0 bps5 bps10 bps20 bpsDaily one-way turnoverBreak-even cost (bps)
PFL / OSCAR2.0771.3910.703−0.6730.62415.7
DF-STPO1.8260.891−0.046−1.9210.81710.4
RA-STPO2.3571.5590.759−0.8420.64615.5
1/N (dense)1.3791.3681.3571.3350.006738

Break-even cost: the proportional cost at which the mean net return becomes zero.

The improvement does not survive realistic costs under daily rebalancing: at 10 bps, no learned sparse method outperforms 1/N. When the daily signals are held for a month through overlapping tranches, RA-STPO remains the best at 10 bps, but the margin is small (1.50 vs. 1.48 for PFL/ASMP-AFBA and 1.37 for 1/N). With simple monthly rebalancing, PFL/ASMP-AFBA is better (1.56 vs. 1.49). We acknowledge this as a fundamental limitation of the current formulation.

Thank you again for the precise and constructive questions.

Reviewer_RK5ROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a neural network to predict returns and then uses the predicted returns in sparse optimization. The sparse optimization formulation is quite standard. There are some concerns regarding the formulation settings. Also, from my understanding, portfolio design is concerned with medium- or long-term asset management, not daily trading. There is no ablation study.

Critical strengths

The paper applies “regularization” to the predicted returns between two consecutive windows. However, it may not work if there are large changes between the two consecutive windows. The method can explicitly control sparsity, but the solver is a kind of NNOMP. To enforce the sum-to-one constraint, it uses normalization.

Critical weaknesses

There are insufficient ablation studies: RA-STPO + Solver A RA-STPO + Solver B RA-STPO + Solver C Does the improvement come from RA-STPO or the proposed solver?

Since the approach rebalances daily and performance is evaluated cumulatively, I do not understand why log returns are not used in either the experiments or the formulation.

The novelty of the paper does not lie in sparse optimization. Rather, it lies in the estimation of returns. The sparse solver is quite old. With some reformulation, the solver is identical to NNOMP. I do not understand why the authors present this analysis.

Some recent results on sparse solvers based on projected gradient methods or iterative hard thresholding should be mentioned.

The authors should select some comparison methods that directly handle nonnegativity constraints.

Why is there no risk-free return in the formulation?

The experimental setting has a major problem. The portfolio is rebalanced daily, and the portfolio weights may change significantly between consecutive rebalancing periods. Although the simulation results may be good, frequent portfolio adjustments could lead to high transaction costs and execution risks in practice. Also, how are the trading prices within each day selected in the simulation? If the close price is used, it is no good for daily trade.

The proposed work is a short-term return prediction and trading system. However, portfolio design typically concerns medium- or long-term asset management. Daily changes in portfolio weights may improve performance in simulations, but they may not lead to a practical investment strategy.

Questions for authors
  1. Why is K = 10 selected?
  2. Why is there no risk-free return in the formulation?
  3. Why are log returns not considered in the formulation?
  4. The Sharpe ratio is a performance indicator. Since the simulation considers daily rebalancing, why are turnover costs and other trading costs not included?
2 comments
AuthorsOpenReview ↗

Response to Reviewer RK5R (1/2)

We sincerely thank the reviewer for a careful and practically grounded review. The comments on the solver, return conventions, the risk-free rate, transaction costs, and the investment horizon identify real weaknesses of our formulation, and we ran additional experiments for each point. Unless stated otherwise, entries are five-market mean annualized Sharpe ratios (SR) with k = 10.

W1, W3, W4, W5 (solver ablation, NNOMP, projected-gradient/IHT methods, nonnegativity). Thank you for asking whether the improvement comes from the return estimates or from the solver, and for the pointers to modern sparse solvers. We crossed each signal with five long-only allocators. One of them is projected gradient with exact projection onto the k-sparse simplex (Kyrillidis et al., 2013), which enforces nonnegativity, the budget, and cardinality in a single step. mSSRM-PGA and ASMP-AFBA are natively long-only and were run with their original allocations.

