acceptodds
Under review as a conference paper at ICLR 2027

DelayCast: Event-Gated Low-Rank Residual Adaptation for Time-Series Foundation Models

Abstract

Time-series foundation models achieve strong forecasting performance from historical numerical observations, while external event notifications may provide additional information about future changes. Existing approaches incorporate external information through numerical covariates or multimodal fusion, yet effectively utilizing sparsely released textual notifications remains challenging when their availability and subsequent effects are temporally uncertain. To address this challenge, we propose DelayCast, an event-gated low-rank residual adaptation framework that incorporates structured information extracted from event notifications into frozen foundation-model forecasts. DelayCast combines historical numerical states with cutoff-available event attributes to generate constrained multi-horizon residual corrections, while preserving the original backbone forecast when no eligible event information is available. On the SpaceWeather dataset, DelayCast reduces the 72-hour forecasting mean absolute error of frozen Chronos-2 by 8.97% on the 2025 evaluation and 3.50% on a partial-year 2026 cohort. Beyond forecasting performance, we employ hierarchical comparisons to distinguish numerical calibration from the incremental contributions of event information, and further investigate the gap between the predictive value of retrospectively known response timing and the ability to recover such temporal structure from sparse event supervision.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Time-Series Foundation Models, Event-Conditioned Forecasting, Low-Rank Residual Adaptation, Structured Event Information, Multi-Horizon Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_ykzMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to improve time-series foundation model forecasts by leveraging sparse external event notifications characterized by uncertain timing. The proposed approach keeps the core forecasting backbone frozen and learns to predict the residual—integrating historical numerical states with event attributes available at the time of prediction—thereby bypassing the need to explicitly estimate the onset and duration of event-driven responses. We introduce "DelayCast," a method that generates forecast corrections through deterministic event gating, low-rank cosine basis representations, and constraints on residual magnitude, while preserving the original forecasts in the absence of relevant events.

Critical strengths
  1. Focusing on the integration of information regarding irregular external events with foundational time-series models holds significant practical research value.
  2. The use of lightweight residual adaptation ensures that, in the absence of available event data, the predictive performance of the backbone model is precisely preserved.
  3. Implementing a hierarchical comparison between numerical calibration and event-based information helps to discern the specific sources of predictive gains.
Critical weaknesses
  1. The sparsity, asynchronous nature, and temporal uncertainty of event notifications do not directly necessitate a DCT-based low-rank residual design; there is a lack of clear motivation for why this specific design is well-suited to address these challenges.

  2. Existing comparisons primarily focus on validating the effectiveness of adding an adapter, failing to provide a robust benchmark against covariate-based prediction methods, which leverage the same event attributes, or simple residual regression.

  3. The experiments lack a thorough comparison between predicting the full residual directly and using a DCT-based low-rank approximation; furthermore, the individual contributions of the shrinkage and magnitude constraints have not been isolated or verified.

  4. The attribution of gains across different layers is primarily based on the development set (which was used for hyperparameter tuning), and there is a lack of comprehensive control experiments for the independent test years; consequently, the observed year-over-year performance gains may largely stem from numerical calibration.

  5. The inherent difficulty in accurately predicting the onset time of a response does not imply that explicit response modeling cannot improve final prediction accuracy. The absence of an end-to-end prediction baseline leaves the experimental justification for adopting a direct residual adapter insufficient.

Questions for authors

see weaknesses

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

The paper presents DelayCast, which augments a frozen Chronos-2 model with a constrained residual correction. Using numerical-state features and structured attributes from event notifications available at forecast time, a Ridge regression predicts correction coefficients in a truncated DCT basis. A deterministic availability gate ensures that DelayCast reproduces the base model’s predictions exactly when no eligible event information is available. The method is evaluated on geomagnetic forecasting at 48-, 72-, and 96-hour horizons, with encouraging improvements over Chronos-2, particularly on the 2025 evaluation set.

Critical strengths
  • DelayCast represents the residual correction using only K=4K=4 DCT coefficients and limits its magnitude through shrinkage and a smooth cap. These constraints reduce the adapter’s flexibility and may help prevent overfitting when event data are scarce.
  • DelayCast uses only event notifications available at forecast time. Its deterministic gate disables the correction when no eligible event is available, preserving Chronos-2’s original prediction exactly.
  • Encouraging 2025 results and chronological evaluation across multiple years.
Critical weaknesses
  • Comparing DelayCast with Chronos-2 using the same event information and with simpler residual corrections would help show whether its specific design provides additional benefits.
  • DelayCast is evaluated only on SpaceWeather with Chronos-2. It remains unclear whether the gains would extend to other domains, where event frequency and effects may differ, or to other forecasting backbones. Testing additional datasets and models would strengthen the applicability of the approach.
Questions for authors
  • Would a learned correction weight, combined with the deterministic availability gate in Equation (2), improve performance over applying the correction to every eligible event?
  • Does training include windows without events? How exactly do Numeric and Presence differ in their inputs and gating?
  • How much improvement comes specifically from the DCT representation, shrinkage, and amplitude constraint?
Reviewer_pxo9OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DelayCast, an event-gated low-rank residual adaptation framework for TSF foundation models. The authors aim to address the challenges of incorporating sparse, asynchronous external event notifications with uncertain temporal effects into time series forecasting. Specifically, DelayCast combines historical numerical features with structured event attributes to predict low-dimensional DCT coefficients, which are then used to reconstruct residual corrections for the frozen Chronos-2 backbone. A deterministic event gate ensures that the original backbone forecast is preserved when no eligible event information is available. Experiments demonstrate TSF accuracy improvements across different evaluation periods.

