acceptodds
Under review as a conference paper at ICLR 2027

Who Contributes to Whom? Cross-Entity Attribution in Panel Time Series Forecasting

Abstract

A forecast indicates what may happen, but understanding the factors that shape the prediction can provide additional evidence for interpreting and assessing it. In panel time series, multiple entities may contain predictive information about one another. Identifying these entity-level contributions can therefore reveal which entities are relevant to a target forecast and help domain experts investigate cross-entity relationships. However, most existing panel forecasting models focus primarily on prediction and do not directly quantify how individual entities contribute to each target forecast. To our knowledge, XPanelNet is the first panel time-series forecasting architecture to integrate entity-level predictive attribution directly into the forecasting process. XPanelNet decomposes each entity's history into trend, seasonal, cyclic, and residual components to construct a base forecast, and then augments it with signed cross-entity contributions computed through a latent vector-autoregressive (VAR) coupling over a low-dimensional shared representation to obtain the final prediction. This additive formulation makes entity-level attribution an intrinsic part of the forecasting process rather than a post hoc interpretation derived from attention or similarity weights. Across nine real-world panel datasets spanning finance, energy, climate, hydrology, transportation, and public safety, XPanelNet achieves the best result of any model, including 10 deep forecasting baselines (TimeKAN, TimeMixer, iTransformer, CrossGNN, PatchTST, TimesNet, DLinear, SCINet, Non-stationary Transformer, and Autoformer), on 7 of 9 evaluated datasets by MAE and 4 of 9 by MSE, while simultaneously providing entity-level predictive attributions.

Primary area:
learning on time series and dynamical systems
Keywords:
Panel Time Series Forecasting, Cross-Entity Attribution, Explainable Forecasting, Predictive Attribution, Interpretable Machine Learning

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_3gJROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Key contribution is to add a Latent VAR component to a base (complex but sound) forecast. The VAR component involves contributions from other entities in the panel, and is aimed to provide (minor) forecasing improvements, but more importantly attempts to add some explanability in the form of contributions from other panel members ("entities).

Critical strengths

The Latent VAR/IRF may be novel, not sure. The base model pretty standard in deep learning approaches.

A wide variety of (DL based) existing approaches are compared against, and the contribution of the additional (VAR) component is identified in Table 3.

Critical weaknesses

I was put off by two main issues:

  1. There seems to be much more use of AI than conveyed by the "AI USE STATEMENT", e.g. see remenants of AI output in Table 5 caption, last phrase of "Router scalability", pretty much all o the last para in the appendix, etc.

  2. Though explainability is a major motivation, the nature of this is pretty basic: a single number (based on correlation, not causality) per other time series (the "entity). See fig 2. There is anecdotal support in form of deletion curves, fig. 3. I think this would be highly inadequate for any domain user; the relative importance of key entities as a whole is likely already known at this very coarse level (e.g. weather is geographically correlated, see first image of fig 2. Practitiioners are more interested in what specific part of another entitie's series caused a surprising read, causation, match with domain knowledge, but the authors do not even consider this and I infer that they did not work with domain experts to understand the value of their results.

There are other issues as well, but the two above are show-stoppers for me.

Questions for authors

should work with domain experts and understand what is already known about the application to have an impact.

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

XPanelNet adds a latent VAR router to entity-local forecasts built from decomposed time series. Source-specific propagation gives signed additive contributions for each target and horizon, with pairwise tensors constructed only for explanation. Evaluation covers nine panel datasets, forecasting ablations, reconstruction, raw-input occlusion, deletion and router randomization. The claimed interpretation is predictive, not causal.

Critical strengths
  1. The explanations are actual terms in the predictor rather than attention weights relabeled as contributions. Linear propagation makes the source accounting exact under the same gate and readout, and the normal forecasting path avoids materializing a dense pairwise tensor. This is a useful design for inspecting a fixed model.

  2. The evaluation covers diverse panels and reports important limitations. Table 1 shows best average MAE on seven datasets and MSE on four; Table 3 includes cases where routing hurts. The explanation tests replace raw histories and rerun the full model, going beyond additive reconstruction alone.

Critical weaknesses
  1. The explanatory claims need a consistent definition and evaluation reference. Equation (1) excludes the target's own source, but Equation (16) and the implementation include it. Zero-diagonal lag matrices do not eliminate multistep self-return paths: A1 squared can have a nonzero diagonal. Table 3 therefore removes both self-feedback and other-source information. Separately, swapping an off-target source's history changes its contribution from P(X) to P(Xswap), so the forecast difference is P(X)-P(Xswap), not P(X). The evaluator correlates that difference with P(X) alone, although the donor contribution need not vanish. Its ranking also uses abs(mean_h P), not mean_h abs(P), allowing horizon-wise signs to cancel. These distinctions matter for the central signed, horizon-specific explanation claim; exact reconstruction does not settle them.

