acceptodds
Under review as a conference paper at ICLR 2027

Adversarial Training for Deep Hedging in Nonstationary Markets

Abstract

Deep hedging learns trading policies from historical or simulated market trajectories, yet under nonstationarity these training paths may not represent future market conditions. We propose WRAP (Wasserstein–Reweighting Adversarial Perturbation), a drift-aware adversarial training framework derived from a two-budget distributionally robust optimization (DRO) formulation. The formulation is anchored to a weighted empirical reference distribution whose fixed baseline weights are chosen to balance sampling uncertainty against temporal drift. Around this reference distribution, the ambiguity set addresses two complementary forms of distributional misspecification by allowing an adversary to reweight the observed trajectories subject to a -divergence constraint and perturb their paths subject to an optimal-transport (OT) constraint. We derive a joint first-order expansion in which the leading-order increase over the nominal expected loss decomposes into a reweighting contribution determined by the dispersion of hedging losses across trajectories and a transport contribution determined by the sensitivity of the loss to path perturbations. This expansion yields an explicit finite-dimensional adversarial attack that replaces the distributional inner supremum with a tractable first-order approximation. Across stationary and nonstationary Heston dynamics and a generalized affine diffusion (GAD), the experiments show complementary benefits from reweighting and transport, with joint adversarial training providing the largest gains under nonstationarity.

Primary area:
optimization
Keywords:
deep hedging, distributionally robust optimization, nonstationary markets, adversarial training, optimal transport

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_39xkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WRAP, a framework for deep hedging in nonstationary markets that combines drift-aware empirical weighting, ϕ\phi-divergence scenario reweighting, and Wasserstein path perturbations. The authors derive next-period coverage guarantees and a neat joint small-budget expansion that yields an explicit first-order adversarial attack. The method is then evaluated on both Heston simulations and market-calibrated GAD models.

Critical strengths

Overall, the paper presents a clean conceptual framework backed by solid empirical work. The evaluation is quite thorough for this domain—spanning stationary and drifting Heston dynamics, transaction costs, alternative risk metrics, and ablations across five calibrated assets—delivering consistent and convincing gains under controlled distribution shifts.

Critical weaknesses
  1. My main reservation is that the statistical coverage theorem doesn't quite seem to cover the settings actually used in the experiments. Specifically, the theorem relies on a finite-diameter path space and bounded consecutive W∞W_\infty drift. However, both Heston and GAD paths have unbounded support, and shifting volatility distributions generally won't satisfy finite W∞W_\infty drift unless there is some explicit truncation. Unless I missed it, I couldn't find any path-clipping procedure in the experimental details. Also, the theoretical constants and drift radius needed for calibration aren't actually estimated from data—instead, the weights and budgets are just tuned via validation. This somewhat undercuts the central claim that WRAP provides a principled finite-sample foundation for nonstationary hedging.
  2. Along similar lines, there is a bit of a leap from the theorem to the algorithm actually run in practice. The theory is local and first-order (assuming vanishing budgets, full-sample linear tilting, and the exact DRO objective). In contrast, the experiments use fairly noticeable budgets, exponential tilting, minibatch-normalized weights, and a 50/50 blend of clean and adversarial losses. Right now, there's no finite-budget approximation bound or optimization guarantee bridging this practical recipe back to the robust value. Finally, a minor point on the empirical framing: the "market-data" experiments are still evaluating simulated paths from fitted GAD models rather than true out-of-sample realized market trajectories.
Questions for authors
  1. Did you explicitly clip or truncate the Heston and GAD paths in your implementation? If so, what bounds were used and how much probability mass was affected? If not, how should we think about Theorem 2.3 applying here, given that the finite-diameter and W∞W_\infty assumptions seem to be violated?
  2. Practically speaking, how would someone estimate ρ\rho and the concentration constants required by your weighting rule? In your experiments, are any parts of the weights and ambiguity radii actually calibrated from theory, or are they treated purely as validation hyperparameters?
  3. Could you share some evidence that the exponential, minibatch-based attack remains a faithful approximation to the stated DRO problem at the larger budgets used in your experiments?
  4. In the GAD experiments, WRAP searches over a 2D ambiguity grid, whereas the single-budget baselines only search over 1D. Do the gains hold up if the baselines are given an equivalent hyperparameter search budget (e.g., finer grids or tuning an extra regularization parameter)?
1 comment
AuthorsOpenReview ↗

