Learning Predictive Ambiguity Sets for Decision-Focused Distributionally Robust Optimization

Junjie Guo

REVIEW 3 major objections 5 minor 25 references

Learning Predictive Ambiguity Sets for Decision-Focused Distributionally Robust Optimization

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A deep model can learn how large a robustness ball should be for each market state, recovering most of the gains of strong fixed-radius DRO while using a smaller average radius.

desk verdict Clean, usable recipe for adaptive Wasserstein radii in decision-focused DRO; the math is standard and the portfolio gains look real, but the superiority claim still sits on one chronological path of one 20-name basket. read the letter →

arxiv 2607.09820 v1 pith:AKVF544V submitted 2026-07-10 cs.LG q-fin.CP

Junjie Guo This is my paper · ORCID

classification cs.LGq-fin.CP

keywords distributionallyrobustoptimizationWassersteinambiguitysetsdecision-focusedlearningportfoliopredictiveuncertaintyadaptiverobustnessscenariogeneration

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Predict-then-optimize systems treat a point forecast as reliable and can amplify small prediction errors into large decision mistakes. Classical distributionally robust optimization protects against that by optimizing against a whole ball of distributions, but the ball is usually centered on historical samples with one fixed radius, so it is often too conservative in calm regimes and still brittle under regime shift. This paper proposes learned predictive ambiguity sets: from context, a neural model outputs a finite nominal scenario distribution and a state-dependent Wasserstein radius that together define the ambiguity set for a robust decision layer. The radius is trained by conditional quantile calibration, size regularization, and realized decision loss so that robustness expands when forecasts are unreliable and contracts when they are trustworthy. On long-only portfolio optimization with 20 S&P 500 stocks from 2018–2026, the method substantially beats equal-weight, pure predict-then-optimize, and historical Wasserstein DRO, and nearly matches a strong deep fixed-radius baseline (26.28% annualized return, Sharpe 1.30, final wealth 1.61) while using a smaller average radius and adapting better across volatility and drawdown regimes.

What carries the argument

Learned predictive ambiguity sets (LPAS): a contextual Wasserstein ball whose center is a neural finite nominal scenario distribution and whose radius is a state-dependent network; the dual of that ball supplies a tractable robust decision layer that is trained jointly with the radius.

What would settle it

Re-run the same portfolio experiment over multiple rolling windows, random seeds, and a larger asset universe and find that the learned-radius model no longer matches fixed-radius DRO on return and Sharpe while keeping a smaller radius, or that the learned radius fails to rise with market volatility.

Watch

Extended reading notes

Core claim

Learned predictive ambiguity sets—a contextual finite scenario distribution plus a state-dependent Wasserstein radius trained by quantile calibration, size regularization, and downstream decision loss—can make distributionally robust optimization adaptive rather than globally fixed. On the reported 20-asset portfolio task they recover most of the out-of-sample performance of a deep fixed-radius DRO baseline while using a smaller average radius, slightly better tail metrics, and stronger regime adaptivity.

Load-bearing premise

The out-of-sample superiority and regime-adaptivity claims rest on a single chronological train/validation/test split of one 20-stock S&P 500 universe, without multi-seed or multi-fold checks.

Editorial extensions

If this is right

Robust optimizers can shrink the ambiguity radius in calm regimes without giving up protection when forecasts are unreliable.
How large the robustness ball should be can be driven by decision quality, not only by predictive coverage.
Historical fixed-radius Wasserstein DRO underuses context and can be outperformed by predictive centers plus adaptive radii.
Most of the reported gains can be obtained by staged training: pretrain scenarios, calibrate the radius, then decision-focused fine-tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same radius-learning pattern should transfer to inventory, routing, and energy dispatch, where context likewise modulates forecast reliability.
When decision-aware tuning leaves empirical coverage below the nominal quantile, conformal post-calibration of the learned radius can restore strict risk control without fully undoing performance.
The optional anisotropic ground metric (proposed but not tested) could further cut conservatism by stretching the ball only in decision-sensitive directions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The useful takeaway is simple: treat the ambiguity radius as a contextual output of the same model that produces the nominal scenarios, train it with pinball calibration plus size and decision loss, and feed the resulting Wasserstein ball into a dual DRO layer. That package is what they call LPAS.

