acceptodds
Under review as a conference paper at ICLR 2027

Constraint-Driven Measure Transformation of Pretrained Diffusion Models

Abstract

Adapting a pretrained diffusion model to a target known only through expectation constraints requires solving for the tilt that reaches it. We solve for it by minimum relative entropy against the learned prior, giving the weighted Monte Carlo measure, and carry it into the frozen sampler through the Doob -transform. Both are prior work. Because the reference is learned, it can be varied, and the construction becomes an instrument for a question central to derivative pricing: what does a finite set of market constraints determine about a path-dependent price? On Heston paths, 53 constraints span 97% of an Asian option's payoff variance but only 76% of a barrier's, and that unspanned variance sets, to within 2.4%, how fast the admissible price range opens with relative-entropy budget. On a real AAPL option surface, every prior we test (learned, parametric and Black–Scholes alike) misprices around a scheduled earnings date in the same direction, and the pipeline recovers the known earnings-jump signature from the quotes. Carrying the tilt into the sampler, which the identification results do not require, works only at mild tilts. The learned correction is accurate as a function, but the sampler does not realise the measure it implies.

Primary area:
generative models
Keywords:
diffusion models, change of measure, Doob h-transform, inference-time adaptation, risk-neutral pricing, quantitative finance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mgn4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies adapting a pretrained diffusion model for financial time-series generation to satisfy constraints on sample averages of specified quantities. It first reweights samples from the diffusion model using importance sampling, then learns a Doob h-transform correction that guides the frozen diffusion sampler toward the resulting target distribution. On synthetic data, the corrected sampler is less accurate than retraining on reweighted samples but requires less adaptation time.

Critical strengths

The financial time-series application is interesting. Instead of steering individual samples with a reward function, this paper provides a useful setting for studying distribution-level calibration and whether a corrected diffusion sampler preserves the required averages.

Critical weaknesses

The readability of the paper is not good. Could the author provide a human readable text explain the contribution matters for the paper? It is very hard for the reviewer to evaluate claims like this "Finite constraints span most of an Asian’s payoff but leave a quarter of a barrier’s unspanned."

Something like we developed a method for adapting diffusion based models for a set of numerical constraints on the expectation of samples. On this financial dataset, we show that …

The proposed approach combines existed entropy-based reweighing and diffusion model guidance but the technical contribution remains unclear for novelty. The method section is even more uncomfortable for the reader, "We call the construction Score Measure Transformation (SMT) and claim nothing for the mechanism, which is the Doob h-transform." what is the point introducing this new term, to make it sound novel?

The real-data experiment also evaluates only reweighted priors, with no clear advantage over the student‑t baseline and no evaluation of the learned correction.

Questions for authors
  • What practical value does the learned corrected sampler bring? How is that being useful for real-world financial application given you can reweight the samples and your are not constraint by timing.

  • Could the author show that the corrected sampler reproduce the calibrated distribution on real data, i.e. the APPL case?

  • Why does the learned correction become less accurate when the target distribution moves further from the prior? Is this because of the importance sampling? Or because the correction itself becomes harder to learn?

Reviewer_tXsEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper uses a path-valued DDPM trained on returns as the reference measure for entropy projection onto market constraints (weighted Monte Carlo). It measures how much of an exotic's price the constraints determine via a span / residual-variance decomposition, compares priors on an AAPL surface, and tests amortizing the tilt via a Doob h-transform. The claimed contribution is that a learned prior "earns its place" where no correctly specified model exists. absent.

Critical strengths

The paper is careful and candid: falsified predictions are reported and limitations are stated. The span decomposition is a clean way to frame identification.

Critical weaknesses

1. The AAPL prior may be memorizing, which would remove any reason to use a diffusion.

  • The prior has about 580k parameters and is trained for 450 epochs on about 2,819 stride-1 overlapping 21-day blocks (only about 135 non-overlapping months), with no validation split.
  • Its largest generated daily return (0.146) sits almost exactly on the data's largest day (0.143).
  • The authors themselves attribute its edge to having "seen many earnings windows."

The literature predicts memorization in this regime. The denoising score-matching objective has a closed-form optimum that only replicates training data, and memorization tends to occur on smaller datasets (Gu et al., TMLR 2025). Memorization occurs when the number of samples is small at fixed parameter count (Bonnaire et al., 2025). With an optimal score, backward trajectories collapse onto memorized training points (Biroli et al., Nature Communications 2024).

A diffusion that learns the empirical score samples training blocks, or a Gaussian KDE around them. It is then just an expensive block bootstrap or KDE, with no advantage over classical nonparametric methods.

2. The baselines are weak, and the relevant literature is missing.

The comparison uses an i.i.d. Student-t, Black-Scholes, and a Heston fitted by method of moments. The Student-t, despite generating 241% daily moves, ties the diffusion. Entropy projection of historical returns is established as canonical valuation: Stutzer (1996, Journal of Finance), and Gray & Newman (2005), who found that under stochastic volatility canonical valuation outperforms Black-Scholes. Neither is cited.

Missing baselines, each cheap to project onto the same constraints:

  • historical blocks or a block bootstrap (Politis & Romano, 1994)
  • a Gaussian KDE on blocks
  • GARCH / GJR-GARCH (Duan, 1995)
  • filtered historical simulation (Barone-Adesi, Engle & Mancini, 2008)
  • a GARCH with an earnings-day variance term
Questions for authors
  1. Memorization diagnostics: nearest-neighbour distances from generated paths to training blocks, compared with a held-out time period.
  2. The baselines listed above, projected onto identical constraints.
Reviewer_rPDjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how to adapt a frozen diffusion model to a distribution specified by expectation constraints. It combines entropy-based calibration with a learned Doob h-transform. Financial experiments examine calibration accuracy, sensitivity to the prior, and the accuracy of the corrected sampler.

Critical strengths
  1. The overall two-step approach:
  • a. Find minimum-relative-entropy weights that satisfy the constraints.
  • b. Learn a Doob h-transform correction so the frozen generator can sample from the reweighted distribution. is natural and reasonable.
  1. The financial applications are interesting and important.
Critical weaknesses
  1. Dynamic no-arbitrage is not enforced. In the specific application of generating option dybamics, the most important constrant is dybamic no-aribitrage constratin, which is expressed in terms of meausure-based constraint. The finite moment tests do not guarantee the conditional martingale property. The calibrated measure need not be risk-neutral, and the reported ranges are not no-arbitrage bounds. This limits the actual applications of the proposed method.

  2. The methodological novelty is limited. The construction combines standard entropy calibration with Doob h-transform guidance. The paper does not clearly establish what is new beyond applying these tools to a learned prior.

  3. The experiments lack generality. The financial evaluation centers on a Heston setup with 21 daily steps and one AAPL option surface from a single date. This does not establish robustness across assets, longer horizons, or different market conditions. The real-data comparisons also lack confidence intervals, so the observed differences do not establish a reliable advantage over simpler baselines

Questions for authors
  1. Can richer martingale tests prevent the increase in held-out martingale errors when option-price constraints are added?

  2. Can the authors separate gradient-estimation errors from errors caused by the guided sampler visiting states different from its training inputs?

  3. Are there other important applications?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.