Relaxed assumptions and practical choices in adversarial training

Thank you for these questions. This allowed us to improve the results significantly!

Q1. The original Heston experiments do not clip paths, but GAD has a PnL floor. We checked the impact of clipping in the nonstationary Heston Experiment (Table 1) with price/variance clipping boxes [50,200]/[0,1], [25,400]/[0,2] and [10,1000]/[0,4]. All leave the datasets unchanged. Retraining WRAP-U at (ε,τ)=(0.3,0.6)(\varepsilon,\tau)=(0.3,0.6) with clipping also leaves the result unchanged.

Upon your question, we managed to generalize the W∞W_\infty drift bound to a finite WrW_r drift bound, r≥pr\ge p. The finite diameter assumption can be relaxed by using finite moments. However, this comes at the cost of a weaker concentration rate. The new Proposition C.6 (Appendix C.2) gives coverage without finite diameter, assuming the revised drift bound and a finite target moment of order s>ps>p. Weighted Wasserstein expectation bounds (Fournier and Guillin, 2015) and Markov's inequality yield a radius scaling as β−1/p\beta^{-1/p}. A revised version with changes in red has been uploaded. We thank you for improving the paper significantly.

Q2. Computing the empirical W∞W_\infty distance between consecutive blocks reduces to a matching problem that minimizes the largest matched distance; see Bansil and Kitagawa (2021). For finite r≥pr\ge p, empirical WrW_r between equally sized, uniformly weighted blocks is computed by minimizing the average rr-th power of the matched path distances.

We select the reference-weighting parameter and adversarial budgets using cross validation. As in Keehan et al. (2025), the optimal weights depend only on ϱ/εϱ\varrho/\varepsilon_{\varrho}: scaling both quantities equally leaves the weights unchanged. Tuning this ratio avoids redundant calibration. This follows established DRO practice. Mohajerin Esfahani and Kuhn (2018) recommend cross-validation or bootstrapping because concentration-based radii can be overly conservative; see also Gao (2023) and Blanchet, Murthy, and Nguyen (2021).

Q3. We understand this question as asking how the practical attack relates to the local construction, particularly the replacement of linear by exponential tilting. For fixed path losses, exponential tilting solves the KL-constrained reweighting subproblem exactly. It agrees with linear tilting to leading order at small budgets. We compare it with the direct small-budget linear formula on 4,000 validation paths per seed, holding the hedger and attack threshold fixed. This formula approximates KL quadratically; it does not apply the manuscript's largest-feasible-step adjustment. The table reports attained KL divergence:

KL budget τ\tauLinear approximationExponential tilting
0.10.0850.100
0.30.2430.300
0.60.4850.600
1.00.840 (3 valid seeds)1.000

Means cover five seeds except the marked entry. At τ=1\tau=1, the linear formula produces negative weights in two seeds, making their KL undefined. Exponential tilting uses the full budget with valid weights throughout. The KL budget is an upper bound, so using less does not violate the constraint.

Our response to Reviewer BR6s, Q1 also compares one-step and multistep training, which in theory should approximate the stated DRO Problem better. The difference in the objective however remains very small while the hedging performance remains essentially constant or worsens: One step captures 96.30%/96.63% of the best observed attack gain over 1, 5, 20 and 50 steps for WRAP-U/W and 20-step training has higher CVaR than one-step training.

