acceptodds
Under review as a conference paper at ICLR 2027

Forecasting under Changing Response Delays: Prediction and Memory When Relations Disappear

Abstract

An upstream signal may remain useful when its response arrives later, but rising forecast error can also reflect a broken relation or increased noise. We study what a forecaster should predict and remember across these changes. A target-history model fitted while the relation is active can absorb the source effect it must later predict without. Evaluating the shared relation at a training-input reference improves cancellation prediction in two controlled mechanisms. We derive how the inactive prediction also changes effective delay refresh through posterior mass. Matching conditional delay weights preserves the cancellation benefit, while different return rules change its size in opposite directions. Remembering inactive delays helps old-delay recurrence and costs accuracy when a new delay returns. A fixed reference/adaptive configuration reduces noise error by 31.3% on new synthetic streams and 10.6% on semisynthetic streams driven by household measurements, relative to a fixed-scale joint tracker. It meets the synthetic criteria but fails the joint test because external delay and stable-window errors exceed their tolerances. A fixed refresh control is competitive. Changing the inactive target level reverses the reference's advantage; learning an offset mitigates this at a transient cost. These results identify a practical coupling: the forecast assigned to inactivity determines both immediate error and how much delay information survives for future use.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, changing response delays, inactive prediction, probabilistic filtering, memory

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

The paper studies what a forecasting model should predict when an upstream sensor stops helping and whether it should remember the previous delay. It compares using the target's history alone with keeping the original model but replacing the sensor input with its training average. This choice affects both prediction accuracy and how much delay information the model retains.

Critical strengths
  • The paper raises a potentially useful forecasting problem, although its real-world relevance remains unclear.
Critical weaknesses
  • The writing is difficult to follow. The abstract and introduction introduce technical details before explaining the basic idea.

    • The first two sentences do not clearly connect the possible causes of forecast error to what the model should "predict and remember."
    • Terms such as "inactive prediction," "cancellation benefit," and "inactive delays" appear without clear explanations.
    • The introduction also needs clearer explanations of "active state" and "inactive state."
  • The problem setting feels ad hoc. The paper does not establish that the specific relationship-loss setting occurs in the real-world datasets it uses.

  • The experimental scope is limited. The household experiment uses real sensor inputs, but its targets are generated using the same two controlled mechanisms. Although the paper includes some real-data and baseline comparisons, broader evaluation against forecasting baselines on standard benchmarks is needed.

  • The reference assumption is not well justified. Losing a sensor's influence is not necessarily equivalent to setting its input to the training average, even when the model still uses the target's recent history.

  • Given the lack of clarity in the current presentation, a clearer overview figure of the problem setting and method is needed to help readers follow the paper.

Questions for authors
  • Can the authors test the full method against forecasting and online-adaptation baselines on real data with verified delay changes and relationship loss?
Reviewer_GeigOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies forecasting when a source–target relationship temporarily disappears and later returns, possibly with a different delay. It asks what to predict during inactivity and whether to retain previously learned delay information. A state-based framework compares prediction and memory strategies on synthetic, semi-synthetic, and natural forecasting tasks.

Critical strengths
  1. The distinction between what to predict during inactivity and what to remember for a later recurrence is useful. It identifies a forecasting problem that standard stationary or smoothly drifting settings do not capture well.

  2. The experiments include cases where memory retention hurts and simple strategies remain competitive. These results help establish when the proposed choices are useful and when they fail.

Critical weaknesses
  1. The choice of inactive prediction remains difficult in practice. Reference-based prediction depends on the assumed cancellation behavior and can fail under a different inactive offset. It is unclear how to choose it before observing the inactive regime.

  2. The strongest evidence comes from synthetic or semi-synthetic settings. The natural-data results give less support for broad practical usefulness, so I view the contribution primarily as an analysis of a particular forecasting mechanism.

  3. Transition rules, delay refresh, variance adaptation, and recurrence handling introduce several configuration choices. The component studies do not make clear how to select a configuration for a new dataset, which limits ease of use.

Questions for authors
  1. What observable information could guide the choice between reference-based inactive prediction and a history-based fallback before the inactive regime has been observed?

  2. Which findings in the natural-data experiments support the proposed disappearance-and-recurrence mechanism, and which benefits have only been established in controlled settings?

Reviewer_SUM3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper considers forecasting yty_t from its past and a source $x_t4 when the source’s effect changes delay dtd_t becomes more difficult to detect under noise. The model shares one fitted nonlinear response across candidate delays and tracks a weight for each delay plus an inactive state. The proposed change is to predict inactivity by setting the source to its training mean, rather than switching to an autoregression trained on data that still contain the source’s effect. This choice also affects how quickly the model forgets earlier delays.

Critical strengths

The paper’s strength is the identification and analysis of how the inactive prediction affects delay memory. The controlled experiments are also useful; they separate cancellation, delay changes, and noise, and show that remembering an old delay can help recurrence while obstructing adaptation to a new delay.

Critical weaknesses

Most evidence comes from synthetic data. The household experiment uses generated targets, which limits its value as real-world validation. For the real datasets used, they appear to lack verified delay changes or disappearing relationships, and the best neural baseline performs better on every natural task.

Broader comparisons with methods for changing delays, regime shifts, and recurrent concepts would help establish whether the approach offers an advantage.

The inactive prediction relies on the training-mean input representing the absence of source influence. This assumption is restrictive, the paper’s own experiments show that changing the inactive target level can reverse the benefit. A fixed refresh rule is also competitive, leaving the need for adaptive refresh unclear.

Methodological novelty is very modest, with the main contribution lying in the analysis of existing components.

Questions for authors
  • Can you validate the method on real data with verified delay changes or disappearing relationships?
  • What practical advantage does the approach offer, given that the best neural baseline achieves lower mean MSE on each of the five natural tasks?
  • What methodological novelty do you claim beyond the analysis of existing components?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.