- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes UniTIMS, a pretrained Transformer for zero-shot irregular multivariate time series forecasting. The model combines time-aware patch representations, an additive temporal attention bias, and a hybrid attention mask that allows bidirectional attention over the observed context while retaining causal attention over the forecast sequence. The authors also introduce VersaTSA for pretraining and evaluate the model on irregular forecasting benchmarks, regular multivariate forecasting datasets, and fev-bench. The reported results suggest promising transfer to irregular forecasting, although inconsistencies in the main comparison table and the method description limit the conclusions that can currently be drawn.
Critical strengths
-
The paper addresses a practically relevant limitation of time series foundation models: transferring to observations with irregular intervals and asynchronous variables. Representing actual timestamps at the patch level and using their relationships to bias attention is a reasonable design for this setting. The combination with multivariate context provides a potentially useful approach to zero-shot irregular forecasting.
-
The evaluation covers both irregular and regular forecasting, multiple model sizes, and several forecast horizons. In particular, the UniTIMS averages in Table 2 agree with the detailed results in Table 9 at the reported precision, and the reported ECL and ETTm2 results are competitive. Section 4.4.2 also explicitly specifies the exclusion of the three irregular and six regular evaluation datasets from pretraining, which is an appropriate precaution for assessing zero-shot transfer.
Critical weaknesses
-
Several baseline results in Table 1 are assigned to the wrong datasets, affecting the central performance comparison. In Table 1 of the cited Hi-Patch paper (Luo et al., 2025), Hi-Patch's MSE is 2.57 +/- 0.02 on Human Activity and 4.86 +/- 0.03 on PhysioNet, using the displayed 10^-3 scale. The present submission reverses these assignments. The same issue occurs for several other methods, including iTransformer and PatchTST, whereas GRU-D is assigned correctly; a global exchange of the two column headers would therefore not resolve the table. If UniTIMS's own dataset labels are correct, its Base model's Human Activity MSE of 3.41 is worse than Hi-Patch's 2.57, while the Large model's 2.44 remains better but by a much smaller margin than the current table suggests. These errors do not establish that UniTIMS is ineffective, but they prevent the current table from supporting the stated comparison with specialized full-shot models. A corrected table and confirmation of the evaluation configurations are necessary to assess which conclusions remain supported.
-
The description and validation of the core temporal representation are inconsistent. Section 4.1.2, Equation (3), defines patch-level temporal aggregation through masked attention pooling with a shared learnable query. However, Section 5.4 describes the full model as concatenating timestamp embeddings and applying a linear projection, with averaging used as the ablated alternative. These are different architectures, making it unclear which method produced the reported results and what the ablation establishes. Separately, the comparison in Section 5.5 changes both the attention mask and the available context: the causal setting uses channel-independent inputs, whereas the hybrid setting uses concatenated multivariate context. Consequently, this comparison does not isolate the contribution of the hybrid mask. Together, these issues weaken the evidence connecting the proposed mechanisms to the reported improvements.
Questions for authors
-
Could you provide a corrected Table 1, confirm the dataset assignment of every result, and identify which rankings and conclusions change? Please also confirm whether the dataset splits, observation/forecast windows, preprocessing, and metric aggregation used for UniTIMS and the other foundation models match the cited full-shot baseline protocol, and identify any differences.
-
Which temporal aggregation was actually used for the main experiments: Equation (3)'s attention pooling or the concatenation-and-projection operation described in Section 5.4? Were the ablations evaluated against that same implementation? For the Section 5.5 mask comparison, are results available with identical multivariate inputs and training settings, changing only the mask?
-
Could you provide the expanded VersaTSA source manifest and observation-count definition underlying the 231B figure? The LOTSA rows listed in Table 3 sum to approximately 27.65B observations, so the additional expansion is not accounted for by that table. Please also identify the fev-bench version and clarify its exclusion protocol: sources such as M5, Favorita, and KDD Cup 2022 appear in both the pretraining inventory and fev-bench. Section 4.4.2 explicitly addresses the nine other evaluation datasets, but it is unclear how the fev-bench tasks were excluded. A source-, series-, and time-range-level account would help distinguish legitimate disjoint subsets from evaluation overlap; shared source names alone do not establish leakage.