Q4. We agree that WRAP's two-dimensional grid evaluates more combinations than the single-budget baselines. In our response to Reviewer BR6s, Q3, we restrict GAD selection to six WRAP budget pairs and six OT settings, including clean training. WRAP has lower mean historical test risk than ERM in all ten settings and OT in five.

Furthermore, nonstationary Heston, Figure 4(a–b) shows improvements across a broad region with both budgets positive. The best joint configurations reduce validation CVaR by 12.6% and 10.5% under uniform and drift-aware weights, versus 9.1% and 6.8% along the single-budget axes. This supports complementarity on the tested grid, while on the axis there is little room for improvement.

Lastly, WRAP also contains the single-budget methods and clean training. Figure 3(a) and Table 3 show validation selecting τ=0\tau=0, recovering OT; Table 5 shows the same under M–CVaR. The grid identifies when one mechanism suffices and when their combination helps.

Reviewer_8X4zOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WRAP (Wasserstein–Reweighting Adversarial Perturbation), a drift-aware adversarial training framework for deep hedging in nonstationary markets. The method constructs a weighted empirical reference distribution to balance sampling uncertainty against temporal drift, and combines probability reweighting under a ϕ-divergence constraint with trajectory perturbations under an optimal-transport constraint. A joint first-order expansion yields a tractable adversarial attack. Experiments under stationary and nonstationary Heston dynamics and generalized affine diffusion (GAD) show complementary benefits from reweighting and transport, with joint adversarial training providing the largest gains under nonstationarity.

Critical strengths
  1. The framework addresses three channels of distributional misspecification through non-uniform baseline weights, adversarial probability reweighting, and optimal-transport perturbations.

  2. The theoretical analysis provides finite-sample coverage guarantees and a joint first-order expansion that separates loss dispersion across trajectories from sensitivity to path perturbations.

Critical weaknesses
  1. The GAD experiments use entropic OCE with PnL floored at −1 during training and evaluation. Since entropic risk penalizes large losses exponentially, this floor may suppress extreme losses. It remains unclear whether the reported improvements hold without the PnL floor.

  2. The optimal-transport perturbations move trajectories along locally adverse directions under a time- and component-weighted ground metric. However, the paper does not sufficiently investigate whether the resulting adversarial trajectories remain financially admissible and consistent with realistic market dynamics.

  3. The market-data experiments use trajectories simulated from GAD models fitted to historical equity prices rather than directly realized market trajectories. This limits the evidence that WRAP improves hedging performance under actual market conditions and model misspecification.

Questions for authors

Please address the main concerns in the weakness.

2 comments
AuthorsOpenReview ↗

PnL-floor sensitivity, adversarial path diagnostics, and historical evaluation

We thank the reviewer for the constructive assessment of our work. We address the stated weaknesses below. Tables report means over five seeds.

W1. PnL flooring. You are right that flooring the PnL can affect the measured tail risk. We introduced the floor to stabilise training under the entropic loss and limit extreme gradients. The sensitivity to large losses also depends on the risk-aversion parameter λ\lambda, which we fix in our experiments. For comparability with He, et al (2025), we follow the −1 floor in their GAD implementation.

Keeping the hedgers trained with the −1 floor fixed, we reevaluated them with floors −1, −2, −5 and no floor. On the simulated test paths, no selected policy has PnL below −1, so all four evaluations give identical entropic risk. On historical paths, only AAPL changes for the ERM, OT and WRAP comparisons. Its results are:

TrainingMethodFloor −1Floor −2Floor −5No floor
FixERM0.5960.6240.6240.624
FixWRAP0.5120.5700.5700.570
RobustERM0.6320.6420.6420.642
RobustWRAP0.3730.3730.3730.373

In the GAD tables, WRAP denotes the validation-selected model with both budgets positive; OT can also select clean training. Removing the evaluation floor preserves WRAP's lower mean risk than ERM in 9/10 historical settings and OT in 4/10. Thus, these comparisons show that suppressing losses below −1 do not change the evaluation.

