acceptodds
Under review as a conference paper at ICLR 2027

Diachronic Sample Integration: Robust Tail-Risk Estimation with Generative Models

Abstract

Deep generative models are increasingly used as simulators for downstream decision making under data scarcity, but in risk-sensitive applications their usefulness depends on rare adverse scenarios rather than typical samples. Standard generative objectives prioritize bulk distributional fidelity, leaving low-probability tails vulnerable to localized optimization noise and making tail-dependent functionals unstable under finite simulation budgets. We introduce Diachronic Sample Integration (DSI), a test-time inference framework that ensembles generated samples across checkpoints from a stochastic training trajectory. DSI targets a checkpoint-mixture distribution that averages checkpoint-specific tail fluctuations rather than relying on a single brittle endpoint. We formalize this mechanism through a finite-budget bias-variance theory. Empirically, across multivariate synthetic processes and high-frequency trading data, DSI substantially reduces tail-estimation error compared to single-checkpoint baselines under fixed simulation budgets, outperforming standard diffusion and state-of-the-art tail-aware baselines without modifying the generative objective.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Generative Models, Risk Modeling, Diffusion Models, Financial Risk

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PbcjOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3(+2)
Soundness
2
Read review
Paper summary

The authors propose Diachronic Sample Integration (DSI), a method for estimating tail-risk measures (VaR and ES) with generative models. Instead of drawing all synthetic samples from one trained checkpoint, a fixed sampling budget is split across several checkpoints from the same training run. The method itself is not tail-specific, but the authors argue that tail behaviour is learned poorly and fluctuates between checkpoints, so pooling lets these checkpoint-specific errors partly cancel. This is supported with a bias-variance decomposition, in which the error splits into a persistent bias that pooling cannot remove and a fluctuation term that shrinks as more weakly correlated checkpoints are pooled. The experiments follow the TailGAN benchmark, and VaR and ES are estimated from generated paths on synthetic and NASDAQ data.

Critical strengths
  • The method is simple and cheap to use. There is no need to change the model, training objective or architecture. It can in principle be applied to any generator. The comparison in Table 7 is a nice result: the pooling of checkpoints from only one training run reaches about the same accuracy as an ensemble of independently trained models.
  • The paper is open about when the method is assumed not to help (persistent bias, strongly correlated checkpoint errors), and the diagnostics in Table 2 match the results in Table 1. I appreciated that the paper explains its failure cases instead of only reporting where it works.
Critical weaknesses
  1. It is hard to understand what is being generated and why.
  • The introduction motivates the method in general terms ("rare adverse scenarios", "tail-dependent functionals") but never describes the actual setup: the model generates multi-asset price paths, these are passed through trading strategies to get profit-and-loss distributions, and VaR and ES are estimated from those. I only understood this once I reached Section 5.
  • The paper also does not really explain why a generative simulator is needed instead of estimating risk directly from historical data. This matters because historical simulation does about as well as DSI in Table 1.
  • A short, concrete description of the pipeline and its motivation early in the paper would make the contribution much easier to follow.
  1. Some results in Table 1 are hard to interpret, and a few claims go further than the results.
  • Some reported numbers need more explanation, in particular the standard deviation of DSI and the backtest results for the Oracle (see Questions 1 and 2).
  • The abstract says DSI outperforms "state-of-the-art tail-aware baselines", but TailGAN is the only tail-aware method in the comparison.
  • Section 5.2 says DSI keeps the dependence structure "without visibly degrading" it, while Figure 2(a) suggests it is somewhat worse than the single-checkpoint DDPM (though still clearly better than the GANs).
Questions for authors
  1. In Table 1, DDPM (DSI) has a standard deviation of ±0.2 on synthetic data, while SE(1000), which samples directly from the true distribution with the same budget of 1000, has ±2.3. How can DSI vary less across runs than sampling from the true distribution? Are the generated samples redrawn in every repetition?
  2. The Oracle is rejected by the Kupiec coverage test in 16.32% of trials, and HSM in 17.19%. For the true data-generating process I would expect a rejection rate close to the nominal level of the test. How do you explain this, and how should DSI's 20.26% be read relative to it?
  3. On synthetic data, historical simulation (3.4) has a lower error than DDPM (DSI) (3.7) and a lower coverage rejection rate. In which situations would you expect a generative simulator with DSI to be preferable to historical simulation?
  4. Did you consider comparing against other tail-aware generative methods mentioned in the related work, such as FIDE or heavy-tailed diffusion models?
7 comments
AuthorsOpenReview ↗

R3 (1/3) - Weaknesses

We thank the reviewer for the careful reading and for recognising the practical value of checkpoint reuse and the diagnostics explaining when pooling provides limited benefit. We appreciate the candid feedback on the writing and hope that the revised presentation below makes the setup and motivation clearer. We will upload a revised PDF with changes highlighted in blue before the discussion period ends. We also clarify the reported results and summarise the additional baseline comparisons.

W1: Simulation pipeline and motivation

“It is hard to understand what is being generated and why.”

Following Tail-GAN [1], our application is scenario simulation for portfolio risk management: generating realistic joint asset-price dynamics to estimate the risk of large portfolio losses. For static portfolios, this requires preserving cross-asset dependence, while dynamic trading strategies also depend on the temporal evolution of prices. We therefore evaluate generated scenarios through the Value-at-Risk (VaR) and Expected Shortfall (ES) of the profit-and-loss distributions they induce for benchmark static and dynamic strategies. DSI improves this simulation procedure by dividing a fixed generation budget across checkpoints from one training run and pooling their generated scenarios before estimating these risk measures.

The terms “rare adverse scenarios” and “tail-dependent functionals” refer to extreme losses and quantities sensitive to the tails of these distributions. Our submission emphasised the probabilistic formulation, in line with its primary area of Probabilistic Methods (Sampling and Uncertainty Quantification). We agree an earlier description of the concrete pipeline would make the connection to the application easier to follow, and will update our manuscript accordingly.

“The paper also does not really explain why a generative simulator is needed instead of estimating risk directly from historical data.”

