Adapt Data to Model: Adaptive Transformation Optimization for Domain-shared Time Series Foundation Models
Yunzhong Qiu, Zhiyao Cen, Zhongyi Pei, Chen Wang, Jianmin Wang
TL;DR
This paper introduces TATO, a data-centric framework that adaptively configures transformation pipelines to enable a single pre-trained large time series model to effectively generalize across diverse domains, significantly improving forecasting accuracy.
Contribution
TATO is a novel adaptive transformation optimization method that enhances domain adaptation of large time series models without fine-tuning, using a robust, efficient pipeline configuration approach.
Findings
Achieves up to 65.4% reduction in MSE on diverse datasets.
Completes optimization in under 2 minutes, suitable for real-world use.
Significantly improves forecasting accuracy across multiple domains.
Abstract
Large time series models (LTMs) have emerged as powerful tools for universal forecasting, yet they often struggle with the inherent diversity and nonstationarity of real-world time series data, leading to an unsatisfactory trade-off between forecasting accuracy and generalization. Rather than continually finetuning new LTM instances for each domain, we propose a data-centric framework, time-series adaptive transformation optimization (TATO), that enables a single frozen pre-trained LTM to adapt to diverse downstream domains through an optimally configured transformation pipeline. Specifically, TATO constructs three representative types of transformations, including context slicing, scale normalization, and outlier correction, to help LTMs better align with target domain characteristics. To ensure robustness, we incorporate carefully selected time series augmentations and a two-stage…
Peer Reviews
Decision·ICLR 2026 Poster
The authors provide sufficient ablation studies and comparisons, which enhance the credibility of the findings.
1. The writing is deficient, like the transition from paragraph in line 55 to 57 feels abrupt; Section 2.3 is a seemingly extraneous point, confusing the reader and disrupting the narrative flow; Section 3.1 is repetitive because the same content has been discussed before; In figure 2, it seems like the order of 3 types of operators is reorderable, but it's said in the text that they are fixed, it's confusing which ones are reorderable; In Section 3.2.2, post-process operators are mentioned, but
The paper presents an original paradigm that shifts the focus from model adaptation to data adaptation, introducing the concept of FrozenForecasting for large time series models. The proposed TATO framework is well designed, combining hyperparameter optimization and Pareto based ranking to ensure both robustness and efficiency. The methodology is clearly described with experiments across several state of the art LTMs and diverse datasets. The work provides some evidence of performance improvemen
While the idea of adapting data instead of models is novel, the paper could benefit from a deeper theoretical justification of why certain transformations consistently enhance generalization across domains. The search space for transformation pipelines, though compact, may still be computationally demanding for larger datasets or longer horizons, and scalability analyses beyond 500 samples are limited. Some test-time adaptation methods for time-series forecasting were completely ignored by the a
- The idea of leveraging transformation pipelines to align time series data with forecasting models is novel and promising, as it improves domain-adaptive forecasting performance while keeping the backbone models fixed. - The proposed pipeline optimization process is efficient, typically requiring less than two minutes in most cases. - Experimental results demonstrate that the proposed framework effectively boosts the performance of various backbone models on widely used benchmark datasets.
- Clarity and presentation: - The framework description lacks sufficient detail. For instance, how are individual trials sampled from the pipeline space? Providing pseudocode or an algorithmic overview would greatly improve readability and reproducibility. - It is mentioned that only a portion of the training samples are used in the TATO framework, but the selection strategy for these samples is unclear. A detailed explanation of how these samples are chosen from the full training set would
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsTime Series Analysis and Forecasting · Machine Learning in Healthcare · Stock Market Forecasting Methods