acceptodds
Withdrawn from ICLR 2027

Volatility-Clustering Adaptation for Financial Time Series

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series foundation models are increasingly adapted to new domains through fine-tuning on target data, under the implicit assumption that more target data yields better forecasts. We show that this assumption can fail in financial forecasting, where individual price changes are difficult to predict, but large moves tend to cluster, creating alternating calm and turbulent periods. Using Kronos, a foundation model trained on price bars of open, high, low, close, and volume, we argue that adapting to financial domains requires training signals beyond next-token prediction. We introduce **Volatility-Clustering Adaptation (VCA)**, which augments next-token cross-entropy with a differentiable penalty on the autocorrelation of squared returns, the standard statistical signature of volatility clustering. This additional objective provides a multi-step training signal by matching the resulting dependence structure of autoregressive rollouts to those of the realized future. Across three asset sets and two evaluation conventions, VCA delivers the best average performance and highest win rate against the pre-trained model without fine-tuning, typically improving RankIC at some cost in -MAE. Overall, our results suggest that effective financial adaptation may require objectives that capture domain-specific temporal structure beyond token-level prediction.

Primary area:
learning on time series and dynamical systems
Keywords:
financial forecasting, volatility clustering, stylized facts of returns, foundation model fine-tuning, time-series foundation models

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_WV5DOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper imposes an adaptation specific to financial markets to improve synthetic data generation. The authors augment next-token cross-entropy with a loss matching the batch-mean autocorrelation of squared returns between the generated and realized future paths. This is built on Kronos, and additional experiments are reported for Chronos-T5-small. The authors then evaluate their methodology on CSI300, CSI500m and Crypto across four horizons, using price-path RankIC and realized-variance MAE. Stronger performance among fine-tuning objectives is more consistently supported than improvement over the frozen model.

Critical strengths

The main strength of this paper is that it introduces a concrete, practical, and implementable domain-specific adaptation objective. The paper combines cross entropy with a differentiable rollout-level financial statistics rooted in one of the most salient stylized facts in financial asset returns.

Critical weaknesses

The paper suffers from some key weaknesses. The central claims exceed the reported results. Table 1 shows mixed performance for the proposed VCA when compared to alternatives. Reliable improvement over a strong frozen baseline is central to the adaptation claim in the paper, and outperforming weaker adaptation objectives is a narrower result. There is also no uncertainty across experiments reported.

The finance literature has documented multiple important stylized facts. The motivation for including volatility clustering, in its current form, is not totally clear. For example, there is also a strong “leverage effect” in which volatility changes depend on the sign of returns. This fact is as strong as volatility clustering itself, but isn’t discussed.

Questions for authors
  1. How should the claims of improvement over Frozen across datasets and horizons be reconciled with the negative and near-zero average Scores in Table 1?
  2. Under FORE, is realized variance computed from the ensemble-averaged price path or separately computed per rollout and then averaged?
  3. Do VCA’s advantages persist when CE, MSE, and VCA use a common validation criterion or each uses its own objective-aligned criterion?
  4. Was validation-based subset selection fixed before inspecting test results, and do subset benefits persist under matched optimizer-update budgets?
Reviewer_tXcnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper adapts Kronos, a financial foundation model that generates OHLCV bars as discrete tokens, by adding an auxiliary loss. The loss matches the batch-mean autocorrelation of squared returns of differentiable autoregressive rollouts to the realized ACF. Differentiability comes from a Gumbel-softmax straight-through estimator. The authors also study regime-matched fine-tuning data. They add a proposition showing that a volatility trend can change the level of squared returns without changing the distribution of the sample ACF, which explains a failure mode.

Critical strengths
  1. The core idea is clean and general. Turning a stylized fact into a differentiable rollout loss is a reusable recipe, and the Gumbel-softmax straight-through trick is the right tool for discrete token models. The controlled MSE baseline, which shares the rollout but not the ACF loss, isolates the effect of the ACF term. That is a good ablation design.

  2. The paper is candid about failure. Proposition 1 identifies a real hole in the objective: the ACF of squared returns is invariant to the direction of a volatility trend, so the loss can be satisfied while the forecast variance drifts. The authors confirm this in simulation and in the rollouts, and they report the Chronos blow-up rather than hiding it. This kind of analysis is more valuable than another small benchmark gain.

Critical weaknesses
  1. The gains do not support the claims. The only consistent improvement is a small σ²-MAE gain on Crypto under FORE. On CSI300 and CSI500 the average Score is at or below zero. Under VOL, every VCA score is negative, so the "two evaluation conventions" claim fails.

  2. The regime-matching study uses test-period volatility. Table 2 lists test/first and test/last ratios and uses them to pick the matched slice. The paper calls this a diagnostic, but the abstract and contribution list present "volatility-matched data can match or outperform full-data fine-tuning" as a finding. Calling it a diagnostic does not fix the leak; it only labels it.