Historical simulation is a useful reference when sufficient, representative observations are available. Generative simulators are used in several settings where direct reuse of historical observations is insufficient for the downstream task:

  • Scarce tail observations. Extreme losses are sparsely represented in historical data, making tail-risk estimation and evaluation of trading strategies challenging. For example, a dataset of 1,000 observations contains approximately 50 observations in a 5% tail. A learned simulator can supply additional data and scenarios for downstream analysis. Much of the existing work on synthetic data generation emphasises fidelity to the bulk of the distribution. In these risk-management applications, however, the need for synthetic scenarios arises precisely because extreme observations are scarce, making accurate tail behaviour central to the purpose of generation.
  • Data privacy. Restrictions on access to or sharing of data containing privacy-sensitive information can motivate the use of synthetic datasets for downstream modelling. Their suitability depends on both the required statistical fidelity and the applicable privacy requirements.
  • Stress and scenario testing. Risk management often requires evaluating strategies or pricing models under rare adverse conditions. The use of historical data is sometimes not allowed as the results might overfit on the historical observations. Simulation supports such analyses, with the usefulness of the generated scenarios depending on the model's representation of the relevant risks.

These applications motivate improving the tail fidelity of learned simulators. The challenge extends beyond sample scarcity: empirical VaR/ES computation involves ranking and sorting losses, introducing non-smooth operations into gradient-based training. The “Economic importance” discussion in the general response provides further context.

W2: Interpretation of results and scope of claims

“Some reported numbers need more explanation, in particular the standard deviation of DSI and the backtest results for the Oracle (see Questions 1 and 2).”

We address the standard deviations in Q1 and the Oracle coverage results in Q2 below.

“The abstract says DSI outperforms ‘state-of-the-art tail-aware baselines’, but TailGAN is the only tail-aware method in the comparison.”

TailGAN [1] was the direct tail-aware comparator in the submission. We have since added Score-based Heavy-tailed Diffusion (SHD) [2], which studies multivariate financial returns and VaR estimation. Table G1 in the general response reports this comparison, with DDPM (DSI) achieving lower tail-risk error than both TailGAN and SHD on the synthetic and market benchmarks. Q4 discusses this baseline and the additional comparisons requested by Reviewer 1.

AuthorsOpenReview ↗

R3 (2/3) - Q1-3

“Section 5.2 says DSI keeps the dependence structure ‘without visibly degrading’ it, while Figure 2(a) suggests it is somewhat worse than the single-checkpoint DDPM (though still clearly better than the GANs).”

Following the evaluation in TailGAN [1], correlation and autocorrelation (Corr/ACF) diagnostics serve as checks that generated scenarios retain the basic dependence patterns of the reference data. These linear summaries provide a partial view of the nonlinear dependence relevant to portfolio tail risk. Their intended role is to assess whether the overall patterns in the correlation heatmaps and the shapes of the ACF curves are preserved, rather than to rank simulators by raw distance alone. The distances should therefore be interpreted alongside these plots, with attention to whether generation introduces structural distortions. Although Figure 2(a) reports somewhat higher distances for DSI than for single-checkpoint DDPM, DSI retains the overall dependence patterns while improving VaR/ES estimation. We will clarify this diagnostic role in the final manuscript and explain that the retained patterns, rather than identical distance values, underpin our description of preserved dependence structure.

Q1: Variability across repetitions and sample redrawing

“How can DSI vary less across runs than sampling from the true distribution?”

SE(1000) represents the error of ordinary i.i.d. Monte Carlo sampling from the true distribution. TailGAN [1] introduced it as an empirical reference for average relative error (RE), not a lower bound on the variance of RE. TailGAN [1] also reported standard deviations below those of SE(1000).

Crucially, DSI does not use naive Monte Carlo sampling. It combines checkpoint ensembling with stratified sampling across checkpoints, which can reduce both checkpoint-specific fluctuations and sampling variance. The consistently reduced standard deviations in Table 1 are consistent with this variance-reduction mechanism.

“Are the generated samples redrawn in every repetition?”

Yes, the samples are redrawn in every repetition.

Q2: Oracle and DSI coverage rejection rates

“The Oracle is rejected by the Kupiec coverage test in 16.32% of trials, and HSM in 17.19%. For the true data-generating process I would expect a rejection rate close to the nominal level of the test. How do you explain this, and how should DSI's 20.26% be read relative to it?”

The nominal 5% rejection rate corresponds asymptotically to testing the exact population VaR. Our Oracle instead estimates VaR from 1,000 samples drawn from the true process, so the estimated threshold has sampling uncertainty, which can raise the rejection rate above the nominal level. The Oracle therefore provides a finite-sample reference under the same limited estimation budget.

DSI's rejection rate of 20.26% is 3.94 percentage points above the Oracle's 16.32%, compared with 53.62% for single-checkpoint DDPM. DSI achieves the lowest coverage rejection rate among the learned methods in Table 1 and comes closest to the finite-sample Oracle reference. Historical simulation (HSM), at 17.19%, also remains close to that reference.

Q3: When to prefer a generative simulator

“On synthetic data, historical simulation (3.4) has a lower error than DDPM (DSI) (3.7) and a lower coverage rejection rate. In which situations would you expect a generative simulator with DSI to be preferable to historical simulation?”

The settings described in W1, including restricted access to historical data and downstream scenario-generation requirements, motivate using a generative simulator. DSI improves the tail-risk accuracy of such a simulator through checkpoint reuse under a fixed generation budget. On the synthetic benchmark, HSM has lower RE (3.4 versus 3.7), while on the market benchmark DSI has lower RE (9.6 versus 11.4). These results support DSI as a practical improvement to learned simulation, with the choice between a simulator and historical simulation depending on the available data and the downstream task.

AuthorsOpenReview ↗

R3 (3/3) - Additional Baseline and References

Q4: Additional baseline comparisons

“Did you consider comparing against other tail-aware generative methods mentioned in the related work, such as FIDE or heavy-tailed diffusion models?”

Yes. We considered SHD [2], which we identified after submission, and the heavy-tailed diffusion models of Pandey et al. [3], both of which incorporate Student-t distributions into the diffusion process. We chose to reproduce SHD as the additional heavy-tailed baseline because it was developed for a similar purpose, with experiments on multivariate financial data and VaR estimation that closely match our evaluation setting than the weather-focused experiments of [3]. Table G1 in the general response reports the comparison, with DDPM (DSI) achieving lower tail-risk error on both benchmarks.

We have also added the comparisons requested by Reviewer 1: snapshot ensembling, post-hoc EMA, and dispersion-adjusted single-checkpoint controls, including variance matching and oracle inflation. DSI achieves comparable tail-risk error to snapshot ensembling, while post-hoc EMA and the dispersion controls have higher errors on both benchmarks. The “Additional baselines” section of the general response reports the results, and our response to Reviewer 1 provides the implementation details.

