acceptodds
Under review as a conference paper at ICLR 2027

On the Encoding of Calendar Features in Covariate-Informed Time Series Foundation Models

Abstract

Time series foundation models (TSFMs) increasingly offer covariate support, with calendar information being the most common type of categorical covariate in practice, as it is available for every point in time. However, how these features are encoded is largely treated as an implementation detail rather than as an object of study, despite encoding being one of the few design choices left for pretrained models at inference time. With the present study, we demonstrate that the way information is propagated into the model substantially affects forecast accuracy of TSFMs. Therefore, we systematically compare seven calendar encodings, including three conventional target-agnostic representations (ordinal, one-hot, and cyclic) and four target encodings, across four state-of-the-art TSFMs (Chronos-2, TabPFN-TS-3, TiRex-2, TimesFM-3) on nine datasets from multiple domains in a zero-shot rolling evaluation. The nine datasets are influenced by social activities, such as weekends and holidays, making calendar information particularly relevant for forecasting. We find that encoding choice affects forecast accuracy across all models, with MASE differences of up to 9.9 % on non-holidays and 83 % on holidays within a single model. Effectiveness and utilization vary substantially across models. Improvements are concentrated around holidays, where calendar information provides signals unavailable from target history, while some model encoding combinations also improve non-holiday forecasts. A SHAP-based attribution analysis locates the contribution of the covariates relative to the autoregressive component and explains why some model–encoding combinations degrade performance. Our findings suggest that current benchmarks for covariate-informed TSFMs can distort model comparisons by imposing fixed encoding strategies. Future benchmarks should therefore evaluate multiple encodings or provide model-specific encoding strategies aligned with each model’s best practices.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Machine Learning, Time Series Forecasting, Foundation Models, Benchmarking, Covariates

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_SP6rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper empirically studies how calendar covariates should be encoded for zero-shot, covariate-informed Time Series Foundation Models (TSFMs). It compares seven encoding strategies—three conventional target-agnostic (ordinal, one-hot, cyclic) and four target-based (mean target encodings over hour×weekday with variants for holidays)—across four recent TSFMs on nine hourly datasets with pronounced weekend/holiday effects, and uses a SHAP-based attribution method to analyze how models utilize these covariates. The main findings are that (i) encoding choice materially affects accuracy for all models, (ii) target-based encodings yield the largest gains on holidays and are often more robust overall, and (iii) the best encoding is model-dependent, implying that benchmarks with fixed encodings can misrepresent relative model performance.

Critical strengths
  1. Provides a systematic, architecture-spanning investigation of calendar encoding—an overlooked but impactful design choice in zero-shot TSFMs.
  2. Introduces a suite of simple, context-only target encodings that are leakage-safe (computed from past context) and pragmatically applicable at inference time.
  3. Extends a sampling-free SHAP methodology to multiple TSFMs for diagnosing the relative contribution of calendar covariates versus autoregressive signal.
Critical weaknesses
  1. Target encodings are computed as simple means without regularization/backoff; low-support categories (especially holidays) can yield high-variance encodings
  2. Only hourly datasets with strong calendar/holiday signals are included; lacks negative controls (datasets where calendar is weak) and other cadences
  3. Context length is fixed to 8192 based on Chronos-2’s limit; this may not be optimal for other models and constrains generality
  4. The method assumes each forecast-horizon category combination appears in context; handling of missing combinations (e.g., rare holidays within a 49-week context) is under-specified
  5. There is a lack of clear figures to illustrate the motivation and the pipeline
Questions for authors

See weakness

3 comments

We thank the reviewer for thoroughly reading our paper.

Weakness 1: no regularization/backoff

We deliberately use simple context means to keep the encoding parameter-free and identical across models, so that differences can be attributed to the models rather than to encoding-specific tuning. We acknowledge that holidays are a low-support category. However, our target-encoding variants already span a range of granularities, from coarse, high-support to fine, low-support holiday categories, and thus effectively cover different degrees of backoff. Even the finest variant with the lowest support (Th,W,HT_{h,W,H}) outperforms all target-agnostic encodings on holidays for three of four models [ref. Table 2]. While the low support count of the Th,W,HT_{h,W,H} encoding is a weakness of the specific encoding, we argue that it is not a weakness of our benchmark, as we deliberately include target encodings from high support count (Th,WT_{h,W} (Hol = Sun)) to low support count (Th,W,HT_{h,W,H}).

