acceptodds
Withdrawn from ICLR 2027

Timer-M1: A Multivariate Time Series Foundation Model via Learning Primitives

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We introduce Timer-M1, a pretrained multivariate time series foundation model that learns with primitives for zero-shot forecasting. Across domains, time series share elementary temporal and relational patterns, termed primitives, yet differ in how these primitives manifest and evolve across different contexts. Despite progress in zero-shot and task-general forecasting, existing foundation models may still struggle to generalize to complex real-world scenarios. To this end, we develop a primitive-based data synthesis and pretraining pipeline. The synthesis pipeline generates series with temporal primitives shared across domains and then assembles real and generated series into multivariate samples using relational primitives. Afterwards, samples are organized into episodes by assigning distinct channel roles as target variates, past-only covariates, and known-future covariates, ensuring that the model is optimized on predictable variates using available exogenous information. Technically, Timer-M1 further adapts gated two-dimensional Transformer blocks that dynamically allocate cross-variate attention across layers. Across three large-scale forecasting benchmarks, Timer-M1 ranks first on both FEV and TIME and second on GIFT-Eval among most recent time series foundation models. These results support effective primitive-based pretraining as a route to robust general forecasting technique across domains and task settings.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Time Series Forecasting, Foundation Model, Primitive, Pretraining

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_9KprOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Timer-M1, a multivariate time-series foundation model trained through temporal and relational primitive composition. A data pipeline combines real and generated source series into multivariate records, and forecasting episodes vary target/covariate roles and observation masks. The 121M-parameter model uses patch embeddings, temporal and cross-variate attention, learned layer-wise gates, and a quantile forecasting objective. Evaluation reports competitive results on FEV, TIME, and GIFT-Eval, selected primitive/reliability tests, and gate/data-construction ablations.

Critical strengths
  1. The model supports univariate, joint-target, past-covariate, and known-future settings with a single compact backbone.

  2. The work links data construction to observation availability instead of only increasing corpus size.

Critical weaknesses
  1. The zero-shot generalization claim lacks a concrete training-data eligibility audit. The authors do not provide source identities, temporal cutoffs, deduplication, or target-interval overlap checks.

  2. The primitive-learning explanation is under-specified and not sufficiently isolated. Appendix A gives representative operators rather than an executable generation recipe, and the supplied paper omits core optimization settings, consumed training exposure, source-mixture weights, and long-context post-training details. The data ablations do not document equal update/token budgets, and there is no same-data control isolating episode-role/supervision sampling or a held-out composition test. Thus the results do not yet establish whether its gain comes from primitive composition, different training exposure, or task design.

  3. The improvement on these benchmarks seem marginal. for the primitive evaluation settings, Timer-M1 seems to be more advantageous since this is how its training data is constructed.

Questions for authors

1: Which exact GIFT-Eval and TIME training sources and time intervals enter pretraining and long-context post-training? Please provide the holdout/deduplication policy and target-interval overlap audit, plus results separated by training-source exposure where applicable.

2: What are the actual operator distributions, source-mixture weights, episode sampling distributions, optimization schedule, update/token budget, and compute for the main model and each ablation? Were the recipe comparisons matched for consumed training exposure rather than stored corpus size?

3: Does the primitive benefit survive matched-budget reruns and a held-out primitive-composition test? A same-data control that changes only episode-role/supervision sampling would also clarify whether the gains come from relational generation or the training task design.

Reviewer_7YseOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Timer-M1, a foundation model for multivariate time series. It combines real and synthetic data to construct temporal and cross-variate primitives. During pretraining, it uses an Episode mechanism to organize different variate roles and observation masks. Architecturally, Timer-M1 uses a 2D Transformer with layer-wise learnable gates. Across the FEV, TIME, and GIFT-Eval benchmarks, it achieves overall ranks of 1st, 1st, and 2nd among compared baselines, supported by ablations on data construction, gating, and primitive forecasting.