We also considered FIDE [4] and made repeated implementation attempts. The main difficulty in constructing a comparable baseline is the difference between its generation setting and our multivariate portfolio-risk task:

  • Multivariate adaptation. FIDE is formulated for a univariate series x∈RTx\in\mathbb{R}^T conditioned on a scalar block maximum. Extending it to joint multi-asset paths would require specifying how the conditioning mechanism represents multivariate extremes and their dependence, introducing methodological choices beyond the original formulation.
  • Conditional generation. FIDE requires a block maximum as its generation condition, whereas our evaluation uses unconditional generation. FIDE can sample this condition from a fitted generalised extreme-value (GEV) distribution, but doing so introduces an additional parametric tail assumption. We preferred to avoid introducing such assumptions, since the choice of tail family and its suitability for our data would require additional justification for risk-modeling in regulated environments.

Our implementation attempts did not yield satisfactory performance in this setting. Given the adaptations required, we could not establish a comparison that faithfully retained FIDE's original formulation while matching our task.

References

[1] R. Cont, M. Cucuringu, R. Xu, and C. Zhang. “Tail-GAN: Learning to Simulate Tail Risk Scenarios”. Management Science, 2025. https://doi.org/10.1287/mnsc.2023.00936.

[2] H. Liu, T. Zhu, N. Jia, J. He, and Z. Zheng. “Learning to Simulate from Heavy-Tailed Distribution via Diffusion Model”. Operations Research, 2026. https://doi.org/10.1287/opre.2024.1450.

[3] K. Pandey, J. Pathak, Y. Xu, S. Mandt, M. Pritchard, A. Vahdat, and M. Mardani. “Heavy-Tailed Diffusion Models”. International Conference on Learning Representations (ICLR), 2025.

[4] A. H. Galib, P.-N. Tan, and L. Luo. “FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation”. Advances in Neural Information Processing Systems, vol. 37, 2024.

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

This paper proposes Diachronic Sample Integration, which pools samples generated by multiple checkpoints from a single training run to estimate Value-at-Risk and Expected Shortfall. Under a fixed generation budget, the method improves tail-risk estimation on synthetic and financial market data. The theoretical analysis decomposes estimation error into persistent bias, correlated checkpoint fluctuations, and finite-sample noise.

Critical strengths

The method substantially improves VaR and ES estimation without additional training or a larger generation budget. By reusing checkpoints from one training run, DSI achieves accuracy comparable to an independently trained deep ensemble, offering a reduction in training cost.

Critical weaknesses
  • The methodological and theoretical novelty is limited. DSI applies checkpoint ensembling to tail-risk estimation by pooling generated samples. Section 4 accounts for nonlinear VaR/ES estimation and finite sampling, but its main conclusion remains the familiar ensemble principle that averaging reduces imperfectly correlated fluctuations while preserving shared bias. The additional insight beyond applying this principle to tail-risk estimation is unclear.

  • The empirical contribution is limited to financial tail-risk estimation. The experiments evaluate VaR and ES of trading-strategy returns on synthetic and market data. Given the limited methodological novelty, evidence that checkpoint sample pooling benefits tail-sensitive tasks beyond this financial setting would strengthen the paper’s broader relevance to generative modeling.

Questions for authors

How should the checkpoint window, count, and stride be selected for a new dataset? Does the same selection rule remain effective across different learning-rate schedules?

3 comments
AuthorsOpenReview ↗

R2 (1/2) - Weaknesses

We thank the reviewer for recognising DSI's improvements under a fixed generation budget and its comparable accuracy to an independently trained ensemble using a single training trajectory. Below, we clarify the additional theoretical insight and address the practical checkpoint-selection question.

W1: Methodological and theoretical contribution

“The methodological and theoretical novelty is limited. DSI applies checkpoint ensembling to tail-risk estimation by pooling generated samples.”

The simplicity of DSI is deliberate and is its main practical advantage. DSI leaves the architecture, objective and training schedule unchanged, so it can be applied post hoc to any trained generator whose checkpoints have been retained. Snapshot ensembling [1] instead requires a cyclic learning-rate schedule during training, and a deep ensemble requires one training run per member. With a standard DDPM, DSI achieves lower tail-risk error than TailGAN [2] and SHD [3], which modify the generator specifically for tail risk (3.7 vs 5.1 and 6.0 synthetic, 9.6 vs 12.1 and 26.4 market, general response Table G1). As the reviewer notes, it also matches an independently trained deep ensemble from a single training run (3.7 vs 3.9 synthetic, 9.6 vs 9.8 market). That a procedure requiring no change to training outperforms purpose-built tail models is itself a methodological result, and it questions whether the added complexity of such models is needed for tail-risk estimation.

“Section 4 accounts for nonlinear VaR/ES estimation and finite sampling, but its main conclusion remains the familiar ensemble principle that averaging reduces imperfectly correlated fluctuations while preserving shared bias. The additional insight beyond applying this principle to tail-risk estimation is unclear.”

The ensemble principle says that averaging helps when member errors are imperfectly correlated, but not whether checkpoints from a single training run are decorrelated enough to help. Independently trained members fluctuate independently, whereas successive checkpoints along one trajectory have serially dependent errors. Theorem 1 measures this dependence through the effective checkpoint count KeffK_{\mathrm{eff}}, and Table 2 shows that it decides the outcome. Twenty DDPM checkpoints are worth about 19 independent ones, and DSI reduces RE from 7.7 to 3.7, whereas 20 TailGAN checkpoints are worth about one, and RE barely changes (5.1 to 5.0). Snapshot Ensembles [1] showed empirically that checkpoint ensembles can work, and our analysis explains when they do and when they do not.

W2: Relevance beyond financial tail-risk estimation

“Evidence that checkpoint sample pooling benefits tail-sensitive tasks beyond this financial setting would strengthen the paper’s broader relevance to generative modeling.”

The additional experiments in Table G4 of the general response evaluate energy, traffic, healthcare, and weather data using domain-specific stress functionals. DSI reduces single-checkpoint DDPM tail-risk error in all four domains under the same total sampling budget.

We would also like to clarify that the submission's primary area is Probabilistic Methods (Sampling and Uncertainty Quantification). Financial tail-risk estimation is one of the most challenging problems in this domain because of noisy observations, rare extreme events, and complex tail dependencies [2,4]. Tail-GAN [2] studies the same use case of generating multi-asset scenarios that preserve the tail risk of static and dynamic trading strategies. This setting also offers standardised evaluation protocols based on VaR and ES, shaped by regulatory requirements for financial risk management [5]. These features make finance a demanding and practically relevant benchmark for our sampling framework. The cross-domain results extend the evidence beyond this application.

AuthorsOpenReview ↗

R2 (2/2) - Question and References

Q1: Checkpoint selection and learning-rate schedules

“How should the checkpoint window, count, and stride be selected for a new dataset?"

