acceptodds
Under review as a conference paper at ICLR 2027

Time Series Foundation Models Can Be Effective Forecasters for Irregular Time Series

Abstract

Time series foundation models (TSFMs) offer strong zero-shot forecasting across domains. Their ability to transfer without task-specific training makes them attractive for sparse irregular forecasting, yet they are typically applied to regularly sampled sequences. Inspired by temporal point process modeling, we study an inference-time event-sequence interface that represents each observed event by its value and inter-event interval, enabling frozen TSFMs to produce next-event value and timing forecasts without regular-grid imputation. On 85 Time-IMM series from nine dataset families, the event-sequence interface achieves lower normalized value MAE and MSE than regular-grid interfaces for each of the five evaluated TSFMs; both errors are also lower in 8 of 9 dataset families. Both interfaces use the same frozen TSFM checkpoint and are evaluated on the same held-out examples, providing evidence that the input interface can be an important performance bottleneck for frozen TSFMs on irregular time series. Compared with four trained irregular baselines, frozen TSFMs using the event-sequence interface achieve lower next-event value and timing errors. On seven EasyTPP datasets, frozen TSFMs using the event-sequence interface provide competitive next-event timing forecasts relative to TPP baselines trained on each dataset. Together, the results show that frozen TSFMs can effectively forecast sparse irregular time series through the event-sequence interface.

Primary area:
learning on time series and dynamical systems
Keywords:
time series foundation models, irregularly sampled time series, zero-shot forecasting

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

The paper studies an inference-time interface for applying frozen time-series foundation models (TSFMs) to irregularly sampled time series (or event series). Each event occupies a sequence position, with values and inter-event intervals represented as numeric channels, and the task is to predict the next event's value and timing. The authors compare this interface with regular-grid preprocessing using the same five frozen checkpoints on 85 Time-IMM series from nine families, alongside trained irregular-forecasting and temporal point-process baselines. Overall, the findings make a compelling case that changing the input representation can improve pre-trained TSFM forecasters without updating their parameters.

Critical strengths
  1. A useful, controlled interface comparison. Identical checkpoints and held-out examples are compared with validation-selected grid configurations. Table 1 reports lower average values of MAE and MSE for every evaluated TSFM; Chronos-2 improves from 0.402/0.583 to 0.363/0.473. This supports a practical contribution independently of comparisons with separately trained models.

  2. Broad evaluation with explicit scope and interface qualifications. The Time-IMM and EasyTPP results cover multiple different models and domains. Appendix B.2 distinguishes independent value/interval forecasting from multivariate conditioning; Appendix G acknowledges one-step point predictions and unaudited pre-training overlap. The paired design still supports within-checkpoint interface gains. Weaker categorical-event results are also disclosed.

Critical weaknesses
  1. The closest pretrained irregular forecasters are neither compared in the experiments nor excluded with a concrete justification. ISTS-PLM (KDD 2025) and VIMTS (ICML 2025) predate this work. Their adaptation procedures differ, but that alone does not establish that a fair forecasting comparison is impossible. The existing oracle-time value protocol offers a potential common task. Each should be compared under aligned information access, with adaptation costs disclosed, or its exclusion specifically justified.
  2. The added value over simple event-indexed predictors is not yet clear enough. Appendix E's normalized one-step "Value change" appears to provide persistence MAE if its targets, masks and aggregation match the main evaluation. This is useful partial evidence. An explicit persistence MAE/MSE baseline, simple autoregression, and basic interval predictors would show how much the TSFMs add on data where recent observations are already strongly predictive.
Questions for authors
  1. Can you compare ISTS-PLM and VIMTS in the existing oracle-time value setting, using the same observed histories, splits, query times and metrics as the other oracle-time baselines? Please report adaptation costs and make clear that these baselines receive the true query time, whereas the event-sequence TSFMs predict the next event directly. If either method cannot be compared fairly, explain the specific obstacle and how it limits the comparison.