W2. Adversarial path plausibility. We agree that the transport constraint alone does not guarantee financially admissible or realistic paths. We therefore compare clean and adversarial Heston trajectories at (ϵ,τ)=(0.3,0.6)(\epsilon,\tau)=(0.3,0.6) for WRAP-U and (0.4,1.0)(0.4,1.0) for WRAP-W. Each statistic averages 4,000 paths per seed, counting paths equally to isolate the effect of the path perturbations.

Path statisticOriginalAttacked: WRAP-UAttacked: WRAP-W
Annualized realized volatility (%)34.7637.2437.93
Maximum drawdown (%)9.619.809.84
Correlation of consecutive returns-0.036-0.086-0.097
Correlation of price and variance changes-0.693-0.687-0.685

The attacks increase volatility and drawdown while preserving the negative relationship between price and variance changes. All ten runs satisfy the checked transport and reweighting budgets and preserve initial prices and variances, with no price or variance floor hits. These diagnostics describe the stress imposed by the attack; they do not establish consistency with a diffusion model or absence of arbitrage.

W3. Evaluation on historical paths. You are right that our original GAD test trajectories are simulated. The benchmark is calibrated separately to five equities over the training, validation and test periods. A small historical record can yield tail-risk estimates dominated by a few extreme losses. Calibrated simulation permits evaluation on 20,000 paths per asset, improving precision conditional on the fitted models. Historical evaluation tests transfer to observed markets.

We now evaluate the simulation-trained hedgers on 34 overlapping, 30-step historical windows per stock from 2024–2025, with model selection on observed 2020–2023 windows. The tables report entropic risk at the original −1 floor.

The ERM model under Robust training is our robust deep hedging baseline, adapted from Lütkebohmert et al. (2022, Remark 4.2): at each simulation step, we sample a complete parameter vector from rolling historical estimates. WRAP adds adversarial training to the same training distribution.

Fixed-parameter training

StockERMOTWRAP
AAPL0.5960.2450.245
AMZN0.3780.3050.305
BRK-B0.2200.1710.198
GOOGL0.2880.2720.272
MSFT0.2440.2030.203

Parameter-robust training

StockRobust deep hedging (adapted)OTWRAP
AAPL0.6320.3210.321
AMZN0.4150.3210.321
BRK-B0.2730.1570.187
GOOGL0.3220.3040.304
MSFT0.1960.2010.201

Here, validation selects WRAP from all 42 budget pairs, including zero budgets. In our response to Reviewer BR6s, Q3, we restrict WRAP to six pairs, matching OT’s search size.

WRAP has lower mean risk than ERM in nine settings. Validation selects τ=0\tau=0 in eight, recovering OT. The two joint-budget selections have higher test risk than OT. One possible explanation is that transport already captures the main shifts in these windows, leaving little benefit from reweighting. Our daily Yahoo Finance sample provides few non-overlapping windows, which may limit how reliably we can distinguish the methods.

Reviewer_8X4zOpenReview ↗

Reply to the rebuttal

Thank you for your rebuttal. I maintain my original scores.

Reviewer_meM9OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a distribution-robust adversarial training framework called WRAP for deep hedging under non-stationary market dynamics. It integrates drift-aware empirical weighting, ϕ\phi-divergence scenario reweighting, and Wasserstein path perturbations. The theoretical contributions include a distributional coverage bound, a joint first-order robust-loss expansion, and an explicit adversarial attack. Experiments on synthetic Heston and market-calibrated diffusion models report reduced hedging risk under distribution shift.