We recommend selecting a window where training MSE has broadly stabilised, then choosing the checkpoint count and stride through a small validation sweep, stopping when further increases provide little improvement. Our sensitivity analysis in Figure 2(c) and Table 6 (Appendix D.4) illustrates this principle: with M=2M=2, increasing KK beyond 20 provides no consistent improvement, and larger strides do not improve on this configuration. This supports our use of K=20K=20 and M=2M=2 as a practical balance between accuracy and checkpoint count, and provides a starting point for a new dataset.

"Does the same selection rule remain effective across different learning-rate schedules?”

To test whether the same checkpoint-selection rule remains effective across learning-rate schedules, we repeated the checkpoint-count and stride sweep in Figure 2(c) and Table 6 under constant, step-decay, and cosine-decay schedules. We varied K∈{5,10,20,50}K\in\lbrace 5,10,20,50\rbrace and M∈{1,2,5}M\in\lbrace 1,2,5\rbrace. The table compares these new results with the constant-LR results from the manuscript and reports average VaR/ES relative error (%). Bold entries indicate our default configuration.

Learning-rate scheduleStride MMK=5K=5K=10K=10K=20K=20K=50K=50
Constant14.80 ± 0.464.74 ± 0.344.29 ± 0.183.84 ± 0.11
Constant24.91 ± 0.634.23 ± 0.413.63 ± 0.213.85 ± 0.22
Constant54.76 ± 0.624.47 ± 0.383.82 ± 0.253.84 ± 0.12
Step decay14.75 ± 0.344.81 ± 0.333.48 ± 0.293.66 ± 0.45
Step decay24.24 ± 0.504.12 ± 0.643.39 ± 0.253.37 ± 0.23
Step decay54.65 ± 0.373.93 ± 0.373.59 ± 0.293.42 ± 0.44
Cosine decay14.70 ± 0.344.71 ± 0.293.45 ± 0.293.60 ± 0.43
Cosine decay24.79 ± 0.484.52 ± 0.613.37 ± 0.253.34 ± 0.21
Cosine decay54.72 ± 0.314.59 ± 0.393.49 ± 0.323.44 ± 0.43

Across all three schedules, the sweep shows a similar pattern to Figure 2(c): increasing checkpoint count initially reduces error, followed by diminishing returns. At K=20K=20, M=2M=2 achieves the lowest mean error among the tested strides under every schedule. Increasing KK from 20 to 50 at this stride provides no improvement under constant LR and reduces RE by only 0.02 and 0.03 percentage points under step and cosine decay, respectively. Thus, the new schedules support the same practical selection rule illustrated in Figure 2(c): use a small validation sweep over checkpoint count and stride, stopping when further increases yield little improvement. Our default K=20,M=2K=20,M=2 remains a favourable accuracy–checkpoint-count trade-off across the tested schedules.

References

[1] G. Huang, Y. Li, G. Pleiss, Z. Liu, J. E. Hopcroft, and K. Q. Weinberger. “Snapshot Ensembles: Train 1, Get M for Free”. International Conference on Learning Representations (ICLR), 2017.

[2] R. Cont, M. Cucuringu, R. Xu, and C. Zhang. “Tail-GAN: Learning to Simulate Tail Risk Scenarios”. Management Science, 2025. https://doi.org/10.1287/mnsc.2023.00936.

[3] H. Liu, T. Zhu, N. Jia, J. He, and Z. Zheng. “Learning to Simulate from Heavy-Tailed Distribution via Diffusion Model”. Operations Research, 2026. https://doi.org/10.1287/opre.2024.1450.

[4] Alexander J. McNeil, Rüdiger Frey, and Paul Embrechts. Quantitative Risk Management: Concepts, Techniques and Tools. Princeton University Press, 2015.

[5] Basel Committee on Banking Supervision. Minimum Capital Requirements for Market Risk. Bank for International Settlements, January 2019. https://www.bis.org/bcbs/publ/d457.htm.

Dear Reviewer hDVd,

Thank you again for recognising the practical value of DSI. As the discussion period closes today on November 18, we would like to briefly summarise how we addressed your concerns in the updated manuscript.

  • W1. We substantially revised Sections 1–3, and 6 to clarify DSI’s methodological and theoretical contributions and explain how they differ from prior work.
  • W2. We added experiments on energy, traffic, healthcare and weather data in Section 6 (Cross-Domain Tail-Risk Estimation).
  • Q1. We added practical guidance on selecting checkpoints and comparisons across learning-rate schedules in Appendix D (Checkpoint Selection and Schedule Robustness).

We hope these revisions clarify the method’s contribution and practical scope. We would be grateful if you could reconsider your rating before the discussion period closes. Or let us know if we could make any further improvements.

Reviewer_qA1JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Diachronic Sample Integration (DSI), a test-time procedure for estimating tail-risk measures. It pools generated samples from multiple checkpoints along a single training trajectory and estimates Value-at-Risk (VaR) and Expected Shortfall (ES) from the pooled sample under a fixed budget. Experiments on synthetic processes and NASDAQ data show its effectiveness.

Critical strengths
  1. DSI reuses checkpoints produced during training, so it is easy to use in practice.
  2. The paper includes EMA, SWA, SWAG, and Deep Ensemble controls and uses within-model diagnostics to explain why the adversarial trajectories benefit little.
Critical weaknesses
  1. Pooling samples across checkpoints is essentially snapshot ensembling “Snapshot ensembles: Train 1, get m for free” applied to generative models. EMA also explored similar ideas. A mixture is also smoother than its components, so part of the improvement may simply come from here. The paper is missing three important baselines: a snapshot ensemble, a dispersion-inflated single checkpoint, and a tuned or post-hoc EMA ”Analyzing and Improving the Training Dynamics of Diffusion Models”.
  2. This paper provides little evidence that training satisfies the theory’s assumptions.
  3. The evaluation is limited. It only includes one synthetic process, one short market split, and one budget, and the improvements are primarily observed in DDPM.
Questions for authors
  1. How does DSI differ algorithmically from a snapshot ensemble of generative distributions? Can you add the baselines listed in Weakness 1, and report the EMA decay and how the “best” EMA checkpoint was selected?
6 comments
AuthorsOpenReview ↗

R1 (1/5) - W1

We thank the reviewer for recognising the practical value of checkpoint reuse, the aggregation controls, and the diagnostics explaining the limited gains in adversarial models. Below, we clarify the relationship to existing ensemble methods, explain the evidence supporting the theoretical mechanism, and report the additional comparisons requested in the review.