Weakness 2: only hourly data, no negative controls

We agree that the generalizability of our results is limited by the lack of another resolution. Therefore, we will add another experiment run with 4 of the 9 datasets that are available in 15 min resolution to the revised paper.

Regarding the suggested negative control, we thank the reviewer for this idea. However, the scope is deliberate: we study how the encoding of calendar information affects forecast performances, which presupposes that this information is relevant. Datasets with little calendar influence therefore fall outside this scope.

Weakness 3: context length fixed to 8192

Context length is not fixed to 8192, as Appendix A.3.11 contains the overall results with all models and context length 2048.

Weakness 4: missing combinations in context

We fully acknowledge that our description of unseen calendar combinations in the paper was insufficient and appreciate this comment. For context length 8192, one target encoding (Th,W,HT_{h,W,H}) encounters a forecast-horizon combination that does not exist in the context in 13 of the 110 holidays of the benchmark dataset. With context length 2048 this problem occurs substantially more frequently, with 65/110 holidays containing a combination not present in the context. On non-holidays, this is never the case for both context lengths, and all other target encodings do not suffer from this problem at all. The fallback solution in this scenario is to replace the encoding for the affected day by the context mean. This is a fallback that is possible in every situation. We will integrate a description of this fallback, as well as the support counts and the number of affected windows in the revised paper.

While the differences in support count show in the comparison between target encodings, the Th,W,HT_{h,W,H} encoding still shows comparable or superior performance to the non-target encodings across both context lengths. An ablation study using Th,WT_{h,W} (Hol = Sun) as a fallback would likely improve performance. However, this would make the overall encoding more similar to the fallback, complicating the attribution of gains to the explicit Th,W,HT_{h,W,H} encoding. We therefore retain its original form.

Weakness 5: missing figures for motivation and pipeline

We thank the reviewer for this suggestion. We will include a figure illustrating our pipeline in the appendix, as the limited space in the main paper requires us to prioritize figures that we consider more relevant to the central contributions of our work.

I thank the authors for their rebuttal. Although the clarification on Appendix A.3.11 is noted, the response reinforces several of my original concerns.

Specifically, the fact that 65/110 holidays require a naive global-mean fallback at L=2048L=2048 highlights the practical fragility of unregularized target encodings under shorter contexts.

Furthermore, declining to include negative controls leaves open the question of whether target encodings cause performance degradation on datasets without strong seasonality.

Coupled with the fact that the 15-minute cadence results are only promised for a future revision, I do not find sufficient grounds to raise my score and will maintain my rating of 2.

AuthorsOpenReview ↗

Response to Reviewer

We apologise for any misunderstanding. These additions were planned for this revision from the outset, and many new results have now been incorporated. Computing them, however, took considerable time, which caused the delay.

Reviewer_Tjt9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how calendar feature encoding affects the forecasting performance of time series foundation models (TSFMs). The authors compare seven encoding strategies, including ordinal, one-hot, cyclic, and four target-based encodings, across four TSFMs and nine datasets in a zero-shot rolling forecasting setting. The experiments show that encoding choices can substantially affect prediction accuracy, particularly on holidays, and that the effectiveness of different encodings varies across models. The authors further employ SHAP to analyse the contributions of calendar covariates relative to historical target values. Based on these findings, the paper argues that existing TSFM benchmarks using fixed encoding strategies may lead to misleading model comparisons and recommends considering multiple or model-specific encodings in future evaluations.

Critical strengths

The paper investigates how calendar feature encoding affects the performance of time series foundation models, an important but often overlooked factor in forecasting evaluations. This is relevant because encoding choices may influence the reliability of model comparisons.

The authors compare seven encoding strategies across four foundation models and nine datasets, with additional SHAP-based analysis. The experiments provide useful empirical evidence that encoding effectiveness varies across models and forecasting conditions, particularly on holidays.

Critical weaknesses

My main concern is the limited novelty of this work. The paper mainly compares several existing encoding methods, such as ordinal, one-hot, cyclic, and target encoding, but does not introduce a new method or provide much new understanding of why these methods behave differently. The work feels more like an empirical comparison than a research paper with a clear technical contribution. Although the results might be useful in practice, I am not fully convinced that the contribution is strong enough for ICLR.