Signal \ allocatorGreedy (NNOMP)OSCARmSSRM-PGAASMP-AFBAPGD + sparse-simplex projection
Historical mean1.3441.1871.3121.1991.316
PFL, 32 members2.0542.0771.8521.9002.057
RA-STPO, corrected mean only2.2601.8801.9272.0412.130
RA-STPO, full deployment2.3572.2302.2232.3262.261
  • With every allocator, RA-STPO (2.22–2.36) remains above the best prediction-focused combination (2.08, PFL/OSCAR). With prediction-focused forecasts, greedy is not the best allocator (OSCAR 2.08 vs. greedy 2.05). The gain therefore comes from the learned signal, not from the solver.
  • We agree with the reviewer that the selection rule is nonnegative OMP (Yaghoobi et al., 2015) and that its guarantee follows from weak submodularity (Das & Kempe, 2011; Elenberg et al., 2018). Presenting this analysis as a contribution was our mistake.
  • Projected-gradient and IHT-type methods (Blumensath & Davies, 2009; Beck & Eldar, 2013; Kyrillidis et al., 2013) are indeed the relevant modern comparators, and the PGD column above is our attempt to include one.

W2, Q3 (log returns). Thank you for this question. We used simple returns because a one-period portfolio return is linear in simple returns (r_p = wᵀr) but not in log returns, which is the standard convention for mean-variance and Sharpe formulations. Because cumulative performance matters, we also computed log-return metrics. The ranking is unchanged.

MethodSRSR of log returnsCAGRAnn. volatilityMax drawdown
PFL / OSCAR2.0881.95557.8%24.7%19.3%
DF-STPO1.8231.71651.3%23.6%18.7%
RA-STPO2.3692.25561.7%22.1%16.9%
1/N1.3961.29323.4%16.2%16.5%

Computed with our post-hoc evaluator; SR differs from the main table by at most 0.02.

We also recognize that a CAGR above 60% is not credible as an investment result. It reflects daily rebalancing without costs and with close-to-close execution, which the reviewer rightly questions below.

W6, Q2 (risk-free rate). Thank you for pointing this out. We recomputed SR on excess returns using daily short-term rates (€STR, SONIA, KRW 3-month, JPY call rate, US 3-month T-bill). We also subtracted the risk-free rate inside the decision layer and re-optimized.

MethodExcess SR, decision with rf = 0Excess SR, rf in the decision
Historical / OSCAR1.0511.013
Historical / mSSRM-PGA1.1781.188
PFL / OSCAR1.9341.944
PFL / greedy1.8651.890
DF-STPO1.6771.656
RA-STPO2.1942.161

The ranking is preserved, and each change is within 0.04. This is because daily risk-free rates (0.3–4.5% per year) are small relative to cross-sectional differences in daily expected returns. For reference, the excess SR of 1/N is 1.168. We agree that excess returns are the correct formulation and should have been used throughout.

AuthorsOpenReview ↗

Response to Reviewer RK5R (2/2)

W7, Q4 (transaction costs, turnover, execution price). We fully agree that this is the main limitation of the paper, and we thank the reviewer for pressing on it. We charged proportional costs on the traded amount relative to drifted holdings (DeMiguel et al., 2009), keeping the paper's daily rebalancing.

Method0 bps5 bps10 bps20 bpsDaily one-way turnoverBreak-even cost (bps)
Historical / mSSRM-PGA1.3121.1911.0690.8260.12153.0
PFL / OSCAR2.0771.3910.703−0.6730.62415.7
DF-STPO1.8260.891−0.046−1.9210.81710.4
RA-STPO2.3571.5590.759−0.8420.64615.5
1/N (dense)1.3791.3681.3571.3350.006738

Break-even cost: the proportional cost at which the mean net return becomes zero.

At 10 bps, no learned sparse method, including ours, outperforms the equal-weight portfolio. On execution prices, the paper's backtest uses close-to-close returns, which assumes execution at the closing price of the decision day. Because this is optimistic, we also delayed execution by one full trading day (no costs):

MethodExecution at decision-day closeOne-day delay
Historical / mSSRM-PGA1.3261.413
PFL / OSCAR2.0881.412
DF-STPO1.8230.859
RA-STPO2.3691.598
1/N1.3961.385

Computed with our post-hoc evaluator; values differ from the main table by at most 0.02.

RA-STPO stays the highest, but about one third of its gross SR disappears within one day, so much of the signal is short-lived. We accept the reviewer's conclusion that the simulation results do not translate into a practical daily strategy.

W8 (portfolio design is a medium- to long-term problem). Thank you for this fundamental comment, with which we agree. Because it predicts next-day returns and rebalances daily, our formulation is closer to a short-term trading system than to portfolio design. To check how much of the result survives at longer horizons, we held the daily signals for a week or a month. Holdings were either rebalanced all at once ("periodic") or replaced at 1/H of capital per day through overlapping tranches (Jegadeesh & Titman, 1993).