Critical strengths
  1. The paper is decently organized.
  2. The derivations appears to be technically sound.
  3. The empirical evaluation includes some helpful ablation studies, and varied risk metrics that clarify the framework's mechanics.
Critical weaknesses
  1. Theoretical novelty appears to be limited. The core framework appears to adapt established results of drift-aware weighting and coverage bounds from [1], standard transport–divergence ambiguity sets, and known single-budget sensitivity expansions. The central theoretical result is an additive combination of existing terms that fails to characterize cross-mechanism interactions.
  2. Theory-implementation gap is non-trivial. The empirical instantiation diverges largely from the theoretical derivations. The coverage radius is never numerically calibrated, parameters and budgets are tuned via validation, linear tilting is replaced by exponential tilting, and the training objective optimizes a heuristic clean/adversarial mixture. Additionally, relatively large perturbation budgets undermine the local first-order approximation, which lacks empirical verification.
  3. Experimental design and comparisons are potentially unfair. Synthetic benchmarks test only monotonically increasing drift and volatility, potentially biasing results toward the proposed perturbation direction. Moreover, WRAP receives a substantially larger hyperparameter tuning budget than baselines.

References
[1] Keehan, D. S., Anderson, E. J., & Wiesemann, W. (2025). Don't Look Back in Anger: Wasserstein Distributionally Robust Optimization with Nonstationary Data. arXiv preprint arXiv:2510.18566.

Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

Clarifying the theoretical contributions, implementation, and experimental design

We thank the reviewer for the constructive assessment of our work. We address the stated weaknesses below.

W1.Theoretical novelty.

We agree that Wasserstein DRO, ϕ\phi-divergence DRO, and drift-aware weighting have been studied previously, including formulations combining optimal transport (OT) and ϕ\phi-divergence constraints. Our contribution is a tractable two-budget DRO framework for path-dependent hedging losses, together with a joint sensitivity analysis and an explicit first-order accurate adversary.

Importantly, the joint expansion does not follow from existing single-budget results. First, we establish an exact reduction of the robust value to an outer optimization over scenario probabilities and an inner transport problem over trajectory distributions (Proposition Exact empirical reduction, Appendix C.3.1).

Second, the transport problem depends on the adversarially chosen scenario probabilities. Thus, a Wasserstein expansion for fixed probabilities does not justify optimization over reweightings. We resolve this by proving a uniform OT expansion over probabilities near the baseline distribution (Lemma Uniform OT expansion, Appendix C.3.1). We then establish that all divergence-feasible reweightings lie in this neighborhood as τ↓0\tau\downarrow0, yielding the joint first-order expansion (Theorem 3.6). Consequently, the additive leading-order terms are derived rather than assumed, while the mechanisms remain coupled at finite budgets. Their interaction can give rise to higher-order cross terms, which are not captured by the first-order expansion. Finally, we construct an explicit adversary with one adversarial weight and one perturbed trajectory per observation, prove its feasibility under both budgets for sufficiently small budgets, and establish that it attains the robust value up to an error of o(ε+τ)o(\varepsilon+\sqrt{\tau}) (Theorem 3.7; Appendix C.3.2). We will make these technical contributions and their distinction from existing results more explicit in the manuscript.

W2. Theory–implementation gap. We agree that Theorems 3.6–3.7 establish first-order guarantees for small budgets, not finite-budget optimality guarantees. We would like to clarify the three implementation choices highlighted.

First, exponential tilting solves the KL-constrained reweighting problem exactly for fixed trajectory losses, whereas linear tilting provides its small-budget approximation. At larger budgets, the nonnegativity constraints 1+ατYn≥01+\alpha_\tau Y_n \geq 0 can prevent linear tilting from exhausting the KL budget. We therefore use exponential tilting, computed through one-dimensional bisection, to avoid this limitation (Appendix D, Experiments).

Second, combining clean and adversarial losses follows established adversarial training practice, including He et al. (2025, Eq. 5.1), who use a weighted combination of these losses for deep hedging. We adopt equal weights (1:1) and acknowledge that this practical training objective differs from the exact DRO minimax objective. We will clarify this distinction in the manuscript. Third, we distinguish the theoretical coverage guarantee from empirical parameter selection. The coverage theorem provides a conditional statistical guarantee, while the baseline weights and ambiguity budgets in our experiments are selected using held-out validation performance. This is consistent with Keehan et al. (2025), who advocate cross-validation for both the Wasserstein radius and the drift parameter governing the weights, as well as He et al. (2025, Appendix D), who select adversarial training budgets through validation.