What is actually new is the joint object—finite predictive scenarios plus a state-dependent radius—and the staged training that makes it trainable. The dual (Eq. 14–15) and the portfolio specialization (Eq. 19–20) are textbook under affine loss and Euclidean cost; they are written cleanly and used correctly. The ablations are honest: kill calibration and the radius collapses and performance dies; size and decision fine-tuning matter less at the chosen scale. Against equal-weight, P2O, and historical WDRO the gains are large. Against a deep fixed-radius DRO that shares the same Transformer scenarios, LPAS keeps most of the return and Sharpe with a clearly smaller average radius and slightly better worst-month/CVaR numbers, and the radius tracks volatility the way you would want.

The soft spot is exactly the one the stress note flags, and the paper already owns it in §7: one chronological 1129/410/515 split on one 20-name S&P universe, no seeds, no rolling folds, no second universe. Decision-aware validation on that same path can favor the adaptive story. Coverage also undershoots the nominal τ after decision tuning. Those are real limits on how hard you can lean on the 26% / 1.30 / radius-24-vs-35 numbers; they are not reasons to dismiss the architecture.

Citations are appropriate (Wang, Sun, Chenreddy–Delage, classic Wasserstein DRO). No circularity in the objective. Free parameters are the usual portfolio/ML set and are selected on validation.

This is for people who already run predict-then-optimize or Wasserstein DRO pipelines and want a practical adaptive-radius layer, especially in finance-style sequential decisions. It is not a theory paper. I would send it to referees: the method is coherent, the dual is checkable, the empirics are directionally informative, and the limitations are stated. Multi-fold/seed results and code would turn a conditional accept into a much stronger one. Worth engaging if you work in this stack; not urgent if you do not.

Referee Report

3 major / 5 minor

Summary. The paper proposes learned predictive ambiguity sets (LPAS): a contextual deep model that outputs a finite nominal scenario distribution and a state-dependent Wasserstein radius (optionally an anisotropic ground metric), which define a contextual ambiguity set for a DRO decision layer. The radius is trained by combining pinball quantile calibration, size regularization, and downstream decision loss (Eqs. 24–28; Algorithm 1). The finite dual of the Wasserstein DRO layer is derived (Eqs. 14–15) and specialized to a long-only portfolio problem with Euclidean cost, yielding a closed-form robust objective (Eqs. 19–20). On a single chronological split of 20 S&P 500 names (2018–2026), LPAS-W reports 26.28% annualized return, Sharpe 1.30, final wealth 1.61, and a smaller average radius than a deep fixed-radius DRO baseline while remaining competitive on tail metrics and improving some regime slices (Tables 2–4).

Significance. If the empirical claims hold under stronger validation, the work would be a useful bridge between decision-focused learning and Wasserstein DRO: instead of a hand-tuned fixed radius centered on historical samples, both the nominal distribution and the radius become contextual and trainable. The dual specialization for the portfolio layer is standard and correctly applied, and the staged training objective (prediction + calibration + size + decision loss) is a concrete, implementable recipe. The main contribution is therefore architectural and empirical rather than theoretical; its value hinges on whether adaptive radii reliably reduce unnecessary conservatism while preserving robustness. The paper is transparent about the single-split limitation (§7), which is appropriate.

major comments (3)

§6.1–6.3 and Tables 2–4: The central superiority and regime-adaptivity claims rest on a single chronological 1129/410/515 split of one 20-name S&P 500 universe, with no multi-seed, multi-fold, or multi-universe results. Because the Transformer scenario generator, radius network, and decision-aware validation score are all selected on this path, the reported edge (26.28% return / Sharpe 1.30 / radius 24.3 vs Fixed-DRO’s 35.4) and the high-ρ / drawdown regime gains in Table 4 could be path-specific. §7 already flags this; for the claims as stated, at least one additional rolling fold or multi-seed summary is load-bearing.
§4.5 and §6.5 / Fig. 4–5: After decision-aware tuning the empirical coverage of LPAS-W is 0.755 versus the nominal τ=0.9 target. The paper notes that size regularization and decision loss trade exact coverage for performance, and suggests conformal post-calibration if strict coverage is required. That is fine as a design choice, but the abstract and introduction still present the radius as “calibrated”; the manuscript should either report a post-calibrated variant or qualify the calibration claim so that readers do not over-read statistical coverage guarantees.
§3–4 and experiments: The optional anisotropic ground metric c_ψ (Eq. 7) is part of the stated framework and contributions but is never evaluated; all results use fixed Euclidean cost. Either evaluate a learned-metric ablation or narrow the contribution statement so that the unevaluated component is not presented as part of the empirical method.

minor comments (5)

