acceptodds
Under review as a conference paper at ICLR 2027

Explaining Time Series Forecasting with Horizon-Resolved Attribution

Abstract

Recent advances in explaining time series (TS) models have produced methods that identify which past values a prediction depends on. However, most existing methods return a single importance vector, assuming that every predicted step depends on the same past values. In this paper, we show that this assumption does not hold, as different forecast steps depend on different past values. Motivated by this observation, we propose Horizon-Resolved eXplanation (HRX), which adds a horizon axis to the explanation, so that every forecast step receives its own importance map. HRX is a simple yet effective plug-in framework with three components: 1) an estimator that reads these maps out of any differentiable forecaster without modifying the TS backbone, 2) an evaluation protocol that validates the horizon axis by measuring how much a single forecast step changes when the inputs an importance map ranks highest are removed, and 3) a rank criterion that predicts in advance whether the axis is worth resolving on a given TS. We further show that this stepwise dependence is low-dimensional, as the explanations of all steps are built from a few shared maps whose number grows far more slowly than the forecast length. Extensive experiments across various backbones and datasets show that the improvement comes from the horizon axis and holds for estimators of previous explanation methods. Code is available at https://anonymous.4open.science/r/HRX2-BA0B.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, explainability, attribution, horizon-resolved explanation, faithfulness evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VTApOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Time series forecasting models can be explained by identifying which past values drive the predictions. However, recent works generate a single importance scores vector to explain the whole predictions, which can be too broad. The paper argues that different forecast steps should depend on different past values. Therefore, the work proposes a new method called Horizon-Resolved eXplanation (HRX), which adds the horizon axis to the explanation so that every forecast step has its own importance attribution map. From this perspective, a new evaluation protocol is proposed to measure how much a single forecast step changes when the most important features are removed. The new study also shows that step-wise dependence is low-dimensional. Extensive experiments with a setting of 56K runs across 26 backbones and 15 benchmarks show that improvement of the method comes from the horizon axis rather than any particular estimator.

Critical strengths

The paper proposes a clear, simple but novel idea on explaining time series forecasters that each forecast has H values so each step should have its own importance map. Table 1 can show how the work is different from previous works while Figure1 can show the motivation for HRX.

The method is simple and efficient with Gradient of a single forecast step. This help the method still be practical in real applications.

The paper is well written and easy to follow with the definition of each term clearly defined before going to the formula.

The experiments are comprehensive with 26 backbone and 15 benchmarks with 56K runs to show comparison.

The results show that explanations with the horizon axis can have better quality than previous work with only a single explanation.

The analysis on Time vs. channel axis, and how many maps a forecast needs in Section Analysis can show how Horizon should contribute.

Critical weaknesses

The behaviour is expected that if we explain each time step prediction independently, we should see different importance maps. However, the objective of explanation is normally for the whole period, then the features drive that seem enough in some applications. It is great if the paper can show an explanation on each forecast time step and can contribute to the whole explanation for the whole steps at once.

In reality, the motivation of HRX from the user's perspective is not clear. A practitioner or a developer may need explanation for each forecast step, but what this brings them more value than a single vector for all steps is not clear.

Explanations appear hard to read for users. The predictions can be interpreted by humans easier with segments or see the whole period instead of checking individual time points and then checking each explanation. For example, in Figure 1, some users can see the output as a trend or a segment with temporal relationship and ask why past segments or trends can lead to this rather than look at each time step.

The evaluation estimates the shared vector and estimates g_own, g_shuffled, and g_margin. However, as I understand, all need a horizon axis to estimate AUC, and the shared vector also depends on the attribution map of each forecast step. However, compared to previous works, this evaluation can not be used since there are no horizon steps in the setting and the output of HRX is not the same format as others. The paper should consider a shared vector as a single vector to compare or use other metrics to compare to other methods that return a single vector.

The motivation is good but methodology of HRX seems to apply existing gradient-based methods to explain each forecast step. The method should also take the forecast steps around one step to see the relationship and improve the method.