MethodWeekly, periodicWeekly, tranchesMonthly, periodicMonthly, tranchesMonthly, tranches, 20 bps
Historical / mSSRM-PGA1.3621.3251.3741.3261.281
PFL / OSCAR1.3051.3371.1961.3571.259
PFL / ASMP-AFBA1.3101.3691.5631.4771.375
DF-STPO0.2860.9710.7071.1461.006
RA-STPO1.3251.4591.4911.5041.389
1/N1.3601.3631.3631.3681.360

Costs are 10 bps unless stated otherwise.

With tranches, RA-STPO remains the best, but its margin over 1/N is small (0.10–0.14 at 10 bps and 0.03 at 20 bps). With periodic rebalancing, other methods are better. Because the model was trained on next-day returns, these numbers do not demonstrate a good monthly strategy either. As the reviewer suggests, a medium-horizon problem with costs would have to be formulated and trained as such from the start.

Q1 (choice of K = 10). Thank you for the question. K = 10 follows common choices in the sparse portfolio literature and was fixed for comparability across markets. We retrained the learned methods for each cardinality.

Methodk = 5k = 10k = 20
Historical mean, best allocator1.4231.3441.390
PFL, best allocator1.9402.0772.091
DF-STPO, retrained per k1.4391.8261.878
RA-STPO, retrained per k2.2852.3572.294

RA-STPO has the highest mean SR at each k, so k = 10 was not a choice that favors our method.

Thank you again for the time and expertise you devoted to our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all three reviewers for their careful and constructive reviews. While the reviewers recognized some merits of the work, such as the explicit control of sparsity (RK5R), training directly on sparse, long-only allocations (Geha), and the supporting theoretical analysis (6rQG), the reviews center on concerns about the problem formulation, the novelty, and the evaluation. Our responses focus on these concerns, and we ran additional experiments for each of them. The main findings are summarized below, and full tables are in the individual responses. Numbers are five-market mean annualized Sharpe ratios (SR).

  • Transaction costs and daily rebalancing (RK5R, 6rQG, Geha). We agree that this is the main limitation of the paper. With proportional costs, RA-STPO's SR falls from 2.36 to 0.76 at 10 bps, where no learned sparse method outperforms 1/N (1.36). At the same cost, holding the daily signals for a week or a month leaves RA-STPO at most 0.14 above 1/N. A medium-horizon problem with costs would have to be formulated and trained as such from the start.
  • Solver and novelty (RK5R, 6rQG). Across five long-only allocators, RA-STPO stays at 2.22–2.36, above the best prediction-focused combination (2.08), so the gain comes mainly from the learned signal rather than from the solver. We agree that the selection rule is nonnegative OMP and that the methodological novelty relative to the sparse Sharpe ratio literature is limited.
  • Fairness of the comparison (Geha). The comparison does not change when the checkpoint criteria are crossed (RA-STPO reaches 2.35 when selected by rank IC) or when DF-STPO's α and β are tuned on validation (1.857, against 2.36 for RA-STPO). However, with a one-year validation window, validation SR does not predict test SR (rank correlation −0.03), which is a genuine weakness of our evaluation protocol.
  • Source of the gain (Geha, 6rQG). In matched controls, most of the gain comes from the bounded correction around a frozen anchor, which adds 0.24–0.27 over decision-focused fine-tuning of the predictor itself. Resampled sparse paths add essentially nothing over a deterministic anchor face (2.357 vs. 2.350), so the paper's attribution of the gain to resampling is not supported. Adaptive blending acts as a safeguard rather than as an improvement in every market.

The remaining points (log returns, the risk-free rate, the choice of k, computational cost, and survivorship) are addressed in the individual responses. We thank the reviewers again for their time and expertise.

Reviewer_6rQGOpenReview ↗

Comments

Thank you for your responses and additional experiments. The new experimental results address some of my concerns. However, I still have two main concerns: 
1. As the authors state in the rebuttal, the improvement comes mainly from the learned signal, and the methodological novelty is limited relative to the sparse Sharpe ratio literature. Therefore, I think the overall novelty is limited.
2. The transaction cost experiments show that the method’s practical effectiveness remains questionable. 
I maintain my overall assessment.

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

Related papers

Discussion (0)

Sign in to comment.