acceptodds
Under review as a conference paper at ICLR 2027

Forecast residuals for the diagnosis of hidden traffic events: the ranking of normal models depends on the alarm rule

Abstract

Hidden traffic events such as lane closures, demand surges, signal faults and sensor failures are commonly detected by forecasting normal traffic and thresholding the residual. We ask what the two stages of such a detector should condition on: the normal model that produces the residual, and the rule that reads it over sensors and time. The study uses a simulated signalised corridor in which every event episode has an event-free twin sharing all random variables, so that the event's effect on each sensor is known exactly, and radar measurements from Interstate 24 with a located event log. The central finding is that forecasting accuracy alone does not determine detection quality: the ranking of normal models depends on the residual score and the alarm rule. Under an instantaneous threshold a forecaster that conditions on recent observations predicts the event it is meant to expose. On the corridor the seasonal prior is not improved upon by the selected hybrid under this score, while under a rule that integrates the residual the accurate penalised autoregression decisively becomes the best linear pairing, at a longer delay and a longer time in alarm; the freeway ordering does not reverse. The relation between accuracy and detection differs between model families under one rule, and a nonlinear network and a zero-shot foundation model are poor inputs under either rule for different reasons. Two results of narrower scope follow: adding the nearest sensor to the detector's input raises recall substantially in this benchmark, and the normal model that is best for detection under an instantaneous rule is the worst among the tested strategies for reconstructing the event-free counterfactual. The generator, configurations, splits, full sweep tables and scripts accompany the submission; the corpus and checkpoints are to be released on publication.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series anomaly detection, Unsupervised model selection, Forecast-based event detection, Sequential change detection, Synthetic perturbations, Traffic incident detection.

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_vYZ6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how the choice of traffic forecaster and residual-based alarm rule affects event detection. Using paired simulations and real I-24 traffic data, it compares forecasting models and alarm rules in terms of recall, detection delay, and alarm duration. It also examines spatial context, changes in traffic conditions, and reconstruction of event-free trajectories.

Critical strengths

S1. The paper examines a practically relevant question: whether forecasting accuracy is a suitable criterion for selecting normal models in residual-based traffic event detection. Considering the interaction between the forecaster and the alarm rule provides a useful perspective on detector evaluation.

S2. The paired simulation design offers a controlled setting for analysing event effects. Event and event-free trajectories generated under the same random conditions provide a reference for evaluating counterfactual reconstruction.

S3. The paper shows that the ranking of normal forecasting models can change with the residual-based alarm rule at a matched false-alarm rate. This observation motivates evaluating the forecaster and alarm rule jointly rather than selecting the forecaster solely by prediction accuracy.

Critical weaknesses

W1. The paper is difficult to follow due to an unclear central motivation and a fragmented presentation. It shifts between model selection, alarm rules, spatial context, and counterfactual reconstruction without clearly connecting these analyses to the main contribution.

W2. The scope of this observation remains insufficiently established. The main ranking reversal is demonstrated in a single simulated corridor and is not reproduced on I-24. Although the paper reports some recall–false-alarm curves, it does not clearly establish whether the central reversal persists across a range of false-alarm budgets and alarm-burden measures. Further analysis of this sensitivity would help distinguish a robust model-selection insight from a finding specific to the evaluated setting and operating point.

W3. The contribution beyond established residual-based process control is not sufficiently clear. The paper acknowledges that adaptive forecasts can absorb sustained shifts, but does not clearly distinguish its new methodological or conceptual insights from an empirical examination of this known mechanism in traffic settings.

Questions for authors

Q1. Does the main ranking reversal persist across different false-alarm budgets and when methods are matched by time in alarm rather than alarm count?

Q2. What explains the absence of the ranking reversal on I-24, and under what conditions should it be expected?

Reviewer_ew4QOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a case study of detection of traffic events (such as lane closures) from sensor and radar data using both a simulated freeway environment and one real-world dataset. A variety of modeling approaches, including classical autoregressive methods and time-series foundation models, are evaluated and analyses are presented from different perspectives. The main issue with the paper is that its methodological contribution will be of limited interest to an ICLR audience: it would be much better suited to an audience interested in applications of machine learning to traffic data modeling.

Critical strengths

None

Critical weaknesses
  • the main weakness of this paper is that it does not make clear how the results from this specific application (detecting unobservable events in traffic data) would be of interest to a broader audience of ICLR readers who are not directly working on this topic. For example, there is limited discussion in the paper of how the results in the paper could be relevant to readers interested in time-series anomaly detection more broadly. As a result the paper is likely to be of limited interest to an ICLR audience. I recommend that the authors consider a more domain-specific publication venue, such as one of the Transportation Research journals or IEEE Transactions on Intelligent Transportation Systems, where readers (and reviewers) are likely to find it much more relevant and interesting

  • the writing is very dense and wordy: it buries relatively simple ideas in complex long-winded sentences and paragraphs. As an example (and this is just one example) if you read both the abstract and the conclusion, it is very difficult to figure out what precisely is the intended contribution of the paper. I strongly recommend to the authors to rewrite their paper in plain, simple, and precise language. I don't know to what extend AI was used in writing of the paper (beyond the description in the "AI Use Statement"), but I would encourage the authors to dial the AI-usage back for future versions of this paper if you want to present a more readable paper to your audience.

Questions for authors

None

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

The paper studied the problem of gap between accurate forecasting model and deviation detector. The paper reveals that forecasting model that depends on the observation data of recent trend would ignore the deviation and reduce the redisual signal that would give the warning. The validation of this idea is primarily validated in the open source traffic simulator and real I-24 freeway data.

Critical strengths

The paper raises an important research question of gap between accurate forecasting and useful deviation detector. This is a topic that should be studied well to make forecasting model applicable to daily life application. The study also found that CUSUM would be a better metric to detect the deviation.

Critical weaknesses

The paper is poorly written with AI-generated content without human check. The main experiments are conducted in the simulation. Although the authors include a validation on real I-24 data, the central finding that the ranking reversal under CUSUM does not appear on the real data, which raises conerns about the real world applicability of the results.

Questions for authors

1.Table 2 reports the seasonal-minus-hybrid recall gaps of 0.12 and 0.22 under the instantaneous score. But the analysis text on page 5 is not in accordance with this result. Which one is the real experimental result.

  1. Why does tne reanking reversal of CUSUM does not appear on I-24 data?

  2. On page 9, the author mentioned that forecast absorption of sustained shifts is a known limitation in Statistical Process Control. What is the new insight that could be obstained from this paper? By far, it looks like just a reproduction of this limitation in the field of traffic forecasting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.