acceptodds
Withdrawn from ICLR 2027

When and Where to Transfer: Learning Transferable Historical Contexts for Short-History Multivariate Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Related sources can improve multivariate time series forecasting when the target history is short, but their usefulness varies across source groups and historical periods. A key challenge is to identify informative historical contexts from the sources. This paper proposes Hierarchical Context Transfer (HiCT), a framework that adaptively weights source groups and their historical contexts to determine both where and when to transfer. A shared context encoder and cross-attention are used to learn weights over source groups and historical contexts. These weights guide regularized transfer estimation, followed by an optional target-specific correction. This paper establishes a target excess-risk bound for the context-weighted transfer estimator. Experiments demonstrate competitive predictive performance across four multivariate forecasting datasets and a favorable accuracy–runtime trade-off against the evaluated baselines. Source-corruption ablations illustrate the roles of group-level and within-group weighting in adapting to unreliable source histories.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time series forecasting, Transfer learning, Cross-attention, Hierarchical context weighting

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

HiCT addresses forecasting with a short target history by learning target-conditioned hierarchical weights over source groups and historical contexts, followed by weighted predictor fitting and optional target-specific correction.

Critical strengths

The short-history transfer problem is practically meaningful, and separating source-level and context-level transferability is intuitive.

Critical weaknesses
  1. The experimental scope is too narrow to support a general MTS forecasting contribution. Only four datasets are evaluated, and the main forecasting setup is one-step-ahead prediction. This is substantially narrower than standard multistep MTS forecasting evaluation.
  2. The advantage is confounded by the rolling-refit protocol. At test time, HiCT regenerates context weights and repeatedly refits prediction parameters as new target observations become available. Several neural baselines are not given an equivalent online adaptation opportunity. Consequently, the performance gain cannot be cleanly attributed to hierarchical transferability weighting.
  3. The core hierarchical weighting mechanism is insufficiently isolated. The source-corruption experiments are mostly conducted on Stock, and the variants replace learned weights while retaining already fitted scorers/predictors rather than retraining each restricted model independently.
  4. The method is evaluated against a limited set of forecasting architectures and does not establish relevance to recent strong transfer/foundation forecasting models. For an ICLR 2027 submission, comparison with only PatchTST, TimeMixer, AdaRNN, DTr-CNN, and ROSE is not sufficient.
Questions for authors

Does HiCT retain its advantage if all competing methods receive the same rolling observations and online refitting budget? Can the authors evaluate multi-step forecasting and more standard MTS benchmarks? What is the performance of independently retrained group-only and context-only variants?

Reviewer_PMmVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studied multivariate time series forecasting for target variate groups with short historical records by leveraging related source groups with longer histories. The authors propose Hierarchical Context Transfer (HiCT), which uses a shared Transformer encoder and cross-attention to learn target-conditioned weights over source groups and historical contexts. These weights guide a regularized transfer estimator, followed by an optional target-specific correction. The paper conducted experiments on open multivariate datasets.

Critical strengths
  1. The paper is clear and easy to follow.
  2. short-history forecasting via other hierarchical context transfer is practical and reasonable in time-series analysis.
Critical weaknesses
  1. Questionable Alignment Assumptions: The method relies on Procrustes analysis (orthogonal matrix estimation) to align PCA-reduced latent spaces of different source and target groups. The assumption that latent coordinates of distinct multivariate time series are "approximately comparable up to rotation and reflection" is extremely strong and likely violated in complex, real-world datasets (e.g., non-linear and non-stationary). This rigid geometric alignment severely limits the generalizability.
  2. Weak Forecasting Baselines and Architectures: The primary forecasting backbone used in HiCT is a simple linear Vector Autoregressive (VAR) model or a basic RNN. The paper fails to compare against modern, highly optimized short-history or few-shot forecasting foundation models (e.g., TOTO-2, PatchTST, or Moirai, LightGTS, in zero-shot/few-shot/transfer settings). Demonstrating superiority only over older or less capable baselines weakens the claim of performance.
  3. Limited Methodological Novelty and nothing new from this paper: short-history forecasting via other hierarchical context transfer is well studied in time-series analysis.
Questions for authors

See Weaknesses

Reviewer_T2jdOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies time series forecasting when training samples are scarce but external data are abundant. It designs a hierarchical context transfer framework to guide joint training on external and target data. The research problem is practically meaningful, especially for settings with limited historical data or cold starts.

Critical strengths

1, The research question is meaningful. How to effectively leverage external data when historical data are insufficient is an important problem. The paper further considers differences in transferability across different periods from the same source, which is more fine-grained than mere data selection. 2, Weight learning is directly tied to target forecasting performance. The method trains a weight generator using target validation error, so that external data selection accounts for the forecasting performance after weighted fitting rather than relying only on predefined distances or similarities. 3, The experimental results are reasonably compelling. HiCT achieves the best error on three of four datasets, and the runtime on Stock also shows good computational efficiency.

Critical weaknesses

1, The baselines are insufficient. On the one hand, the authors should compare with conventional time series transfer learning approaches, such as training on external data and then fine-tuning on the target domain. On the other hand, they should compare with time series foundation models pretrained on large-scale data, such as Chronos-2, which also have strong few-shot capabilities. 2, The ablation study is insufficient. The authors should compare training on target data alone, uniform source weighting, simple similarity weighting, and the full HiCT.

Questions for authors

1, How demanding is the alignment step with respect to the shared observed history? When the shared history is very short, the underlying relationship is not an orthogonal transformation, or there is no temporal overlap, does the method still apply? Can the authors provide an alignment ablation or a sensitivity analysis with respect to overlap length? 2, Can the authors add comparisons with publicly available pretrained time series foundation models such as Chronos-2?

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

Related papers

Discussion (0)

Sign in to comment.