Critical strengths
  • The research problem is practical, and the data construction design is reasonable. The paper focuses on how time-series foundation models generalize across domains, multivariate relations, and diverse covariate settings. Combining complex real-world dynamics with controlled cross-variate relations offers a valuable way to build diverse pretraining data.
  • The experiments are broad and show competitive forecasting performance. The authors compare Timer-M1 with strong recent models like Chronos-2 and TimesFM-3 across the FEV, TIME, and GIFT-Eval benchmarks, alongside analyses of data construction, gating, and controlled primitive forecasting. Joint data generation yields positive gains on several metrics, supporting its effectiveness.
Critical weaknesses
  • The data generation method has limited novelty. Prior work like Chronos-2 already builds multivariate relations on synthetic data, such as linear mixing, lags, and cointegration. Timer-M1 mainly adds real series into this process using existing primitives and operators, which feels like an extension of current frameworks rather than a new method.
  • The unique benefit of the Forecasting Episode design is unclear. Its core steps include window sampling, setting target or covariate roles, and masking which are very similar to existing time-series foundation models. While the authors use known generated relations to build tasks, there is no separate ablation to show how much this helps over random task sampling.
  • The data recipe lacks clear ablation controls and details for reproduction. In the Full Recipe vs. No-Joint ablation, it is unclear if all setups use the exact same training tokens, steps, and compute budget. Also, the paper does not give the actual training sampling ratios or exact rules for key generation parameters.
  • The paper does not fully prove generalization to unseen primitive combinations. The controlled tests use the exact same primitive types (seasonality, trend, lag, and cointegration) as the pretraining data. Without a clear split between training and testing parameters or combinations, the results only show that the model learns seen patterns, not unseen combinations.
Questions for authors
  • Compared to existing methods like Chronos-2, what unique learning abilities do Timer-M1's real-synthetic joint relation generation and Episode design provide?
  • Do the Full Recipe and No-Joint setups use the exact same training tokens, optimization steps, and compute budget? Can the authors share the parameter distributions of the main generation operators and the actual training sampling ratios to clarify the standalone benefit of joint generation?
  • How do the generation rules, parameter ranges, and primitive combinations in the Primitive Evaluation differ from those in the pretraining data? Does the test include combinations unseen during training, and how well do current results prove primitive reuse across tasks and generalization to unseen combinations?
  • Please check whether data leakage is present in the evaluations for the current GIFT-Eval and TIME datasets.
Reviewer_33WxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Timer-M1, a foundation model for time series forecasting. Its central proposal is a data construction pipeline that combines temporal generators with relational operators applied to real, synthetic, or mixed-source series. The resulting records are converted into forecasting episodes with target, past-only covariate, and known-future covariate roles. The model uses factorized temporal and variate attention, with a learned scalar gate on each variate-attention residual branch, and is trained with a quantile forecasting objective. Evaluation covers FEV, TIME, GIFT-Eval, controlled primitive diagnostics, and ablations. The reported results show competitive forecasting performance, with promising results for historical-covariate tasks and periodic extrapolation.

Critical strengths
  1. Timer-M1 uses a shared 121M-parameter checkpoint across three benchmarks and supports multiple target/covariate configurations. Its aggregate performance matches recent baselines.

  2. The FEV past-only covariate results are encouraging: relative to TimesFM-3, the reported MASE and SQL skill scores correspond to approximately 5.3% and 5.1% lower normalized errors, respectively. This suggests practical value, especially for applications where related historical series are available but future covariates are not.

  3. Applying relational operators to real and mixed-source series is a useful direction for expanding multivariate pretraining data without relying exclusively on synthetic trajectories. The Full Recipe versus No-Joint ablation provides initial evidence for this idea: on FEV, the full recipe reduces normalized MASE and SQL errors by approximately 3%. The controlled diagnostics also make specific strengths and failure modes more visible than aggregate benchmark scores alone. With a sufficiently specified and reproducible recipe, this could be a useful contribution to time series pretraining practice.

Critical weaknesses
  1. The proposed approach overlaps with Chronos-2 in its use of temporal/group attention, heterogeneous target and covariate roles, observation masks, and synthetic multivariate dependencies. The most plausible incremental contributions are therefore the extension of relational construction to real and mixed sources and the layer-wise residual gates. However, the paper does not sufficiently isolate these contributions from other differences in data composition and training. The benefit of Full Recipe over No-Joint is concentrated in FEV; the improvements on GIFT-Eval and TIME are much smaller, and TIME CRPS slightly worsens. Furthermore, the primitive diagnostics do not clearly separate training compositions from held-out test compositions. Several relational groups contain only two scoring records, which are not necessarily independent trajectories, and Appendix C.3 states that the displayed groups are a selected subset of the diagnostic suite.

  2. Appendix C.4 states that shared GIFT-Eval and TIME training collections remain in all data ablations, while Appendix B.4 notes that training exposure differs across methods. The manuscript does not clearly identify whether these collections are independent pretraining corpora or training splits from the evaluation datasets, nor does it provide an auditable dataset-level overlap account. This distinction matters for interpreting the zero-shot claims; it is not, by itself, evidence of test leakage. In addition, the data ablations do not adequately specify whether training steps, consumed samples/tokens, and compute are matched when generation routes are removed. The paper also omits key details on generator distributions, mixture weights, optimization, and long-context post-training. These gaps are consequential because the central contribution is the training recipe, and the aggregate advantage over TimesFM-3 on FEV/TIME is small enough that uncertainty estimates and precise evaluation settings are needed to assess its significance.

Questions for authors
  1. Were any primitive combinations or generator parameter ranges explicitly held out during training? Is there a comparison between joint and independent prediction using the same Timer-M1 checkpoint?

  2. Can the authors provide per-configuration results and paired uncertainty estimates for the comparisons with TimesFM-3 and for Full Recipe versus No-Joint, accounting for related configurations from the same dataset? Besides, are the aggregate gains and the FEV past-only covariate gains broadly distributed across datasets or concentrated in a small subset?

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

Related papers

Discussion (0)

Sign in to comment.