acceptodds
Under review as a conference paper at ICLR 2027

SAEF: State-Aware Extreme Event Forecasting for Multivariate Time Series

Abstract

Extreme values are often treated as isolated tail observations, although the dynamics leading to an extreme event may change substantially before the event occurs. However, existing extreme-event forecasting methods mainly rely on rare-sample modeling and tail-supervision optimization, making it difficult to characterize system state transitions and evolving inter-variable dependency structures before extreme events occur. To address this issue, we propose SAEF, a state-aware enhanced framework for extreme-event forecasting. The framework combines local stochastic-dynamics modeling with critical slowing down theory to detect precursor states before extreme values appear, and injects the resulting risk-state information into subsequent forecasting. Furthermore, a Lyapunov consistency constraint converts the second-order statistical reference of local stochastic dynamics into a differentiable training constraint, while a unified risk-state variable is used to regulate dependency structures and match similar extreme-event patterns, thereby enabling state-conditioned forecasting. Experiments show that SAEF consistently improves extreme-event F1 and extreme-point MAE while maintaining competitive overall forecasting performance. The code is available at https://anonymous.4open.science/status/SAEEF-2026.

Primary area:
learning on time series and dynamical systems
Keywords:
Extreme Event Forecasting, Early Warning Systems, State-Aware Modeling, Dynamic Graph Learning, Prototype Learning

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_mjuDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAEEF, a state-aware enhancement framework for multivariate extreme-event forecasting. It combines input-window statistical descriptors and a dynamical risk prior into a shared risk variable, which conditions interaction graphs through SCIG and extreme-event prototype matching through SCPM. Training includes point forecasting, extreme-event supervision, prototype losses, and a Lyapunov consistency term. Experiments on 5G network data and standard forecasting benchmarks report competitive overall accuracy and improvements in extreme-point errors and event-detection metrics.

Critical strengths

S1. A coherent integration of risk-state estimation into forecasting. Rather than using risk only as a post-hoc detection score, SAEEF uses the same risk variable to adjust cross-variable interactions and extreme-event prototype matching. Retaining the base predictor while adding a small state-conditioned correction provides a concrete and coherent way to incorporate risk information into forecasting. This integration is a meaningful aspect of the proposed framework.

S2. Promising improvements on metrics relevant to extreme-event forecasting. The reported gains extend beyond overall prediction error. For example, on ETTh1 with H=96 and q=0.98, the full model reduces extreme-point MAE from 4.237 to 3.684 and increases event F1 from 0.433 to 0.568; AP also improves from 0.504 to 0.661. These task-specific improvements support the potential practical value of the framework, subject to clarification of the mechanism concerns below.

Critical weaknesses
  1. Pre-peak warning does not establish detection before event onset. The reported lead time is measured relative to the event peak, so an alert raised after an extreme event has already started can still count as early. This does not directly support the central claim of detecting precursor states before extreme values appear. Onset-based evaluation and comparisons with simple risk baselines at matched false-alarm rates are needed.

  2. The Lyapunov ablation results require reconciliation. On ETTh1, Table 5 reports event F1 of 0.552 for the configuration preceding the addition of the Lyapunov loss, whereas Table 10 reports 0.418 for “w/o Llyap\mathcal L_{\mathrm{lyap}}.” These appear to represent comparable configurations, but no difference is specified. This discrepancy makes the contribution of a central component difficult to assess and should be resolved with explicit experimental settings.

  3. Comparisons with specialized extreme-event forecasting methods are limited. The direct comparison with NEC+, DAN, SR-MTL, ReFine, and PFformer is restricted to ETTh1 with H=96H=96 and q=0.95q=0.95. The broader overall-forecasting comparisons and internal ablations do not establish superiority over these specialized methods across domains or rarer thresholds. Extending the matched comparison to the main 5G setting and higher quantiles would strengthen the evidence for the proposed approach.

Questions for authors

See the weaknesses. Maybe I will improve my assessment if authors can clarifying these issues well.

Reviewer_um5DOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

SAE²F adds several modules to a forecasting backbone: a risk prior built from window statistics, a risk variable α, an α-weighted graph residual (SCIG), α-conditioned prototype matching (SCPM), tail losses, and a "Lyapunov consistency" loss motivated by critical slowing down. The paper reports the best average MSE/MAE on 9 of 10 datasets. It also reports better extreme-event metrics than NEC+, DAN and others on ETTh1.

Critical strengths

The problem is relevant. The paper also reports event-level metrics (AUROC/AP, lead time, false alarms and calibration), which this area needs more of.

Critical weaknesses

One finding drives my recommendation: in the released code, one of the three claimed contributions cannot affect the model.