  2. Does Table 14's "Value change" use the same targets, masks and averaging as the main value MAE? Averaging its rounded family entries gives about 0.474, compared with TSFM's MAEs of 0.353-0.370. On EPA-Air, however, the diagnostic is 0.109, below the TSFM MAEs of 0.113-0.121 in Table 18. If it is true, persistence is already competitive on that family. Please confirm it, add persistence MSE, and compare simple event-indexed autoregression and last-interval or training-median interval predictors.

  3. How sensitive is the joint-timing comparison to the value/interval loss weight and checkpoint-selection criteria? Appendix C.4 fixes the interval-loss weight at 1.0 and selects checkpoints by validation loss. Raw and log interval targets can have different loss scales. Appendix D.3 studies the encodings, but a check of loss weighting or selection by value MAE would help establish robustness. The oracle-time results already provide useful evidence without the joint loss.

  4. How uncertain are the smaller paired gains? Table 14 reports family MAE gaps of 0.001 on FNSPID and 0.009 on ILINet. Intervals that account for dependence within each series, together with seed variation for the trained baselines, would help distinguish stable gains from small fluctuations.

Minor issue: Line 324 "sections B.2 and D.4 gives", change "gives" to "give".

Reviewer_mthSOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how to apply pretrained time series foundation models (TSFMs) to irregular time series forecasting. To address the reliance of existing TSFMs on regularly sampled inputs, the authors propose an event-sequence interface that represents each observed event using its value and the inter-event interval. This enables frozen TSFMs to directly predict the value and timing of the next event without regular-grid interpolation, model fine-tuning, or architectural modifications. The authors conduct experiments on the Time-IMM and EasyTPP benchmarks, evaluating five TSFMs. The results show that the proposed event-sequence interface reduces value forecasting errors compared with conventional regular-grid interfaces. Furthermore, the method outperforms several trained irregular time series baselines in next-event value and timing prediction.

Critical strengths

The paper proposes a simple event-sequence representation that uses each observed event's value and its inter-event interval as inputs to TSFMs, enabling pretrained TSFMs to handle irregular time series forecasting without additional training or architectural modifications. Experimental results show that this approach achieves lower forecasting errors than conventional regular-grid input methods across all five evaluated TSFMs, demonstrating its effectiveness.

Critical weaknesses
  1. The paper simply combines observed values and inter-event intervals into an event sequence and directly feeds it into existing TSFMs for irregular time series forecasting, without introducing substantial technical innovations. Furthermore, the paper mainly compares the proposed approach with regular-grid methods that introduce numerous artificially imputed values. The performance improvements over these relatively simple baselines are somewhat expected.
  2. The paper only evaluates one-step next-event forecasting, without considering multi-step irregular time series forecasting. Furthermore, the comparisons with specialized irregular forecasting methods are limited to CRU, Latent ODE, Neural Flow, and tPatchGNN, with the most recent method published in 2024.
Questions for authors
  1. Existing irregular time series forecasting methods have widely incorporated timestamps or inter-event intervals as input information to improve forecasting. This paper simply combines observed values and inter-event intervals into event sequences and feeds them into TSFMs. What are the substantial methodological innovations or differences compared with existing time-aware forecasting approaches?

  2. All experiments in this paper focus exclusively on predicting the next observed event. Could the authors further evaluate multi-step irregular time series forecasting, where both the values and timestamps of multiple future events need to be predicted?

  3. Could the authors include comparisons with more recent irregular time series forecasting methods?

Reviewer_Gy6hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to use time series foundation models (TSFMs) to forecast irregular time series. The authors propose a simple method that uses the time interval between observations as an additional input feature. The model predicts the next value and time interval, which is then converted back into the next timestamp. Experiments show that this method reduces MAE compared with regular-grid methods and several trained baselines.

Critical strengths

The paper provides extensive experimental evaluations of zero-shot forecasting across multiple time series foundation models and datasets. The results consistently demonstrate the effectiveness of the proposed approach.

Critical weaknesses
  1. Using the time interval between observations as an input feature is not new. This idea has been widely used since the LSTM era and in temporal point process models. The proposed method is mainly applying an existing idea to time series foundation models.

  2. The foundation models are much larger and pretrained on massive datasets, while the compared models are much smaller. Are the improvements mainly due to the stronger foundation models or the proposed method? More experiments are needed to clarify this.

Questions for authors

see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.