Another concern is whether the comparisons are entirely fair. The encoding methods produce different numbers of input features, which may affect model performance. Also, target encoding uses information from historical target values, while the other methods only use calendar features. This means that the performance differences may come from the additional information or feature dimensions, rather than the encoding strategy itself. I would like to see more controlled experiments to separate these effects, as they are important for supporting the paper's main conclusions.

I also have some concerns about the SHAP analysis. SHAP can help show which features contribute to a model's predictions, but it does not necessarily explain why one encoding method performs better than another. The authors seem to use these results to explain the underlying reasons for the performance differences, which I think goes beyond what SHAP can directly support. More targeted experiments would be needed to make these explanations convincing.

Questions for authors

What are the main research contributions beyond systematically comparing existing encoding techniques?

How do the authors isolate the effect of encoding representation from differences in feature dimensionality and the use of historical target statistics?

2 comments

We thank the reviewer for thoroughly reading our paper.

Weakness 1 and Question 1: novelty

We agree that the encodings themselves are not novel and we do not claim a methodological contribution in this respect. Our contribution is to identify and systematically quantify a problem that is relevant for both research and practice but currently overlooked. For models trained on the target task, encoding choice matters little, as they learn to exploit covariates during training. Zero-shot pretrained models, however, cannot adapt and thus benefit only from encodings they can exploit out of the box. As pretrained models become increasingly important, encoding choice becomes a relevant design decision. Yet, key benchmarks such as fev-bench and FETS provide calendar covariates in only one fixed encoding. Our results show that this can bias model rankings in two ways:

  1. Encoding fit: an encoding may benefit one model more than another, depending on its pretraining or architecture.
  2. Covariate harm (more significant): for calendar information that reflects the autoregressive structure of the target (hour, weekday), some models improve with the covariate while others degrade relative to their no-covariate variant. Supplying covariates uniformly therefore penalizes models that already capture this structure from the target alone. Since non-holidays make up the vast majority of days, supplying such covariates by default can substantially harm overall performance.

New experiments with models released after the submission deadline (t0-gamma, Timestral) confirm this: the overall model ranking changes across encodings (Table 3 of the revised paper). This underlines the relevance of our work, as it shows that current benchmarks distort performances of pretrained models by using fixed encodings. Our work thus sits at the intersection of ICLR areas "Foundation or frontier models, including LLMs" and "Datasets and benchmarks".

Weakness 2 and Question 2 (part 1): dimensionality and fairness

Dimensionality is an inherent property of each encoding, and a practitioner choosing an encoding implicitly chooses its dimensionality. However, we agree that we currently cannot with certainty attribute superior performance of low-dimensional encodings to the encoding strategy alone. Therefore, we will conduct further experiments for the revised paper.

Weakness 2 and Question 2 (part 2): use of historical target statistics

We agree that target encodings carry target-derived information, even though they are computed only from the context (Eq. 1). However, our goal is not to separate calendar information from target statistics. We study how models use calendar covariates under each encoding, which is the choice practitioners face at inference time. From this view, target statistics are part of the encoding, not a confound. As a complement, we will add a SHAP analysis for the conventional encodings in the revised paper. This will show whether models rely on calendar information differently when no target-derived statistics are available.

Weakness 3: SHAP does not explain performance differences

We agree with the reviewer. SHAP decomposes a forecast into additive contributions, but it does not decompose the forecast error. It therefore shows how strongly a model relies on the covariates, not why an encoding performs better. We have revised the manuscript accordingly: the SHAP analysis now describes the models' reliance on the covariates, and our statements on why encodings differ in performance are phrased as interpretations consistent with the results rather than as explanations. These are supported by the controlled comparisons described in our response to W2, which hold either the dimension or the use of target statistics fixed. Where attributions and performance results agree, we report them as consistent observations rather than causal explanations.

Thanks for your valuable response. I will keep my score as a weak reject, since the contribution is more technical than research-oriented. However, I agree that this is a complete work, but I do not think it is suitable for ICLR, a highly reputable conference.

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

The paper compares calendar encodings across four pretrained forecasting models on nine hourly series, separating holiday and ordinary-day performance. It finds substantial model–encoding interactions and uses grouped Shapley values to examine input-channel reliance.

Critical strengths
  1. The within-model comparisons keep forecast windows, context, horizon, and metrics fixed. They show that a routine preprocessing choice can materially change model comparisons.

  2. The holiday split exposes effects hidden by overall averages. Per-dataset results, a shorter-context experiment, and seasonal/profile baselines add useful checks.