Critical strengths
  1. The paper addresses a practically relevant challenge in time-series forecasting: incorporating sparse and asynchronous external event information whose effects may be delayed or temporally uncertain.

  2. DelayCast provides a computationally efficient adaptation strategy without modifying the pretrained foundation model. Its low-rank residual correction and deterministic event gate make the framework relatively simple and easy to implement.

  3. The authors conduct cross-year evaluations, hierarchical ablation studies, and event-response timing analyses to investigate the sources of forecasting improvements. These analyses provide useful insights into the potential value and limitations of external event information.

Critical weaknesses
  1. The paper lacks an illustrative figure in the Introduction to clearly motivate the research problem, making it difficult for readers to understand the specific challenges the authors aim to address. Moreover, the overall framework diagram (Figure 1) is overly cluttered. The actual methodological contributions, such as the Event Gate and Low-Rank Residual Adaptation, are buried among numerous small boxes that provide little meaningful information. The figure gives the impression of being entirely AI-generated without sufficient manual inspection or refinement. However, the authors explicitly state in their AI-use declaration that generative AI was used solely for language polishing. This raises serious concerns about the accuracy and completeness of their AI-use statement.
  2. The comparisons with competitive models in the main experiments are severely insufficient. Throughout the paper, the proposed DelayCast is essentially compared against only a single frozen Chronos-2 model. Although the authors extensively discuss time-series foundation models such as TimesFM and Moirai in the Related Work section, none of these models are included in the main experimental comparisons. The current experimental results are insufficient to demonstrate the superiority of DelayCast over existing foundation-model adaptation methods.
  3. The experiments are conducted on only a single SpaceWeather dataset. Although the authors divide the evaluation into three periods, different temporal splits of the same dataset should not be considered distinct benchmark datasets. Consequently, the paper fails to demonstrate the generalization capability of the proposed method across other time-series forecasting scenarios influenced by external events, such as finance, energy, and transportation.
  4. The proposed low-rank residual adaptation predicts only a small number of DCT coefficients rather than directly modeling the full residual trajectory. This design may limit the model's ability to capture fine-grained variations and abrupt changes. Moreover, the authors do not appear to introduce any specific mechanism to compensate for this loss of expressive capacity. I encourage the authors to further justify this design choice and investigate whether the restricted residual representation limits forecasting performance.
Questions for authors
  1. Please provide a justification for the absence of background images, propose a solution, and clarify the primary focus of Figure 1.
  2. Include additional results from the main experiments, specifically incorporating a competition model and benchmark datasets.
  3. Clarify how Low-Rank Residual Adaptation effectively models complex trends and abrupt changes.
Reviewer_bDYJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DelayCast, a lightweight adapter that uses structured event attributes to correct forecasts from a frozen time-series foundation model. A Ridge regressor predicts low-rank residual corrections, which are applied only when an eligible event notification is available. Experiments on SpaceWeather examine forecasting performance, the contribution of event features, and the recoverability of event-response timing.

Critical strengths
  1. The paper addresses a relevant forecasting setting involving sparse event notifications and uncertain response times.
  2. The adapter is lightweight and preserves the original backbone forecast when no eligible event is available.
  3. The hierarchical comparisons and feature ablations help distinguish numerical calibration from the potential contribution of event information.
  4. The authors report episode-clustered confidence intervals and acknowledge uncertainty in the incremental gains from event features.
Critical weaknesses
  1. The evaluation uses one dataset and one foundation-model backbone. It is unclear whether the approach transfers to other domains, event types, or forecasting models.

  2. The experiments mainly compare variants of the proposed adapter with frozen Chronos-2. Comparisons with other lightweight adaptation and covariate-aware forecasting methods would better establish the value of the approach.

  3. On the development folds, structured attributes improve MAE by 0.58% over the presence-only model, with a 95% confidence interval of [−0.37%, 1.49%]. The evidence for an additional benefit from structured event information is therefore limited.

  4. Event information has previously been used in forecasting, and the method combines established components: DCT basis expansion, Ridge regression, residual correction, and a deterministic gate. The practical combination is reasonable, but its contribution relative to existing adaptation methods needs clearer justification.

  5. The paper motivates a low-rank, smooth residual under sparse supervision, but does not show whether the DCT basis is necessary. A direct full-horizon residual predictor and comparisons with other low-dimensional bases would clarify whether the benefit comes from DCT specifically or from restricting the residual’s degrees of freedom. The effect of the number of basis functions and the ability to capture abrupt responses also deserve further analysis.

  6. The 2026 improvement has a confidence interval that includes zero, and the retrospective numerical data do not provide a fully archived point-in-time replay. The hindsight oracle uses realized response information, so its gain does not necessarily isolate the value of timing alone.

Questions for authors

How does it compare with existing lightweight forecast adaptation methods and models that directly incorporate event covariates?

How does performance change when the DCT basis is removed and the adapter directly predicts the full-horizon residual? Does DCT outperform other low-dimensional or smooth residual parameterizations?

How sensitive are the results to the number of DCT components, particularly for abrupt or short-lived responses?

How is the event-window eligibility rule determined using only information available at the forecasting cutoff?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.