  2. The evaluated model and experimental settings are not sufficiently traceable. Training selects an adaptive Fourier count but saves the original configuration value; the checkpoint loader uses that saved value. If the two counts differ, reloading changes the decomposition. This code path is identifiable, but its impact on the reported checkpoints is unverified. Table 4 also omits its horizon and aggregation, and its final-model entries differ from both Table 1 and Table 7's H=96 entries. Table 6 discloses MAE training versus MSE baselines, different stopping settings and single runs; these comparisons do not isolate architectural gains. Loss-matched reruns of the strongest baselines would help. Presentation compounds the ambiguity: Figure 1's router inset does not match the stated VAR recurrence/readout, and Appendices A.3/A.5 retain editorial notes and a hyperparameter placeholder.

Questions for authors
  1. Can you separate self-feedback from off-target contributions and report horizon-wise diagnostics with a stated replacement reference and consistent magnitude aggregation?

  2. Which resolved Fourier counts and checkpoints produced the faithfulness results? Please identify Table 4's horizon/configuration and provide loss-matched comparisons and variability for the closest baselines.

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

Panel time series forecasting is a subset of multivariate time series forecasting problem, where data from multiple entities are available together. A panel forecasting model can leverage the dependencies between different entities to forecast their futures. There are a bunch of deep learning forecasters to solve this problem of multivariate forecasting. However, it is unclear how an entity contributes to the forecast. Post hoc methods like SHAP can be costly. This paper introduces XPanelNet which is interpretable by construction. The model represents the target as a linear combination of the target predicted by the entities' own history, and signed contributions of all other entities, modeled using a deep VAR(2) model.

Critical strengths

The paper solves an important problem. It reminds me of the NBEATs class of linear models, which used explainable basis functions for forecasting. In this case, the authors use a deep VAR(2) model to model interactions between time series from different entities. To the best of my knowledge, the framing is novel, and I study and a many of design choices makes sense. The problem itself is important, although I do believe that post-hoc explanation methods might suffice in most cases.

Critical weaknesses

The paper is quite hard to follow, especially from Section 3.5 onwards. The authors assume their readers have a concrete understanding of vector auto-regressive models. As a results they use multiple symbols (scale$_k$, AkA_k) and abbreviations (e.g., IRF) without any supporting definitions or intuitions. This makes it incredibly hard to follow, and appreciate the paper and its contributions.

My second reservation is that the experimental settings do not inspire confidence. Recent work [1] has shown that not all multivariate time series benefit from multivariate / panel modeling. It also compares multiple recent methods, including one proposed by itself, on a larger number of datasets. I would recommend that the authors consider those settings for their analysis. Their baselines are old, primarily cover transformers, instead of NBEATS-like linear models, and their numbers do not match either. For example, PatchTST numbers appear to be very different from the original paper. The errors are almost 2x. Also the datasets themselves are known to have known issues. For example, in Exchange, repeating the last value can beat most forecasters. Most recent work use fev-bench, GIFT-Eval etc. for a much better read on forecasting performance. I'd recommend the authors to start with those datasets, just like [1] did. Moreover, it is unclear which differences in the results are significant, specifically in Table 3 where differences are of the order 1e-3.

References

  1. Potosnak, Willa, et al. "MICA: Multivariate Infini Compressive Attention for Time Series Forecasting." arXiv preprint arXiv:2604.06473 (2026).
Questions for authors
  • How is panel forecasting different from general multivariate modeling?
  • ``Post-hoc explanation methods such as SHAP and Integrated Gradients can be applied after training, but computing explanations for every forecast can be expensive, and their attributions are obtained separately from the model’s predictive computation." – This is true, but that’s only the case when the relative attribution changes from one forecast to the other. If temperature is an important predictor for electricity consumption in April, why would it be any different in August?
  • What is lil_i in equation 2? Why do we need gating?
  • Subsection 3.5 is extremely hard to follow: What is IRF? Where is sjs_j used? How are UU and VV estimated? Are they learned? What is scale? This section is quite unintuitive without having a primer on VAR.
  • Line 285: “Models are trained for up to 15 epochs with early stopping based on validation loss (patience 5)” – How is the validation dataset chosen?
  • Line 364: “consistent with the sanitycheck principle of Adebayo et al. (2018).” What is this?
  • What is full-model occlusion?
  • Re. evaluation: have you considered fully correlated entities, or entities with a known lags, or those that are white noise, and how that affects the forecasts?
  • Are the differences significant in Table 3?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.