W1

“Pooling samples across checkpoints is essentially snapshot ensembling ‘Snapshot ensembles: Train 1, get m for free’ applied to generative models.”

Please see Q1 for response.

“EMA also explored similar ideas.”

EMA averages model parameters across training iterations and generates samples from the resulting single model [1,2]. DSI instead generates samples from multiple checkpoint models and pools them, yielding a mixture model that retains the component distributions. Weight averaging and ensembling are distinct operations, as also discussed by Wortsman et al. [3]. In our generative setting,

∑kwkpθk(x)⏟Pooling≠p∑kwkθk(x)⏟Parameter averaging.\underbrace{\sum_k w_k p_{\theta_k}(x)}_{\text{Pooling}} \neq \underbrace{p_{\sum_k w_k\theta_k}(x)}_{\text{Parameter averaging}}.

For illustration, with 10 checkpoints and a budget of 10,000 samples, DSI can generate 1,000 samples from each of 10 different models, preserving differences in their tail behaviour. EMA generates all 10,000 samples from one model with an averaged set of parameters, so every sample comes from the same distribution. Random sampling provides variation within that distribution, but does not retain the diversity across checkpoint distributions.

Consequently, EMA targets the correct distribution only if its averaged parameters induce that distribution. Increasing the sample count reduces Monte Carlo error around the EMA model's distribution but cannot correct a discrepancy between that distribution and the target. DSI instead retains checkpoint-specific tail diversity, allowing opposing tail-probability errors to partially cancel through distribution-level pooling. Shared errors can persist, and our analysis characterises when this diversity improves VaR and ES estimation under a fixed sampling budget. Section 4 provides the theoretical analysis, while Tables 1 and 7 and the additional results in Table G1 of the general response compare pooling with parameter averaging empirically.

“A mixture is also smoother than its components, so part of the improvement may simply come from here.”

The relevance of smoothing depends on the distributional property being considered. In particular, greater smoothness does not by itself imply more accurate VaR or ES estimates. We consider two possible interpretations of this explanation below and would welcome clarification if the reviewer has a different property in mind.

(1) Smoothing of extreme observations. Conventional diffusion models can underrepresent extreme values and heavy tails, as the denoising step often treats extreme values as noise. Galib et al. [4] document difficulties in preserving extreme values, with DDPM block maxima consistently below the true values. Pandey et al. [5] show that standard Gaussian diffusion and flow-matching models can fail to capture heavy-tailed behaviour. Liu et al. [6] establish theoretical challenges in training and sampling from heavy-tailed distributions using conventional diffusion models, while Yu and Yu [7] derive tail-dependent convergence rates and quantify approximation bias. These findings indicate that further smoothing need not improve tail fidelity.

(2) Averaging checkpoint-specific distributional errors. This interpretation is closer to the mechanism studied in our work, but the benefit depends on the structure of those errors. Linear pooling can increase dispersion and does not generally preserve calibration [8]. Averaging also cannot eliminate systematic tail errors shared across checkpoints. DSI exploits checkpoint-specific tail fluctuations that can partially cancel, and our finite-budget bias–variance analysis characterises how persistent bias, checkpoint dependence, and sampling variability affect the resulting VaR and ES estimates.

The central question is therefore when checkpoint pooling improves tail-risk estimation under a fixed sampling budget. Theorem 1 provides the finite-budget MSE decomposition, Corollary 1 identifies the variance-reducible, correlation-limited, and bias-limited regimes, and Proposition 1 characterises the failure modes arising from perfectly correlated errors, shared directional tail misspecification, and persistent bias. Together with the checkpoint diagnostics in Appendix D.2, these results address this question directly, including the regimes in which aggregation provides limited benefit.

AuthorsOpenReview ↗

R1 (2/5) - dispersion-adjusted single-checkpoint definition

“The paper is missing three important baselines: a snapshot ensemble, a dispersion-inflated single checkpoint, and a tuned or post-hoc EMA ‘Analyzing and Improving the Training Dynamics of Diffusion Models’.”

We have added all three requested types of baseline. Table G1 of the general response reports snapshot ensembling, post-hoc EMA, and dispersion-adjusted single-checkpoint controls. Q1 below discusses the aggregation comparisons. Here, we describe the dispersion controls and their selection procedures.

The review suggests a dispersion-inflated single checkpoint without providing a citation or implementation. We therefore inferred a concrete implementation from established work on spread adjustment and predictive-variance calibration. Gneiting and Ranjan [8, Section 3.3] formulate spread adjustment by rescaling centred predictive distributions, while Gneiting et al. [9] calibrate predictive variance through ensemble model output statistics. We adapt these principles to generated return increments, centring on sample means and retaining the original downstream trading strategies. The DSI variance target and oracle searches below are our experimental choices for testing the proposed dispersion explanation.

Return-increment transformation. Following the centred-rescaling principle in [8], let ri,t(n)r^{(n)}_{i,t} denote the generated return increment for asset ii, time tt, and path nn, and let μi\mu_i be its mean over all paths and time steps within a sampling repetition. We apply

ri,t(n,c)=μi+c(ri,t(n)−μi).r^{(n,c)}_{i,t}=\mu_i+c\bigl(r^{(n)}_{i,t}-\mu_i\bigr).

using one common coefficient cc across all assets and time steps, then reconstruct prices and apply all 65 benchmark strategies with their original portfolio weights and trading thresholds. The transformation preserves each asset's sample mean and multiplies its increment variance by c2c^2, while allowing nonlinear, path-dependent strategies to respond to the transformed paths.

Variance matching. Motivated by predictive-variance calibration [9], we define a moment-matching control that uses the corresponding DSI sample as its variance target. For each sampling repetition, we set

c=[∑iVar^(riDSI)∑iVar^(risingle)]1/2.c=\left[\frac{\sum_i\widehat{\mathrm{Var}}(r_i^{\mathrm{DSI}})}{\sum_i\widehat{\mathrm{Var}}(r_i^{\mathrm{single}})}\right]^{1/2}.

This matches aggregate marginal increment variance to that of the corresponding DSI sample. It may yield c<1c<1, so this control permits contraction as well as inflation.

Oracle inflation. We evaluate 81 equally spaced coefficients in [1,2][1,2], then refine the search with 21 points within one coarse-grid step of the best coarse coefficient, retaining previously evaluated candidates. One coefficient per trained model minimises average VaR/ES RE at α=0.05\alpha=0.05 across all 65 strategies and ten sampling repetitions. The more permissive per-panel oracle selects a separate coefficient for each repetition from the same evaluated grid. Both oracles select and evaluate against the same empirical test reference, giving them hindsight access to the evaluation target. Coefficients are shared across strategies and frozen before the rejection tests.

