acceptodds
Under review as a conference paper at ICLR 2027

Synthetic to SPX: Model-Based Pre-Training for Diffusion Models of Financial Time Series

Abstract

Diffusion models have achieved remarkable success in image, video, and text generation. We apply them to financial time series. The main obstacle is the scarcity of data: only one realized price trajectory is available for each asset. Finance has, however, a major advantage: fifty years of mathematical models that capture the main mechanisms of markets and can be simulated at will. These models provide unlimited training data, while their known statistical properties enable controlled evaluation of generated trajectories. We therefore propose model-based pre-training: first, train the score network on parametric models of increasing complexity; second, fine-tune it on the single real trajectory. In the first phase, we compare two approaches: one learns the joint path distribution in , the other factorizes it into conditional distributions over . The autoregressive formulation is the most data-efficient, reaching a given fidelity with up to four times fewer trajectories; with abundant data, all architectures perform alike. We then fine-tune on the overlapping windows of a single trajectory, first on a path-dependent volatility model calibrated on SPX and held out from pre-training, where the ground truth is known, and then on SPX itself. Pre-training improves adaptation in this low-data regime, and what matters is the proximity of the source models to the target rather than their diversity. On SPX, the pre-trained model recovers volatility clustering, heavy tails and the leverage effect, which training from scratch misses.

Primary area:
learning on time series and dynamical systems
Keywords:
diffusion models, score-based generative models, generative modeling, financial time series, transfer learning, pre-training, fine-tuning, autoregressive models, synthetic data, memorization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_n7eQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper tackles the problem of learning generative financial time-series models from limited, temporally dependent historical observations. The authors pre-train diffusions models on SPX-calibrated financial models, then fine-tune on SPX itself. The paper evaluates the pipeline on a path-dependent volatility model calibrated to the S&P to attempt to recover known model parameters. The main finding is that autoregressive models use four times fewer training trajectories than the U-Net and twice fewer than the diffusion transformer. The pre-trained models replicate key stylized facts of financial data.

Critical strengths

One strength of this paper is its ability to offer interpretable evidence about learning and transfer in financial data. Given the low-signal environment in financial data, it is difficult to evaluate the merit of synthetic data. This paper gets around this issue by evaluating against a known model from which the pre-training data are generated. This also allows the separation of fidelity and memorization diagnostics.

Another strength is the substantial comparisons offered for the pre-training source ablation. The comparisons include simple priors, individual financial models, mixtures, and an oracle. The authors then show that structured volatility sources outperform simple sources, and selected mixture outperforms the reported full mixture. This offers practical insight into constructing synthetic pre-training distributions.

Critical weaknesses

Perhaps the biggest issue is that there is limited evidence of usefulness of the synthetic data. The synthetic data are important only if they are somehow useful. This is especially important for financial data where there is no ground truth to evaluate the synthetic data.

The main conclusions also rely on an incomplete fidelity metric, as DMLED_{MLE} compares marginal distributions of fitted parameters rather than the full path. This metric does not directly capture parameter dependence or discrepancies discarded by parameter fitting. That is, the architecture and source rankings may not establish broader distributional fidelity.

Questions for authors
  1. How do fine-tuned models compare with pre-training only models and direct calibrated simulations on held-out SPX data?
  2. do the main rankings persist under complementary path-level metrics, and what is the test-versus-test DMLED_{MLE} baseline?
  3. Which historical observations inform simulator calibration, fine-tuning, validation, and evaluation?
  4. Which families constitute each Table 4 mixture, especially “full mixture?” How are they combined?
Reviewer_XcCGOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes model-based pre-training for diffusion models of financial time series. The central motivation is that financial markets provide only a single realized historical trajectory for each asset, making direct training of high-capacity generative models difficult. The authors exploit classical parametric financial models, including GBM, GJR-GARCH, Heston, Bergomi, Bates, and PDV, which can generate unlimited synthetic trajectories. A diffusion model is first pre-trained on trajectories simulated from these models and subsequently fine-tuned on limited target data.

The paper compares three diffusion formulations: a path-space U-Net, a path-space Diffusion Transformer (DiT), and an autoregressive diffusion model with an RNN history encoder. The autoregressive formulation is reported to be more sample-efficient in low-data settings, while the architectures become similar with sufficiently large synthetic training sets.