W3. Experimental Design

We added decreasing variance, partial variance reversion, stable variance and reversed price drift, keeping trained hedgers and validation-selected budgets fixed (see the table in Reviewer BR6s, Q2). Gains extend beyond the original increasing regime. Under declining variance, performance differences narrow, with ERM achieving the lowest CVaR95 under the strongest variance reduction.

Our main focus is robustness under nonstationarity. Figure 4(a–b) shows that the best tested configurations require both budgets to be positive. With the same data, architecture, seeds and validation criterion, joint configurations reduce validation CVaR by 12.6% and 10.5%, versus at most 9.1% and 6.8% along the single-budget axes, under uniform and drift-aware weights, respectively. Increasing the OT budget alone therefore does not recover these gains on the evaluated grid. Improvements also span a broad region with both budgets positive. We acknowledge the unequal candidate counts and discuss tuning effort further in Reviewer 4, Q4. Reviewer BR6s, Q3 compares six WRAP budget pairs with six OT settings for a fairer tuning comparison in GAD. When reweighting adds no benefit, validation can select τ=0\tau=0, recovering OT (Figure 3(a), Tables 3 and 5). The grid thus identifies when one mechanism suffices and when their combination helps.

Reviewer_BR6sOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WRAP trains a deep hedger against an adversary that reweights market trajectories under a ϕ\phi-divergence budget τ\tau and perturbs them under an optimal-transport budget ε\varepsilon. A first-order expansion yields a one-step attack (Theorems 3.6–3.7, Algorithm 1). On a nonstationary Heston model, the paper reports a 19.2% lower far-future CVaR than clean training (Table 1). It also reports results on GAD models calibrated to five stocks (Table 9).

Critical strengths
  1. The problem is relevant and the formulation is clear. A hedge for nonstationary markets must be robust both to which scenarios are likely and to how individual paths could differ, and the two-budget ambiguity set captures both: τ\tau bounds the ϕ\phi-divergence reweighting and ε\varepsilon the optimal-transport perturbation. Theorems 3.6–3.7 give an explicit first-order adversary with one probability multiplier and one perturbed trajectory per observation.
  2. The attack is cheap and connects to prior work. Algorithm 1 needs one loss and gradient evaluation per batch, reduces to the one-step WPGD attack of He et al. (2025) when τ=0\tau=0 and the weights are uniform, and a multistep variant is provided as a stronger benchmark (App. D.3).
  3. The evaluation is broad and the protocol is careful. Budgets and training epochs are selected on validation data, and all results are reported on held-out test data over five seeds. The experiments cover stationary and nonstationary Heston markets, transaction costs, a Mean–CVaR objective and GAD models calibrated to five stocks under two training distributions (Tables 1, 3, 6, 7, 9), and the paper reports openly where a channel does not help (Tables 3, 6, 9).
  4. The mechanisms are made visible and the work is reproducible. Figs. 1–2 and App. D.7 show how reweighting concentrates probability on high-loss paths while transport mainly raises realized variation, and the code, configurations and reproduction instructions are released.
Critical weaknesses
  1. The main experiment favors the attack. Variance and drift rise across the Heston validation and test blocks, and transport mainly raises realized variation (p. 8), so shift and attack point the same way. Under CVaR, ℓDH\ell_{\mathrm{DH}} is flat below mm, so both channels act only on the tail. Neither a shift in another direction nor ERM at a higher CVaR level is tested, and all baselines are ablations of WRAP (only OT-U corresponds to WPGD).
  2. The attacked series are not shown to be plausible: the perturbation at date tt depends on future values through the pathwise gradient and, with κt=1\kappa_t=1, has the same magnitude and a gradient sign at every date (Eqs. 11, 17). Figs. 1–2 show attacked paths but do not examine their realism.
  3. Joint training is not shown to beat either channel alone. In Table 9, the single-budget variants use only positive budgets and are worse than ERM-U on validation in all ten comparisons; with the clean origin admitted, as for WRAP-U, they revert to ERM-U, and WRAP-U then wins 6, ties 3 and loses 1. Tuning is not matched (42 budget pairs vs. one axis), and all evaluation paths are simulated, although the released code documents a historical-window protocol.
  4. The first-order analysis does not cover the attack as used. Theorems 3.6–3.7 are local (the proof needs τ<wmin⁡mδ\tau<w_{\min}m_\delta with wmin⁡≤1/Nw_{\min}\le 1/N, p. 30), yet selected budgets reach τ=1.2\tau=1.2, and τ=0.6\tau=0.6 already cuts the effective sample size from 4,000 to 691 (App. D.7). Training also departs from the analysis (exponential tilting, 1:1 clean/adversarial mix; Remark D.1, Eq. 44), and one-step and multistep attacks are compared only in the stationary setting, where τ=0\tau=0 is selected (Table 10).
Questions for authors
  1. At the selected nonstationary budgets, e.g., (ε,τ)=(0.30,0.60)(\varepsilon,\tau)=(0.30,0.60), how close is the one-step attack to the multistep attack?
  2. Does the WRAP gain hold when variance decreases or reverts in the test blocks? Does it hold against ERM trained at CVaR 0.97/0.99 and against a published robust method?
  3. What are the GAD results on historical windows, with the clean origin admitted for all methods and with matched tuning?