How the method can adapt with time-frequency transformation and segment-based explanation method since the segments can be more interpretable than time points.

How does the explanation for each forecast step cooperate together to explain the whole prediction period? Since each attribution map can explain well in each forecast step but when combined together, the faithfulness or correctness of explanation can be no better than the previous method under the whole period. These comparisons should be considered to show effectiveness of HRX in practice.

The experiments should consider experiment on foundation models which can finetune rather than only zero-shot setting since this can have different behaviour.

Questions for authors

Can the authors explain why users need an explanation for each forecast step rather than a single vector?

Can the authors explain how users can interpret explanations?

Can the authors discuss about the metrics and how the metric can compare between their method with previous ones?

Can the authors explain why they choose gradient only for estimator instead of other methods in time series explanation?

Can the authors show how effective of the method when the explanation presentation is segments or other interpretable form?

Can the authors show how explanation of each forecast step can combine together to explain the forecast and show it better than previous work?

3 comments

Please feel free to let me know if you have any further questions or concerns.

If you feel that your concerns have been adequately addressed, I would sincerely appreciate your consideration of an updated score.

Thanks the authors for the rebuttal. This almost address my concerns. I keep my positive score.

Thank you for your review and constructive feedback!

We look forward to the continued discussion with the other reviewers and AC during the post-rebuttal phase.

Reviewer_pP53OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The work extends time series forecast explainability from being vector based to a matrix based. Therefore, it addresses the contribution of the lookback period to every point in the forecast horizon. The method is agnostic of the forecasting framework as long as the forecasts are differentiable. The main advantage claimed is a more granular explainability spanning the entire forecast horizon. The authors validated their claim using real world datasets spanning multiple domains and also used multiple differentiable forecasting backbones. Overall, the work shows a solid foundation of a new method for forecast explainability but there are some major gaps as pointed out in later sections.

Critical strengths
  1. The authors generated synthetic datasets containing tailored dependencies to test their methodology before moving into real datasets. This is good scientific approach.
  2. The paper is well written with clear explanation and motivation.
Critical weaknesses
  1. There is no citation to prior work most relevant to the one presented here. The ChronoSHAP paper by Calero et al. explicitly computes input attributions for each forecast step, producing an output-step × input-step attribution tensor.
  2. The baseline against which the paper's results are compared is generated by the authors. While this is a legitimate baseline, it is not exactly an established standard based on existing literature. A gradient based method could be used for explaining the mean forecast over the horizon h, in which case a vector is sufficient.
  3. In table 6, the median value of g_margin for Dynamask (0.637) is significantly higher than the authors' method (0.271). There is no discussion on these results and relative merits.
Questions for authors

See weakness.

4 comments

Please feel free to let me know if you have any further questions or concerns.

If you feel that your concerns have been adequately addressed, I would sincerely appreciate your consideration of an updated score.

Dear Reviewer pP53,

With only a few hours left in the discussion period, we would like to thank you for your feedback. If anything remain unclear or require further discussion, we are happy to follow up before the deadline.

Best regards, The Authors

Thank you authors for the detailed response. However, W2 shows marginal gains so I am concerns about the effectiveness of the method.

Hence, I maintain my score.

We sincerely thank the reviewer for reading our response and for telling us the remaining concern.

We would like to briefly put the "size of the W2 gains" in context, using numbers already in the paper.

 

(1) What the numbers mean.

Each gain is a difference of deletion AUCs, where the AUC is the squared change of a forecast step averaged over the deletion sweep and divided by the variance of the forecast (Appendix F.1, Eq. (9)). The gain of +0.148 over the gradient of the mean forecast therefore means that deleting by each step's own row changes that step by an additional 14.8% of the variance of the whole forecast.

 

(2) The W2 gains are as large as our main result.

All values are normalized by the forecast variance, so they are comparable across tables. The W2 gains of each step's own row over the published single vectors range from +0.074 to +0.249, and +0.148 for the gradient of the mean forecast (Table R1, Appendix S.1). This is the same size as the main gain of HRX over the shared vector, +0.143 over 25,517 runs with 95% CI [+0.139, +0.147] (Table 43).