Evaluation and interpretation. We evaluate five trained models per dataset, with ten sampling repetitions per model and N=1000N=1000 generated trajectories per risk estimate. RE is first averaged over sampling repetitions, then reported as mean ± sample SD across the five models. Synthetic-data rejection rates are averaged across strategies and 100 real-data resampling trials per model. This is a separate controlled comparison using raw checkpoints, rather than a replay of the original selected EMA baseline, as indicated in Table G1.

Variance matching leaves synthetic RE unchanged at the reported precision and reduces market RE from 22.3% to 16.4%. The model-level and per-panel oracles yield synthetic RE of 7.4% and 6.9%, and market RE of 13.1% and 13.6%, respectively. Thus, dispersion adjustment can improve tail estimates, particularly on market data, but even these hindsight-selected scalar adjustments do not reproduce the reported DSI errors of 3.7% and 9.6%. These results provide evidence that matching aggregate variance or applying a common scalar inflation alone does not explain the observed improvement from checkpoint pooling.

AuthorsOpenReview ↗

R1 (3/5) - W2-3

W2: Evidence supporting the theory's assumptions

“This paper provides little evidence that training satisfies the theory’s assumptions.”

The assumptions in Appendix B.1 are standard regularity conditions for this type of analysis, and none places a condition on the training objective, optimiser or convergence. Assumptions 1, 3 and 5 are the usual density, moment and linearisation conditions for quantile and ES estimation [10,11,12], and Assumption 4 is the corresponding local condition for first-order expansions of VaR and ES. Assumption 2, that post-burn-in checkpoint errors fluctuate around a persistent bias, rests on the same premise as SWAG [13], a baseline already included in Table 7, which treats post-burn-in SGD iterates as approximate draws from a stationary distribution [14]. Its stationary form is used only to express the fluctuation term through Keff,ψK_{\mathrm{eff},\psi}. The diagnostics in Figure 3, Table 2 and Figure 6 estimate the bias, fluctuation variance and autocorrelation for each trained model, and the resulting regimes match the observed gains (effective checkpoint ratio 0.947 for DDPM vs 0.052 for TailGAN).

W3: Evaluation scope and gains across models

“The evaluation is limited. It only includes one synthetic process, one short market split”

In response to the reviews, we have additionally evaluated SHD, snapshot ensembling, post-hoc EMA, dispersion-adjusted single-checkpoint controls, four non-financial datasets, and different learning-rate schedules. The general response reports the additional baseline comparisons in Table G1 and the cross-domain results in Table G4. R1 W1 describes the dispersion controls, R1 Q1 discusses the aggregation methods, and R2 Q1 reports the schedule comparisons.

We use the same datasets and market split as prior work, TailGAN, to ensure a fair benchmark comparison under a common tail-risk evaluation protocol. The short market split was therefore chosen to preserve comparability with TailGAN. We have added more cross-domain and benchmarks in the general response.

"... and one budget ..."

To assess sensitivity to the sampling budget, we have additionally evaluated DDPM and DSI, from N=100N=100 to N=5000N=5000, using the same budget for both methods at each comparison.

Sampling-budget sensitivity. Average VaR/ES RE (%), reported as mean ± SD. Lower is better.

DatasetMethodN = 100N = 500N = 1000N = 2000N = 5000
SyntheticDDPM15.17 ± 4.0010.53 ± 3.317.70 ± 2.506.06 ± 1.505.80 ± 1.71
SyntheticDDPM DSI9.93 ± 0.944.70 ± 0.733.70 ± 0.202.65 ± 0.172.13 ± 0.15
MarketDDPM38.32 ± 8.9132.58 ± 6.8222.30 ± 4.3016.83 ± 2.2915.71 ± 2.34
MarketDDPM DSI23.30 ± 5.2413.11 ± 2.489.60 ± 1.209.20 ± 1.048.62 ± 0.95

DSI achieves lower average RE than single-checkpoint DDPM at every tested budget on both datasets. Both methods improve as the sampling budget increases, while DSI's advantage persists across the full tested range, showing that the observed gains are not confined to the original budget of N=1000N=1000.

"... and the improvements are primarily observed in DDPM."

The cross-model evaluation examines the distinct regimes predicted by our theory. DDPM illustrates the variance-reducible regime, where checkpoint-specific fluctuations can be averaged out and DSI yields large gains. TailGAN illustrates the correlation-limited regime, where strong checkpoint dependence restricts improvement, while WGAN and the added SHD baseline illustrate the bias-limited regime, where persistent directional errors leave little scope for pooling to help. These contrasting outcomes provide empirical cases for assessing when checkpoint pooling yields substantial gains, modest improvements, or little benefit.

The DDPM results also demonstrate the practical value of applying DSI to a standard generative model. In Table G1 of the general response, DDPM with DSI achieves average tail-risk RE on both datasets. DSI obtains these results by reusing retained checkpoints, without tail-specific fine-tuning, architectural changes, modification of the training objective, or an increased sampling budget. This simplifies implementation and avoids introducing an additional parametric tail family. The latter is useful in regulated risk management, where modelling assumptions require justification and validation, because it removes one additional modelling choice from that process, as discussed in the practical-contribution section of the general response.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response (1/3) - Contributions

We thank the reviewers for their feedback. Subject to the reviewer discussion, we will upload a revised PDF with changes highlighted in blue before the discussion period ends.

Contribution and submission area

Several reviewer comments suggest a misunderstanding of the submission's primary area. We believe this may have caused confusion about the intended contribution and influenced the expectations underlying the evaluation. We therefore clarify the submission area, motivation, and contributions below.

  1. Research area. Our primary area is Probabilistic Methods (Sampling and Uncertainty Quantification). We study tail-risk scenario simulation, the problem addressed by Tail-GAN [1], with generative models serving as simulators for risk management.

  2. Economic importance. Reliable scenarios must preserve the risk of extreme losses for static and dynamic trading strategies [1,2]. Expected Shortfall (ES) measures the severity of losses beyond the Value-at-Risk (VaR) threshold [2] and is central to market-risk capital requirements under the Basel framework [3]. Learning these quantities is challenging because extreme observations are scarce and empirical VaR/ES computation involves ranking and sorting losses, introducing non-smooth operations into gradient-based training. Tail-GAN explicitly addresses this difficulty using differentiable neural sorting [1, Appendix C.3]. Accurate tail-risk simulation therefore has direct implications for financial risk management and capital adequacy.