I would raise my rating to 3 if the gain holds under Question 2.

3 comments
AuthorsOpenReview ↗

Additional experiments: attack strength, regime shifts, and historical data

Thank you for your careful reading and constructive feedback, which helped us strengthen the experiments significantly. All tables report means over five seeds.

Q1. Multi-step performance.

We compare one-step and multistep attacks following the paper's setup. The hedger, paths and attack threshold are fixed. For each seed, we measure the increase in the objective compared with no attack and set the largest increase among 1, 5, 20 and 50 steps to 100%.

Attack stepsWRAP-U relative gain (%)WRAP-W relative gain (%)
196.3096.63
599.5799.90
2099.8899.87
50100.0099.93

The budgets (ϵ,τ)(\epsilon,\tau) are (0.3, 0.6) for WRAP-U and (0.4, 1.0) for WRAP-W. All attacks satisfy the constraints. One-step captures 96% of the best observed gain at these budgets. For WRAP-W, five steps slightly outperform 50 in two seeds, hence 99.93%.

We also train WRAP-U with one-step and 20-step attacks at (ϵ,τ)=(0.3,0.6)(\epsilon,\tau)=(0.3,0.6) for 600 epochs, including 300 warmup epochs.

Training attackNear CVaR95Near CVaR99Far CVaR95Far CVaR99
One step6.6417.6747.2688.466
20 steps7.0368.1417.7079.003

One-step training has lower CVaR in all four comparisons, but admittedly the budgets were selected for one-step training. Comparing both methods after separate tuning would require a full budget search for multistep training. Hence, one-step training captures most of the gains while multistep training worsens the hedging performance.

Q2. WRAP performance.

We train ERM at CVaR97 and CVaR99, then evaluate all models at both CVaR95 and CVaR99.

MethodNear CVaR95Near CVaR99Far CVaR95Far CVaR99
ERM-U, trained at 95%7.95210.0938.99211.526
ERM-U, trained at 97%7.95910.1069.05511.619
ERM-U, trained at 99%7.92110.0279.00111.547
OT-U7.0508.5187.8119.405
WRAP-U6.6417.6747.2688.466

WRAP has lower risk than OT-U, corresponding to WPGD in He et al. (2025), and remains 16–27% below the best ERM baseline even when ERM is trained at higher confidence levels.

We then change only the future test regime, keeping hedgers and selected budgets fixed. The table reports CVaR95 with uniform reference weights.