 

(3) The median mixes forecasters with and without horizon dependence.

When a forecaster uses the same inputs for every step (r(E)≈1r(E) \approx 1), no explanation can gain from the horizon axis, and the gain is +0.015, a level that the W2 gains exceed 5 to 17 times. When the forecaster uses different inputs for different steps (r(E)≥3r(E) \geq 3), the gain rises to +0.323 with 98% of runs positive (Table 10). The rank tells a practitioner in advance which case applies, so the single vector is kept where it suffices and the matrix is used where it helps.

 

(4) With ground truth, the gain is large.

On the synthetic data, which the reviewer kindly described as a good scientific approach, the matrix raises AUPRC from 0.586 to 0.983 for the linear backbone (Table 3).

 

(5) The gains are consistent.

Every 95% CI in W2 excludes zero, the gains are positive in 61% to 83% of runs rather than about half as a marginal effect would be, and each of the seven published methods also improves once the axis is added to it, by up to +0.202 for Dynamask (Appendix S.1).

 

Taken together, these results show that the gains are sufficiently large to demonstrate the effectiveness of our method, rather than being merely marginal improvements.

We hope this clarifies that the gains are consistent and grow exactly where the forecast horizon matters.

Thank you again for your time and feedback.

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

Paper "EXPLAINING TIME SERIES FORECASTING WITH HORIZON-RESOLVED ATTRIBUTION" proposes HRX - plug-in method that explains long-interval time series forecasting using importance map per forecasting step rather than importance map for whole forecasting period. Authors also propose test that checks whether each step is better explained by its own map and rule that tells when this is worth doing based on matrix rank. Experiments use many models, datasets, explanation methods.

Critical strengths
  1. Good practical algorithm for long-horizon forecasting XAI. HRX is simple to use: needs no retraining, works with most models, cheap for long horizon forecasting.

  2. Good scale of experiments. Paper tests many models and datasets including specifically designed synthetic datasets where XAI label is known.

Critical weaknesses
  1. Limited novelty compared to existing works - this is my main concern. There exist many reseach papers devoted to XAI for time series forecasting. Current paper mainly uses existing XAI methods for each forecast step individually. Innovation for ML is small and not strong enough for ICLR. However paper could be more usefull for more specialized venue in XAI.

  2. Most results are based on own tests of paper authors. Standard metrics from prior works are used and improvement gains are small.

  3. It is not demostrated why one step explanations are helpfull. Paper authors state that none of existing works can improve hong-horizon forecasting, but there is no case study to demostrate why is that.

  4. Related work section is short in main paper text. It is hard to see how proposed method is different from existing one step XAI methods.

Questions for authors
  1. How HRX is different from directly running existing XAI time-series forecasting methods for each step individually?

  2. Is there real case (with case study) from open datasets where one step maps helps to make a better decision based on time-series forecast rather than one combined map?

5 comments

Please feel free to let me know if you have any further questions or concerns.

If you feel that your concerns have been adequately addressed, I would sincerely appreciate your consideration of an updated score.

Reviewer_MwWeOpenReview ↗

Comment for "Official Comment by Authors"

Thank you for carefully addressing my concerns, I need to make a final decision based on novelty estimation.

We sincerely thank the reviewer for engaging in the discussion and for offering to reconsider the novelty judgement.

 

Novelty in two sentences:

  • (1) While existing TS explanation methods assume that every forecast step depends on the same past values, we show that this assumption fails and make the forecast horizon an explicit, validated axis of the explanation.
  • (2) Beyond computing per-step maps, HRX provides what per-step attribution alone cannot: a protocol that validates a single forecast step's map, and a rank criterion that predicts in advance when the axis is worth resolving.

 

In short, the contribution is the HORZION axis together with the means to validate and predict it, not a new estimator, which is why it improves every published estimator we tested (Table R1b).