  3. Theoretical contribution. Snapshot Ensembles [4] demonstrated empirically that checkpoints from one training run can approach independent-ensemble performance, leaving a gap in explaining when and why this occurs. We address this gap for nonlinear VaR/ES estimation under a fixed sampling budget, deriving how effective checkpoint count, dependence, persistent bias, and finite sampling govern the benefits and limitations of pooling.

  4. Practical contribution. Simplicity is a strength of DSI: it reuses existing checkpoints without additional training, architectural changes, or a larger generation budget. DSI achieves comparable tail-risk RE to snapshot ensembling (Table G1), while remaining agnostic to the training objective, optimiser, and learning-rate schedule. Unlike the original Snapshot Ensembles method [4], which requires a cyclic learning-rate schedule, DSI can be applied post hoc provided that suitable intermediate weights have been saved. A standard DDPM with DSI also achieves lower tail-risk RE than the specialised Tail-GAN [1] and Score-based Heavy-tailed Diffusion (SHD) [5] methods on both evaluated benchmarks. This shows that additional task-specific complexity is not required to achieve better tail-risk accuracy in these experiments. DSI also imposes no additional parametric tail family. This matters in regulated risk management, where modelling assumptions require justification and validation [6], particularly when scarce tail observations make misspecification difficult to detect. Avoiding an extra tail-family assumption removes one such modelling choice from the validation burden.

AuthorsOpenReview ↗

General Response (2/3) - Additional baselines

Additional baselines

We have evaluated SHD [5], a relevant baseline identified after submission. SHD and the heavy-tailed diffusion models of Pandey et al. [7] share the central mechanism of incorporating Student-t distributions into the diffusion process. We therefore selected one representative of this approach: SHD, whose experiments on multivariate financial returns and VaR estimation are closer to our evaluation setting than the weather-focused experiments of [7].

We additionally report the snapshot-ensemble [4], post-hoc EMA [8], and dispersion-adjusted single-checkpoint comparisons requested by Reviewer 1. Table G1 brings these results together with SHD and the original baselines.

Table G1: additional baselines using the same protocols as table 1 from our submission.

MethodSynthetic REMarket RECoverage rejectionScore-based rejection
DDPM and snapshot ensemble
DDPM (EMA)7.7 ± 2.522.3 ± 4.353.620.06
DDPM (Post-hoc EMA) [8]6.5 ± 2.319.5 ± 4.544.750.05
DDPM (Snapshot ensemble)3.9 ± 0.39.7 ± 1.320.840.00
DDPM (DSI)3.7 ± 0.29.6 ± 1.220.260.00
DSI applied to TailGAN and SHD
TailGAN5.1 ± 1.412.1 ± 1.333.550.29
TailGAN (DSI)5.0 ± 1.311.9 ± 1.232.870.25
SHD6.0 ± 1.526.4 ± 3.237.180.22
SHD (DSI)5.7 ± 0.426.1 ± 1.030.510.23
Single-checkpoint dispersion controls
DDPM (variance-matched) †7.7 ± 2.516.4 ± 2.152.740.07
DDPM (oracle inflation) †7.4 ± 2.513.1 ± 2.153.550.06
DDPM (per-panel oracle inflation) †6.9 ± 0.913.6 ± 1.448.240.05

Note: † R1 W1 gives the full implementation.

SHD has lower average RE than single-checkpoint DDPM on synthetic data but higher RE on market data. DDPM with DSI achieves the lowest RE on both benchmarks. These results show that explicitly modelling heavy tails does not by itself ensure more accurate downstream VaR/ES estimates in this setting.

DSI and snapshot ensembling achieve comparable tail-risk RE on both datasets. DSI obtains this performance by reusing saved checkpoints without modifying training or increasing the total sampling budget, making it a lightweight option for post-training application. Post-hoc EMA improves on the original EMA baseline, but its RE remains higher than that of both sample-pooling methods.

The separate dispersion-control comparison shows that scalar adjustment helps on market data: variance matching reduces RE from 22.3% to 16.4%, and the two oracle procedures obtain 13.1% and 13.6%. Even with hindsight access to the test reference, these scalar adjustments do not reproduce the reported DSI errors of 3.7% and 9.6%. This supports the interpretation that matching aggregate variance or applying a common scalar inflation alone does not account for the observed pooling gains.

Table G2: Checkpoint-level ES bias–variance diagnostics using the protocol of Table 2 in the submission.

ModelbES2b_{ES}^{2}σES2\sigma_{ES}^{2}Sign consensusKeff,ES/KK_{\mathrm{eff},ES}/K
DDPM7.87×10−107.87\times10^{-10}1.79×10−81.79\times10^{-8}0.5800.947
TailGAN3.14×10−63.14\times10^{-6}1.47×10−51.47\times10^{-5}0.7470.052
WGAN1.32×10−71.32\times10^{-7}3.40×10−83.40\times10^{-8}0.8710.650
SHD2.63×10−82.63\times10^{-8}1.07×10−91.07\times10^{-9}0.8310.703

SHD's modest gain from DSI is consistent with a bias-limited regime. Sign consistency, reported as sign consensus in Table G2, is the fraction of checkpoints sharing the dominant ES error sign, averaged across strategy–seed trajectories. SHD's squared persistent component is approximately 24.6 times its checkpoint fluctuation variance, and the sign consensus of 0.831 indicates that checkpoint errors predominantly share the same direction. The small fluctuation component relative to persistent error leaves limited scope for averaging to improve accuracy.

Table G3: Finite-budget MSE decomposition using the protocol of Table 5 in the submission.

ModelMixture-error MSESampling term ANA_NPredicted finite MSEObserved fixed-budget MSE
DDPM1.727×10−91.727\times10^{-9}6.807×10−96.807\times10^{-9}8.534×10−98.534\times10^{-9}8.535×10−98.535\times10^{-9}
SHD2.639×10−82.639\times10^{-8}6.373×10−96.373\times10^{-9}3.277×10−83.277\times10^{-8}3.299×10−83.299\times10^{-8}

SHD's mixture-error MSE accounts for approximately 80.6% of its predicted finite-budget MSE, and predicted and observed MSE agree within 0.67%. Together, Tables G2 and G3 support the theory's explanation that pooling benefits depend on averageable checkpoint fluctuations relative to persistent error, rather than on model family alone.

AuthorsOpenReview ↗

General Response (3/3) - Cross-domain experiments

Cross-domain experiments

To assess applicability beyond finance, we evaluated DSI on four public time-series datasets spanning energy (PJM Electricity Demand [9]), traffic (NYC Taxi Demand [10]), healthcare (ECG5000 [11]), and weather (Monash Weather Dataset [12]).