RegimeHorizonERM-Uϕ\phi-UOT-UWRAP-U
OriginalNear7.9527.2397.0506.641
OriginalFar8.9928.1287.8117.268
Decline .10 to .04Near5.8115.5495.4535.364
Decline .10 to .04Far3.7533.8543.7723.838
Partial .10 to .08Near5.8115.5495.4535.364
Partial .10 to .08Far5.0924.9934.8764.850
Stable varianceNear6.8836.4006.2186.001
Stable varianceFar6.9386.4326.2886.071
Reversed driftNear7.7087.0196.8086.464
Reversed driftFar8.5707.8037.3836.997

Variance pairs are (θNear,θFar)(\theta_{\mathrm{Near}},\theta_{\mathrm{Far}}), with v0=θv_0=\theta. We set θ=.125714\theta=.125714 for stable variance and μ=(−.10,−.20)\mu=(-.10,-.20) for reversed drift. Other parameters are unchanged.

WRAP-U reduces CVaR95 by 4.8–18.4% relative to ERM-U under partial reversion, stable variance and reversed drift without retuning. Only the strongest Far variance decline is a limitation: WRAP-U CVaR95 is 2.3% higher than ERM-U. In this low-variance scenario, adversarial training slightly worsens CVaR95 relative to clean training.

Q3. Historical GAD: restricted search.

Training is simulation-based; validation covers observed 2020–2023 windows (with Covid). We report entropic risk on 34 overlapping 30-day test windows per stock from 2024–2025. For this check, we restrict WRAP's 42-pair grid to six pairs with τ=2ϵ\tau=2\epsilon and ϵ∈{0,.025,.05,.10,.15,.20}\epsilon\in\{0,.025,.05,.10,.15,.20\}. OT has six candidates and ϕ\phi seven, clean included. We select by validation risk for each stock and variant. ERM in the Robust rows is trained on paths with parameters drawn from rolling historical estimates, adapting Lütkebohmert et al. (2022). This provides a separate robust deep hedging baseline for GAD.

TrainingStockERM-UOT-Uϕ\phi-UWRAP-U (subset)
FixAAPL.596.245.591.512
FixAMZN.378.305.378.353
FixBRK-B.220.171.295.171
FixGOOGL.288.272.288.272
FixMSFT.244.203.244.194
RobustAAPL.632.321.597.627
RobustAMZN.415.321.471.327
RobustBRK-B.273.157.314.187
RobustGOOGL.322.304.322.275
RobustMSFT.196.201.314.191

WRAP has lower mean risk than ERM in all ten settings, ϕ\phi in nine and OT in five; gains over OT are mixed.

Reviewer_BR6sOpenReview ↗

Response to the authors' rebuttal

Thank you for the response. The new results address my main concern that the gain relies on a test shift aligned with the attack: it holds against ERM trained at CVaR 0.97/0.99 and when variance stays stable or partly reverts, so I am raising my rating to 3. One point remains: on historical windows with matched tuning, WRAP-U has lower risk than OT-U in three of ten settings and higher risk in five; could the authors discuss when the joint budgets help beyond OT in this setting? As a minor suggestion, an overview figure of the two-budget attack, showing how reweighting and transport act on a batch of paths, would help readers before Algorithm 1.

AuthorsOpenReview ↗

Thank you!

Thank you for the positive reassessment!

Additional reweighting may help when adverse scenarios are underrepresented. Here, however, tuning on the COVID-era validation period may produce already these adverse scenarios. This results in overly conservative budgets for the less volatile test period 2024-2025 by putting too much weight on these adversarial scenarios, hence, the reweighting does not bring any consistent advantages. Furthermore, the small sample of 34 overlapping test windows per stock further limits reliable comparison. In the final version, we will expand the historical evaluation and add the suggested overview figure to better illustrate how reweighting and transport act on a batch of paths. Thank you for that suggestion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.