Questions for authors
  1. The abstract says VCA improves over the pre-trained model under two conventions. Under VOL all three average Scores are negative. Can you restate the claim, or explain why the VOL convention should be discounted?

  2. With K=1, the loss matches only lag-1 ACF. Why call this "volatility clustering" rather than "lag-1 squared-return autocorrelation"? What happens with K=15 at H=48?

  3. The batch-mean ACF can be matched by a mixture of high- and low-clustering paths. Did you check the per-path ACF distribution, or only the batch mean?

  4. The ablation shows K=2 drops RankIC from 0.328 to 0.048 on CSI300 at H=16, and λ=20 drops it to 0.056. Is this a hyperparameter cliff? Were the defaults selected on the test benchmarks?

Reviewer_TF2EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VCA adds a squared-return autocorrelation loss to cross-entropy when adapting a financial time-series model through differentiable rollouts. The paper also studies regime matching and a scale-related blind spot, but its gains over the frozen model are mixed across datasets and metrics.

Critical strengths
  1. A relevant path-level objective. Matching volatility dependence gives the model an explicit training signal for a property of the generated path. This is a sensible addition to teacher-forced token prediction.

  2. A useful rollout control. The MSE baseline uses the same differentiable autoregressive rollout as VCA. This helps distinguish the choice of statistic from the effect of rollout-based training itself.

  3. Informative analysis of a failure mode. The paper shows that matching autocorrelation does not determine the overall volatility level or the direction of a volatility trend. This is a useful qualification of the proposed objective.The improvement claim is too broad. The average Score relative to Frozen is negative on CSI300 and approximately zero on CSI500. Being the strongest fine-tuned method does not establish consistent improvement over leaving the pretrained model unchanged.

Checkpoint selection favours the proposed statistic. The rollout-MSE baseline selects checkpoints using validation ACF gap rather than its own objective or a common downstream criterion. This leaves the advantage over a properly selected alternative unresolved.

Risk-level accuracy remains a concern. The objective leaves volatility level unconstrained, and the paper reports unstable results on the additional backbone. All fine-tuned methods also trail the classical baselines on QLIKE, which limits the broader risk-forecasting claim.

Critical weaknesses

W1. The improvement claim is too broad. The average Score relative to Frozen is negative on CSI300 and approximately zero on CSI500. Being the strongest fine-tuned method does not establish consistent improvement over leaving the pretrained model unchanged.

W2. Checkpoint selection favours the proposed statistic. The rollout-MSE baseline selects checkpoints using validation ACF gap rather than its own objective or a common downstream criterion. This leaves the advantage over a properly selected alternative unresolved.

W3. Risk-level accuracy remains a concern. The objective leaves volatility level unconstrained, and the paper reports unstable results on the additional backbone. All fine-tuned methods also trail the classical baselines on QLIKE, which limits the broader risk-forecasting claim.

Questions for authors

Q1: Reconciliation of the main claim and results. Can the authors reconcile the stated improvement over the pretrained model with the negative or near-zero dataset averages in Table 1? The main conclusion should identify precisely where the method helps.

Q2: A fair checkpoint comparison. Does VCA retain its advantage when each method is selected on its own objective and, separately, on a common downstream validation metric? This would determine whether the result comes from the training objective or the selection rule.

Q3: Treatment of unstable runs. Were all training seeds, including variance-inflated runs, included in the aggregate results? Please state the rule for any failed or excluded runs, since this affects the reliability of the reported gains.

Comments on the paper

1 comment

Withdrawal and corrections

Thank you to all three reviewers. Your questions made us go back, check our evaluation and retrain every method. In doing so, we found problems that change the main conclusions, so we are withdrawing the paper. We summarize them here so that readers are not misled by the submitted numbers.

1. RankIC protocol.

Our RankIC computation is not consistent with the Kronos protocol. In test windows where a symbol has missing bars, our script imputed RankIC as 1 instead of excluding the window. This applies equally to all methods, including the frozen model, but it inflates the absolute RankIC values. Once these windows are excluded, VCA is still the best of the fine-tuning objectives, but it no longer beats the frozen model on average.

2. Data efficiency.

The small data slices were trained for the same number of epochs as full data, which means far fewer training steps. With a matched number of steps, the advantage of the small slices over full data is not robust and the data-efficiency claim only holds under a small training budget.

3. Checkpoint selection.

The paper does not compare with CE, MSE when selecting checkpoint by their own objective and common downstream validation metrics (TF2E, Q2).

What still holds.

Cross-entropy fine-tuning hurts the frozen model on all three datasets. With K=1, the objective targets lag-1 autocorrelation of squared returns, not volatility clustering in general (tXcn, Q2). The leverage effect exists in our data but is much weaker than clustering (WV5D, W2).

A revised version will use the corrected evaluation throughout and report all runs and their variance. Thanks again to the reviewers; these corrections came directly from their concerns.

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

Related papers

Discussion (0)

Sign in to comment.