For each dataset, we train a standard DDPM and compare single-checkpoint sampling with DSI under the same total simulation budget of N=1000N=1000. DSI uses K=20K=20 post-burn-in checkpoints with stride M=2M=2, as in the main experiments. Application-specific stress functionals measure peak electricity demand, congestion spikes, abnormal ECG excursions, and extreme weather events. We report average VaR/ES RE at α=0.05\alpha=0.05, using held-out observations as the reference.

We also attempted to adapt Tail-GAN and WGAN to these cross-domain experiments, consulting their source code and making repeated hyperparameter-tuning attempts. Training remained unstable in our implementations, and we were unable to obtain stable runs suitable for comparison. We therefore report the cross-domain results for DDPM and DSI alongside the statistical baselines in Table G4.

Table G4: Cross-domain tail-risk estimation. Average RE (%) of VaR and ES at α=0.05\alpha=0.05, with N=1000N=1000. Lower is better.

DomainSE(1000)HSMDDPMDDPM (DSI)
Energy (PJM)2.3 ± 1.34.7 ± 2.16.9 ± 2.54.4 ± 1.6
Traffic (NYC taxi)2.1 ± 0.93.2 ± 2.55.8 ± 3.33.0 ± 1.8
Healthcare (ECG5000)2.4 ± 1.37.1 ± 1.221.3 ± 3.68.2 ± 2.1
Weather (Monash)3.1 ± 1.38.6 ± 1.716.6 ± 3.28.1 ± 1.9

DSI reduces the single-checkpoint DDPM error in all four domains, with reductions ranging from approximately 36% to 62%. It also achieves lower RE than historical simulation (HSM) on three datasets. On ECG5000, where the underlying DDPM error is highest, DSI improves RE from 21.3% to 8.2% but remains above HSM's 7.1%. These experiments use the default backbone configuration without domain-specific tuning and support the applicability of checkpoint pooling to non-financial tail-risk estimation.

References

[1] Rama Cont, Mihai Cucuringu, Renyuan Xu, and Chao Zhang. “Tail-GAN: Learning to Simulate Tail Risk Scenarios”. Management Science, 2025. https://doi.org/10.1287/mnsc.2023.00936.

[2] Alexander J. McNeil, Rüdiger Frey, and Paul Embrechts. Quantitative Risk Management: Concepts, Techniques and Tools. Princeton University Press, 2015.

[3] Basel Committee on Banking Supervision. Minimum Capital Requirements for Market Risk. Bank for International Settlements, January 2019. https://www.bis.org/bcbs/publ/d457.htm.

[4] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, Kilian Q. Weinberger. “Snapshot Ensembles: Train 1, get M for free”. ICLR 2017.

[5] H. Liu, T. Zhu, N. Jia, J. He, and Z. Zheng, “Learning to Simulate from Heavy-Tailed Distribution via Diffusion Model,” Operations Research, 2026, doi: 10.1287/opre.2024.1450.

[6] Board of Governors of the Federal Reserve System and Office of the Comptroller of the Currency. Supervisory Guidance on Model Risk Management. SR Letter 11-7 / OCC Bulletin 2011-12, 2011. https://www.federalreserve.gov/boarddocs/srletters/2011/sr1107a1.pdf.

[7] K. Pandey, J. Pathak, Y. Xu, S. Mandt, M. Pritchard, A. Vahdat, and M. Mardani, “Heavy-Tailed Diffusion Models,” in International Conference on Learning Representations (ICLR), 2025.

[8] T. Karras, M. Aittala, J. Lehtinen, J. Hellsten, T. Aila, and S. Laine, “Analyzing and Improving the Training Dynamics of Diffusion Models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.

[9] S.-T. Kim, “PJM hourly energy consumption,” 2025. Available: https://dx.doi.org/10.21227/wqg6-9917.

[10] New York City Taxi and Limousine Commission, “TLC trip record data,” 2024. Available: https://www.nyc.gov/site/tlc/about/tlc-trip-record-data.page. Accessed: 2026.

[11] H. A. Dau, E. Keogh, K. Kamgar, C.-C. M. Yeh, Y. Zhu, S. Gharghabi, C. A. Ratanamahatana, Yanping, B. Hu, N. Begum, A. Bagnall, A. Mueen, G. Batista, and Hexagon-ML, “The UCR time series classification archive,” Oct. 2018. Available: https://www.cs.ucr.edu/~eamonn/time_series_data_2018/.

[12] R. Godahewa, C. Bergmeir, G. I. Webb, R. J. Hyndman, and P. Montero-Manso, “Monash time series forecasting archive,” in Neural Information Processing Systems Track on Datasets and Benchmarks, 2021.

AuthorsOpenReview ↗

Rebuttal Summary

We thank all reviewers for their time and constructive feedback, as the rebuttal period is coming to an end. We have incorporated our responses and additional evidence into the revised manuscript and marked all changes in blue. We received no follow-up from reviewers qA1J and hDVd during the discussion. We believe the additional evidence and clarifications address the concerns underlying their initial ratings and warrant a more favourable assessment.

  • For reviewer qA1J, we have addressed all concerns and questions, including the relationship to existing ensemble methods, support for the theoretical assumptions, and breadth of evaluation (see our response to qA1J and the general response).
  • For reviewer hDVd, we have addressed all concerns and questions about the contribution's originality, its relevance beyond finance, and practical use of the method (see our response to hDVd and the general response).
  • For reviewer Pbcj, following our detailed rebuttal and revisions, including additional benchmark comparisons, reviewer Pbcj raised their rating and praised the improved presentation.

DSI combines a simple method, and transparent theory to address an important mismatch in generative modelling. Standard generative objectives emphasise realism in the bulk of the distribution. However, in downstream tasks accurate tail estimates are essential because errors can have severe consequences in risk-sensitive decision making.

Simplicity is a central strength of our approach: DSI can be applied post hoc without altering training or introducing additional parametric tail assumptions that require justification and validation in regulated applications. On our benchmarks, a standard diffusion model with DSI achieves lower tail-risk error than the evaluated tail-specific baselines without sacrificing the bulk properties. This raises the question of whether their additional modelling complexity is necessary for this task.

The theoretical contribution builds on the empirical success of checkpoint ensembling by explaining when and why checkpoint reuse improves tail-risk estimates under a fixed generation budget. This provides an accessible account of both its benefits and its limitations, and a foundation for further work on checkpoint selection and ensemble inference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.