 

If this clarifies the novelty, we would sincerely appreciate it if the reviewer could kindly reconsider the evaluation and the score of our paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

General Response to All Reviewers

We have revised the paper and uploaded the new version to OpenReview. All changes are highlighted in "GREEN", and we would appreciate it if the reviewers could take a look at the updated PDF.

 

We sincerely thank all reviewers for their thoughtful reviews.

We are glad that the reviewers found HRX

  • to be "a clear, simple but novel idea" with "comprehensive" experiments on 26 backbones and 15 benchmarks (VTAp),
  • a "good practical algorithm for long-horizon forecasting XAI" that "needs no retraining, works with most models" (MwWe),
  • and "a solid foundation of a new method for forecast explainability" whose synthetic validation is a "good scientific approach" (pP53).

The reviewers also found the paper "well written" (pP53, VTAp).

 


Key concerns and our responses

[1] How does HRX differ from per-step attribution such as ChronoSHAP? (MwWe, pP53, VTAp)

Our contribution is not a new estimator but the horizon axis together with the means to validate it: A protocol that scores a single forecast step, with a shuffled control that rules out sparser maps (Section 3.3, Table 2), and a rank criterion that predicts the gain in advance (Section 3.4, Tables 9, 10). HRX is estimator-agnostic and already holds for eight estimators (Table 6).

Nonetheless, we conducted additional experiments in which every one of seven published methods improves once the horizon axis is added (Appendix S.1), as do SHAP and LIME (Appendix S.5). We have added ChronoSHAP to Section 2 and Appendix P.1.

 


[2] Baselines and metrics beyond our own protocol (MwWe, pP53, VTAp)

Our evidence already goes beyond our protocol: On synthetic data with known ground truth, the matrix raises AUPRC from 0.586 to 0.983 for the linear backbone (Table 3) and wins for all eight estimators (Table 5), and all four metrics of prior works are positive under a sign test over 1,126 runs (Table 4).

Nonetheless, we additionally compared each step's own row against the published single vector of seven methods, including the gradient of the mean forecast. Each step is better explained by its own row in every case, with all 95% CIs above zero (Appendix S.1).

 


[3] Why would a user need a map per forecast step? (MwWe, VTAp)

HRX keeps the whole-period explanation, which is recovered exactly as ∑hE[h]\sum_h E[h] (Section 3.1), and the rank criterion tells the user when the per-step view adds anything (+0.015 when r(E)≈1r(E) \approx 1 vs. +0.323 when r(E)≥3r(E) \geq 3, Table 10).

Nonetheless, we additionally studied a use case: Which forecast steps can still be trusted when part of the input is corrupted? A single vector gives the same answer for every step, while HRX identifies the affected steps far above chance (43.3% of the top-25% affected steps vs. 25.0% at random, Appendix S.3, Figure 14).

 


[4] Readability, segments, and the whole forecast (VTAp)

EE is low-dimensional, so a user reads a few shared maps and their per-step weights (Eq. (7), Table 9).

Nonetheless, we conducted additional analyses:

  • (1) The axis survives aggregation into forecast and lookback segments, with every cell above zero (Appendix S.2)
  • (2) Summed into one vector, the matrix matches or exceeds the gradient of the whole forecast (Appendix S.2)
  • (3) Averaging neighboring steps shows that the step-specific part of each row carries the gain (Appendix S.4)

 


[5] Estimators, foundation models, and horizons (pP53, VTAp)

Every row of Table 6 is HRX, and the higher margin with Dynamask shows that the axis brings an even larger gain with a stronger estimator, while the gradient stays the default for its lower cost (18x vs. 592x a forward pass for Dynamask, Table 18). We have clarified the caption of Table 6.

We also additionally showed that the axis improves segment- and frequency-based estimators (Appendix S.5), holds for five fully fine-tuned TSFMs with a positive margin in 95% of runs (Appendix S.6), and holds at HH = 192, 336, and 720 (Appendix S.7).

 


We hope these revisions address the reviewers' concerns, and we would be happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.