Table 1: Candidate numbering and the “Selected” column are hard to parse; a clearer description of the decision-aware validation score (weights on loss, tail, calibration, size) would help reproducibility.
Figures 1–2 legends truncate method names (“Deep Predictive Fixed-radius D”); expand for readability.
Eq. (26): The scenario prediction loss is written as a min-i reconstruction; clarify whether this is the sole L_pred used or whether NLL/energy-score variants were also tried.
§6.1: List of 20 tickers and feature construction details are useful; a short note on data source and any survivorship handling would strengthen the experimental section.
Notation: ρ_t, ρ_ϕ(z_t), and ρ_min appear in several places; a single consistent symbol table would reduce minor ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: dual forms are standard Wasserstein theory, training is staged empirical fitting, and reported portfolio metrics are chronological out-of-sample realizations, not forced by construction.

full rationale

The paper's load-bearing mathematical step is the finite Wasserstein dual (Eqs. 14–15) and the portfolio closed form (Eqs. 19–20), both standard results from the cited external DRO literature (Mohajerin Esfahani & Kuhn, Blanchet & Murthy, Gao & Kleywegt) rather than self-defined identities. The radius is not defined as the quantity later reported as performance: it is trained by pinball calibration on normalized forecast errors (Eqs. 21–23), size/stability regularizers, and a separate decision loss (Eq. 24), then evaluated by realized out-of-sample portfolio return, Sharpe, wealth, and CVaR on a chronological held-out window (Table 2). Hyperparameter selection on a decision-aware validation score is ordinary model selection, not a fitted input renamed as a prediction of the same quantity. There are no self-citations, no uniqueness theorems imported from the author, and no ansatz smuggled from prior own work. Ablations (Table 3) further show that removing calibration collapses coverage and performance, which would be impossible if the main metrics were tautological. Empirical limitations (single split, one universe) affect external validity, not circularity of the derivation chain.

Assumptions & free parameters 7 free parameters · 4 assumptions · 2 invented entities

The central empirical claim rests on standard Wasserstein dual theory, a portfolio modeling choice that robustifies only the linear return term under Euclidean cost, a large set of architecture and regularization hyper-parameters selected on a decision-aware validation score, and the untested assumption that one chronological split of one 20-name equity universe is representative. No new physical entities are postulated; the invented object is the LPAS construction itself.

free parameters (7)

quantile level τ = 0.9
Controls the pinball calibration target for the learned radius; set to 0.9 and used both for training and for interpreting coverage.
risk penalty λ_risk = 8.0
Quadratic risk term weight in the portfolio objective; selected by validation decision-aware score.
transaction-cost objective weight λ_tc = 0.0015
L1 turnover penalty inside the training objective; selected on validation.
scenario count N = 7
Number of discrete scenarios output by the Transformer nominal model; fixed at 7.
radius floor ρ_min = 1e-4
Hard lower bound on the softplus radius network; set to 1e-4.
Transformer architecture (layers/heads/dim/lookback) = 2 layers, 4 heads, dim 48, lookback 63
Model capacity and context length chosen among candidates by validation score (Table 1).
ridge coefficient = 1.0
Regularization strength inside the selected candidate; part of hyper-parameter search.

assumptions (4)

standard math The Wasserstein dual formula (Eq. 14–15) holds for the chosen loss and ground cost, reducing the portfolio problem to −μ̂ᵀw + ρ‖w‖₂ plus deterministic regularizers (Eq. 19–20).
Invoked throughout §4.3–4.4; standard under affine loss and norm cost, but the paper only robustifies the linear return term.
domain assumption A single chronological split of daily returns for the listed 20 S&P 500 names is representative of out-of-sample regime-adaptive performance.
Underpins all of §6; explicitly listed as a limitation in §7.
ad hoc to paper Euclidean ground metric with fixed (non-learned) cost is sufficient; the optional anisotropic metric c_ψ is left unevaluated.
Stated in §3 and §7; experiments use only fixed Euclidean transport.
domain assumption Pinball loss at level τ plus size and stability regularizers produce a radius that is both statistically meaningful and decision-relevant after fine-tuning.
Core of the radius-learning objective (Eq. 24–28) and Algorithm 1.

invented entities (2)

Learned Predictive Ambiguity Set (LPAS)
purpose: Contextual finite nominal distribution plus state-dependent Wasserstein radius (and optional metric) that feeds a DRO decision layer.
Defined in §1 and §4; the paper’s central modeling object. No independent external measurement of ‘LPAS’ exists outside this construction.
Context-dependent radius network ρ_ϕ(z)
purpose: Maps market state to a non-negative ambiguity radius via softplus or sigmoid parameterization.
Eq. 5–6; trained by the composite loss. Its ‘correct’ value is defined only relative to the paper’s calibration and decision objectives.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning Predictive Ambiguity Sets for Decision-Focused Distributionally Robust Optimization." pith.science (2026). https://pith.science/paper/AKVF544V