A parameter-free module (state_features.py) computes σ and ρ from the raw input. The Lyapunov loss compares σ²(1-ρ²) with a scalar, lyapunov_q (saeef_model.py), that nothing else uses. Validation uses only the point loss, and event F1 is computed from the forecast windows. So this loss cannot change the forecasts, α or the reported metrics. Yet Table 5 credits it with +0.016 F1. Table 10 shows F1 falling to 0.418 on ETTh1 without it, a larger drop than removing SCIG or SCPM. Table 10 also contradicts itself. In the progressive path, "+α_prior" is the full model minus L_lyap, and it scores 0.341 / 0.552 / 0.556 (RRC-C1 / ETTh1 / Weather). "w/o L_lyap" scores 0.268 / 0.418 / 0.472, while the seed spread in Table 11 is about ±0.006. Contribution (2) (L93-94) is therefore not supported, and I cannot trust the ablations until this is explained.

Other issues that also matter:

  • The forecasting gains come from a backbone that the paper does not attribute. The backbone code uses DUET's modules (Linear_extractor_cluster, Mahalanobis_mask), but the paper does not say so, and DUET is not in Tables 1 or 7. SAE²F-base alone reaches 0.357 MSE on ETTh1-96, vs 0.354 for the full model, which is already lower than every baseline (the best is 0.367).
  • The critical-slowing-down story is untested. α is supervised mainly through a BCE on extremes in the next H steps (App. E.8). So lead times of 56 and 74.5 steps at H = 96 are expected and do not show critical slowing down.
  • Table 2 covers one dataset, one horizon and one threshold, with no variance.
Questions for authors
  1. How can L_lyap change the predictions or the metrics in the released code? Please rerun "w/o L_lyap" and "+α_prior" with several seeds.
  2. Is the backbone DUET? If so, please state it and add DUET to Tables 1 and 7 under the same look-back protocol. The ETTh1 H = 96 script uses seq_len 512. Please also give each baseline's protocol.
  3. Which score produces the AUROC/AP in Table 17 and the results in Table 4? On which split is τ_ext computed (L137-138)?
  4. What evidence supports critical slowing down? For example, compare α_prior alone with the context score alone, and test for variance or AR(1) trends before events vs control windows.
  5. Are the RRC datasets available?
Reviewer_uAvTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SAEF, a framework for forecasting multivariate time series around upper-tail events. Input-window statistics motivated by local stochastic dynamics feed a risk variable, alpha, which controls cross-variable interaction graphs and extreme-event prototype matching. Training combines forecasting and event-specific objectives with a Lyapunov consistency term. The evaluation covers two mobile-network datasets and eight standard forecasting datasets, with additional analyses of event detection, component contributions, computational overhead, and early warning.

Critical strengths

The shared risk variable gives state estimation a concrete role in forecasting. It changes both the dependency structure and the prototypes used for prediction, making the design more meaningful than adding a separate event classifier. The controlled alpha scan in Table 14 also provides evidence that the conditioning changes prototype selection.

The experiments cover several useful dimensions of the problem. Besides overall forecasting error, the paper reports extreme-point errors, F1, average precision, different quantiles, and component ablations. On ETTh1 at H=96 and q=0.98, the controlled base/full comparison improves extreme MAE from 4.237 to 3.684 and F1 from 0.433 to 0.568. The measured overhead and additional representation comparisons help assess the practical cost of these gains.

Critical weaknesses

The central role of the Lyapunov loss is not clear from its definition. In Eqs. (26) and (41), variance and lag-1 autocorrelation appear to be empirical input-window statistics, with the shared reference q_lyap the only explicitly learned quantity. Under that reading, the loss has no gradient path to the risk estimator or forecasting network, although the paper says it regularizes alpha. Neither the hatted notation nor Algorithm 1 identifies a learned transformation that would establish this connection.

The ablations also need clarification. The stage immediately before adding this loss has F1=0.552 in Table 5, while “w/o L_lyap” has 0.418 in Table 10. These may be different configurations, but their difference is not explained. This matters because Lyapunov regularization is presented as a central contribution.

The early-warning results do not establish warning before an extreme event starts. Appendix E.7 measures lead time from the first alert to the event peak. An alert raised after the extreme threshold has already been crossed can therefore receive positive lead time. Table 4 supports warning before the peak, but does not distinguish anticipation from detection of an ongoing event. That distinction is central to the paper’s precursor-state interpretation.

Questions for authors
  • What exactly are the hatted variance and autocorrelation in Eqs. (26)/(41), and through which trainable operations does their loss affect alpha? Please also explain how the Table 5 pre-Lyapunov configuration differs from “w/o L_lyap” in Table 10, including checkpoint selection.
  • Can the existing early-warning runs be evaluated relative to the first extreme-threshold crossing? Please specify how alerts are matched to event episodes and how events already underway are handled.
  • Which splits are used to fit the event quantile, alert threshold, and calibration parameters? What output is thresholded for event F1: the point forecast, the event-indicator head, or alpha? These details would determine how I interpret the event-detection results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.