The authors first evaluate transfer in a controlled setting using PDV as a held-out target model, allowing access to its known data-generating distribution. They then fine-tune on overlapping historical SPX windows and evaluate generated trajectories using financial stylized facts such as volatility clustering, skewness, kurtosis, and leverage. The results suggest that financially relevant pre-training distributions are more useful than simply increasing the diversity of synthetic source models.

Critical strengths
  1. Interesting application of domain-specific synthetic pre-training to the challenging low-data setting of financial time-series generation.
  2. The held-out PDV experiment provides a useful controlled environment before moving to real SPX data.
  3. Careful analysis of data efficiency and memorization across three diffusion architectures.
  4. The source-model ablation provides an interesting finding that relevance/proximity of synthetic dynamics may matter more than source diversity.
Critical weaknesses
  1. The paper provides a reasonable motivation for model-based synthetic pre-training, but does not sufficiently explain why diffusion models are particularly appropriate for financial time-series generation. The work largely takes diffusion as a given rather than establishing its advantages for this setting.

  2. The central approach consists of pre-training on domain-specific synthetic data generated by parametric financial models and subsequently fine-tuning on limited target data. While the application to financial time series is interesting, the overall pre-training/fine-tuning strategy is conceptually straightforward, and the paper does not clearly articulate what the main methodological novelty is beyond this application.

  3. The presentation and organization make the paper very difficult to follow.

Questions for authors

Please see weaknesses

Justification for rating:

The paper is very difficult to follow. Method doesnt seem well motivated, sections dont connect well, and the contributions of the paper, and the findings for the community are difficult to figure out.

Reviewer_f6SLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies synthetic pre-training for diffusion models of financial time series. The motivation is that only a single historical realization is available for a given financial asset, whereas classical stochastic models of financial dynamics can generate arbitrarily many synthetic trajectories. The proposed pipeline therefore calibrates a collection of existing parametric financial models and uses trajectories simulated from these models to pre-train a diffusion model before fine-tuning it on limited real data.

The paper considers two diffusion formulations. The first models a joint trajectory using either a U-Net or a diffusion transformer. The second factorizes the trajectory autoregressively and uses an RNN to encode the history while a conditional diffusion model generates each subsequent return.

The main empirical conclusions are that the autoregressive model is more sample-efficient than the path-space architectures, that pre-training on financially structured synthetic data improves adaptation in the low-data regime.

Critical strengths

See the summary.

Critical weaknesses
  1. The paper spends much space explaining existing literature and little space isolating precisely what is new. In particular, Section 2 describes standard diffusion and autoregressive diffusion, both of which are well established in the literature, and Section 3 describes financial models that are well established as well. The idea of using synthetic data to pre-train the model is also well established across many domains, including time-series modeling, and is known as synthetic-to-real pre-training. The paper itself acknowledges prior synthetic pre-training approaches such as ForecastPFN and Chronos, as well as prior diffusion models for financial time series. I therefore failed to recognize the novelty of the work.
  2. The experiments compare U-Net, DiT, and AR(RNN) implementations within the proposed framework, but do not compare against existing financial or time-series generative baselines. It is therefore hard to evaluate the performance of the proposed method.
  3. The forecasting example in Figure 5 is also difficult to interpret as positive evidence. The realized COVID path exhibits a drawdown and volatility that are more extreme than almost all generated trajectories, and visually the realized March 2020 decline lies far outside the model's predictive envelope. A pandemic need not be predictable from the conditioning history, but this example appears to illustrate the model's inability to represent the observed tail rather than validate its quality. This again shows the need for baseline comparisons.
  4. Experiments are conducted on a single univariate SPX index. As a result, the proposed framework does not model or evaluate any cross-sectional dependence, such as correlations, sector structure, or tail dependence across assets. This restriction weakens the paper's data-scarcity motivation, since the study does not investigate whether information from multiple related assets can be exploited. The authors acknowledge this limitation.
Questions for authors
  1. In the abstract, the authors claim that "pre-training improves adaptation in this low-data regime, and what matters is the proximity of the source models to the target rather than their diversity." Can you elaborate on the proximity notion? Do you define it in the manuscript? Why does it matter, and which results support this claim?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.