Critical weaknesses

Sparse-category handling is insufficiently specified. Equation 1 assumes that every future calendar combination appears in the context, while Appendix A.1.1 acknowledges that some combinations may be absent. The fallback and smoothing rules, support counts, and affected-window counts are not reported. These details matter for reproducing the holiday results and assessing whether the target-encoding rankings remain stable under reasonable treatment of rare categories.

The explanatory conclusions need clearer limits. The study establishes encoding effects in hourly, calendar-sensitive, day-ahead forecasting. Its SHAP analysis retains target-derived profiles when historical target groups are masked, so the input channels share information. The resulting attributions describe channel reliance, but do not isolate independent calendar information or establish dimensionality-induced signal dilution. The proposed mechanism and its applicability beyond the evaluated setting therefore remain uncertain.

Questions for authors

How are target encodings computed for unseen or sparsely observed calendar combinations? Please report the fallback or smoothing rules and affected-window counts for both context lengths, and check whether the holiday conclusions persist under one reasonable alternative treatment.

Which SHAP conclusions concern reliance on the supplied profiles, and which concern independent calendar information? Could you narrow the dilution explanation accordingly, or provide a focused control supporting that mechanism, while clarifying the scope of the conclusions beyond hourly calendar-driven tasks?

1 comment

We thank the reviewer for thoroughly reading our paper.

Weakness 1 and Question 1: sparse categories and unseen calendar combinations

We fully acknowledge that our description of unseen calendar combinations in the paper was insufficient and appreciate this comment. For context length 8192, one target encoding (Th,W,HT_{h,W,H}) encounters a forecast-horizon combination that does not exist in the context in 13 of the 110 holidays of the benchmark dataset. With context length 2048 this problem occurs substantially more frequently, with 65/110 holidays containing a combination not present in the context. On non-holidays, this is never the case for both context lengths, and all other target encodings do not suffer from this problem at all. The fallback solution in this scenario is to replace the encoding for the affected day by the context mean. This is a fallback that is possible in every situation. We will integrate a description of this fallback, as well as the support counts and the number of affected windows in the revised paper.

While the differences in support count show in the comparison between target encodings, the Th,W,HT_{h,W,H} encoding still shows comparable or superior performance to the non-target encodings across both context lengths. An ablation study using Th,WT_{h,W} (Hol = Sun) as a fallback would likely improve performance. However, this would make the overall encoding more similar to the fallback, complicating the attribution of gains to the explicit Th,W,HT_{h,W,H} encoding. We therefore retain its original form.

Regarding the smoothing rules: the encodings themselves already act as different forms of smoothing. Th,WT_{h,W} (Hol = Sun) pools sparse holiday conditions with a well-supported category, and the graded target encodings represent different trade-offs between support and specificity. We therefore do not apply additional smoothing rules.

Weakness 2 and Question 2: limits of the explanatory conclusions

We thank the reviewer for this comment and agree that our SHAP values quantify the models' reliance on the input channels as provided, not independent calendar information. This is intended. We aim to explain how each TSFM uses the encoding as it is given at inference time, so the target encodings are held fixed across coalitions (A.1.2), and their attributions include the target statistics they summarize. We have revised the manuscript to state this explicitly and to narrow the explanatory claims accordingly.

All our SHAP-based conclusions therefore concern reliance: covariates receive 20–34 % of the attribution on holidays and at most 14 % on non-holidays, most of the holiday attribution shifts to Th,HT_{h,H} when it is provided separately (except for TiRex-2), and Chronos-2 relies less on the profile than the other models. The role of independent calendar information is addressed by the encoding comparison itself, since the conventional encodings carry the same calendar information without target statistics. As SHAP decomposes the forecast and not its error, the attributions alone cannot explain why one encoding performs better than another. We therefore explicitly relate them to the performance results as complementary evidence on how the models use the covariates.

Question 2: Scope beyond hourly tasks

We understand your concern that our conclusions are based solely on hourly resolution data. Therefore, we are going to provide results of an additional experimental run on the 4 out of 9 datasets that are available in 15 minute resolution.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Rebuttal Response by Authors

We sincerely thank all reviewers for their thorough and constructive feedback. The revised manuscript incorporates all changes announced in our responses above. We believe that this work makes a meaningful contribution to ICLR, particularly in connecting rigorous benchmarking with the evaluation of frontier time series foundation models. This contribution has benefited greatly from the reviewers’ comments, which have substantially strengthened both the analysis and the presentation of our results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.