@misc{pith2026260709820,
  author       = {Pith},
  title        = {Pith review of: Learning Predictive Ambiguity Sets for Decision-Focused Distributionally Robust Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKVF544V}},
  note         = {Machine review of arXiv:2607.09820}
}

read the original abstract

Predict-then-optimize systems usually compress uncertainty into a point forecast and then solve a downstream optimization problem as if the forecast were reliable. Distributionally robust optimization (DRO) offers protection against misspecification, but the ambiguity set is often centered at historical samples and uses a fixed radius. We propose \emph{learned predictive ambiguity sets} (LPAS): a deep contextual model outputs a finite nominal scenario distribution, a state-dependent Wasserstein radius, and optionally an anisotropic ground metric. These outputs define a contextual ambiguity set that feeds a DRO decision layer. The radius is trained by a combination of conditional quantile calibration, size regularization, and downstream decision loss, so that robustness is adaptive rather than globally fixed. We derive the finite dual form used by the decision layer, present a staged training algorithm, and evaluate the method on distributionally robust portfolio optimization with 20 S&P 500 constituents from 2018--2026. The proposed method substantially improves over equal-weight, predict-then-optimize, and historical Wasserstein DRO baselines, achieving 26.28% annualized return, Sharpe ratio 1.30, final wealth 1.61, and lower tail loss than a deep fixed-radius DRO baseline while using a smaller average radius. The results show that learned ambiguity radii can recover most of the performance of strong fixed-radius DRO while reducing unnecessary conservatism and improving regime adaptivity.

Figures

Figures reproduced from arXiv: 2607.09820 by the authors.

**Figure 2.** Out-of-sample drawdown. LPAS-W remains close [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 4.** Empirical coverage by market-volatility quartile. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (3 more)

**Figure 5.** Figure 5: Coverage versus radius conservatism. LPAS-W at [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Average turnover by method. The comparison [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Decision regret versus prediction error. Prediction [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 1 linked inside Pith

[1]

Agrawal, A.; Amos, B.; Barratt, S.; Boyd, S.; Diamond, S.; and Kolter, J. Z. 2019. Differentiable Convex Optimization Layers. In Advances in Neural Information Processing Systems

2019
[2]

Amos, B.; and Kolter, J. Z. 2017. OptNet: Differentiable Optimization as a Layer in Neural Networks. In Proceedings of the International Conference on Machine Learning, 136--145

2017
[3]

Ben-Tal, A.; El Ghaoui, L.; and Nemirovski, A. 2009. Robust Optimization. Princeton University Press

2009
[4]

B.; and Caramanis, C

Bertsimas, D.; Brown, D. B.; and Caramanis, C. 2011. Theory and Applications of Robust Optimization. SIAM Review, 53(3): 464--501

2011
[5]

Blanchet, J.; and Murthy, K. 2019. Quantifying Distributional Model Risk via Optimal Transport. Mathematics of Operations Research, 44(2): 565--600

2019
[6]

Blanchet, J.; Chen, L.; and Zhou, X. Y. 2022. Distributionally Robust Mean-Variance Portfolio Selection with Wasserstein Distances. Management Science, 68(9): 6382--6410

2022
[7]

R.; and Delage, E

Chenreddy, A. R.; and Delage, E. 2024. End-to-End Conditional Robust Optimization. In Proceedings of the Conference on Uncertainty in Artificial Intelligence

2024
[8]

Delage, E.; and Ye, Y. 2010. Distributionally Robust Optimization Under Moment Uncertainty with Application to Data-Driven Problems. Operations Research, 58(3): 595--612

2010

Show all 25 references

[9]

C.; and Namkoong, H

Duchi, J. C.; and Namkoong, H. 2021. Learning Models with Uniform Performance via Distributionally Robust Optimization. The Annals of Statistics, 49(3): 1378--1406

2021
[10]

N.; and Grigas, P

Elmachtoub, A. N.; and Grigas, P. 2022. Smart ``Predict, then Optimize''. Management Science, 68(1): 9--26

2022
[11]

Gao, R.; and Kleywegt, A. J. 2023. Distributionally Robust Stochastic Optimization with Wasserstein Distance. Mathematics of Operations Research, 48(2): 603--655

2023
[12]

Gao, R.; Chen, X.; and Kleywegt, A. J. 2022. Wasserstein Distributionally Robust Optimization and Variation Regularization. Operations Research, 70(5): 2939--2961

2022
[13]

Goh, J.; and Sim, M. 2010. Distributionally Robust Optimization and Its Tractable Approximations. Operations Research, 58(4): 902--917

2010
[14]

Koenker, R.; and Bassett, G. 1978. Regression Quantiles. Econometrica, 46(1): 33--50

1978
[15]

A.; and Shafieezadeh-Abadeh, S

Kuhn, D.; Mohajerin Esfahani, P.; Nguyen, V. A.; and Shafieezadeh-Abadeh, S. 2019. Wasserstein Distributionally Robust Optimization: Theory and Applications in Machine Learning. arXiv preprint arXiv:1908.08729

2019 arXiv
[16]

Markowitz, H. 1952. Portfolio Selection. The Journal of Finance, 7(1): 77--91

1952
[17]

Mohajerin Esfahani, P.; and Kuhn, D. 2018. Data-Driven Distributionally Robust Optimization Using the Wasserstein Metric: Performance Guarantees and Tractable Reformulations. Mathematical Programming, 171: 115--166

2018
[18]

T.; and Uryasev, S

Rockafellar, R. T.; and Uryasev, S. 2000. Optimization of Conditional Value-at-Risk. Journal of Risk, 2(3): 21--41

2000
[19]

Romano, Y.; Patterson, E.; and Candes, E. J. 2019. Conformalized Quantile Regression. In Advances in Neural Information Processing Systems

2019
[20]

Shafieezadeh-Abadeh, S.; Mohajerin Esfahani, P.; and Kuhn, D. 2015. Distributionally Robust Logistic Regression. In Advances in Neural Information Processing Systems

2015
[21]

Sinha, A.; Namkoong, H.; and Duchi, J. 2018. Certifying Some Distributional Robustness with Principled Adversarial Training. In International Conference on Learning Representations

2018
[22]

Sun, C.; Liu, L.; and Li, X. 2023. Predict-then-Calibrate: A New Perspective of Robust Contextual LP. In Advances in Neural Information Processing Systems

2023
[23]

N.; Kaiser, L.; and Polosukhin, I

Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention Is All You Need. In Advances in Neural Information Processing Systems

2017
[24]

Wang, I.; Becker, C.; Van Parys, B.; and Stellato, B. 2023. Learning Decision-Focused Uncertainty Sets in Robust Optimization. arXiv preprint arXiv:2305.19225

2023
[25]

Wilder, B.; Dilkina, B.; and Tambe, M. 2019. Melding the Data-Decisions Pipeline: Decision-Focused Learning for Combinatorial Optimization. In Proceedings of the AAAI Conference on Artificial Intelligence

2019

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.

[1] [1]

Agrawal, A.; Amos, B.; Barratt, S.; Boyd, S.; Diamond, S.; and Kolter, J. Z. 2019. Differentiable Convex Optimization Layers. In Advances in Neural Information Processing Systems

2019

[2] [2]

Amos, B.; and Kolter, J. Z. 2017. OptNet: Differentiable Optimization as a Layer in Neural Networks. In Proceedings of the International Conference on Machine Learning, 136--145

2017

[3] [3]

Ben-Tal, A.; El Ghaoui, L.; and Nemirovski, A. 2009. Robust Optimization. Princeton University Press

2009

[4] [4]

B.; and Caramanis, C

Bertsimas, D.; Brown, D. B.; and Caramanis, C. 2011. Theory and Applications of Robust Optimization. SIAM Review, 53(3): 464--501

2011

[5] [5]

Blanchet, J.; and Murthy, K. 2019. Quantifying Distributional Model Risk via Optimal Transport. Mathematics of Operations Research, 44(2): 565--600

2019

[6] [6]

Blanchet, J.; Chen, L.; and Zhou, X. Y. 2022. Distributionally Robust Mean-Variance Portfolio Selection with Wasserstein Distances. Management Science, 68(9): 6382--6410

2022

[7] [7]

R.; and Delage, E

Chenreddy, A. R.; and Delage, E. 2024. End-to-End Conditional Robust Optimization. In Proceedings of the Conference on Uncertainty in Artificial Intelligence

2024

[8] [8]

Delage, E.; and Ye, Y. 2010. Distributionally Robust Optimization Under Moment Uncertainty with Application to Data-Driven Problems. Operations Research, 58(3): 595--612

2010