Out-of-Distribution Generalization in Time Series: A Survey

TL;DR

Time series data are inherently non-stationary, causing distribution shifts that challenge OOD generalization. The paper formalizes TS-OOD with fixed-context and cross-domain settings and surveys methods across data distribution, representation learning, and OOD evaluation, including forward-looking objectives such as $\min_{\theta} \mathbb{E}_{(X,Y) \sim \mathbb{P}_{test}} \left[ \sum_{t=T+1}^{T+\tau} \mathcal{L}(f_{\theta}(X_{t-\Delta:t-1}), Y_t) \right]$. It contributes a three-dimensional taxonomy, a survey of decoupling/invariant/adaptive/LTSM approaches, and an open-source codebase to support reproducibility. The findings inform robust TS models for real-world deployment across domains like transportation, environment, and public health, guiding future research toward invariant representations, multimodal pre-training, and rigorous OOD evaluation.

Abstract

Time series frequently manifest distribution shifts, diverse latent features, and non-stationary learning dynamics, particularly in open and evolving environments. These characteristics pose significant challenges for out-of-distribution (OOD) generalization. While substantial progress has been made, a systematic synthesis of advancements remains lacking. To address this gap, we present the first comprehensive review of OOD generalization methodologies for time series, organized to delineate the field's evolutionary trajectory and contemporary research landscape. We organize our analysis across three foundational dimensions: data distribution, representation learning, and OOD evaluation. For each dimension, we present several popular algorithms in detail. Furthermore, we highlight key application scenarios, emphasizing their real-world impact. Finally, we identify persistent challenges and propose future research directions. A detailed summary of the methods reviewed for the generalization of OOD in time series can be accessed at https://tsood-generalization.com.

Figures (7)

• Figure 1: Examples illustrating how real-world temporal dynamics drive distribution shifts in TS-OOG. The first two columns show covariate shift, where the input feature distribution ($\mathbb{P}(X)$) changes due to underlying dynamics like gradual temporal drifts (e.g., user growth) or abrupt events (e.g., server upgrade). The third column depicts concept drift, a more fundamental change where the input-output relationship ($\mathbb{P}(Y|X)$) itself is altered (e.g., by an algorithm change affecting user engagement). In the social media example, both user growth (temporal drift) and server upgrade (event) cause covariate shifts affecting the input data distribution. Subsequently, the algorithm change induces concept drift by fundamentally altering the relationship between content and engagement. Blue-shaded areas represent historical data; green areas show subsequent OOD data. This highlights that TS-OOG problems stem from temporal dynamics altering either the data's characteristics ($\mathbb{P}(X)$) or its underlying generative rules ($\mathbb{P}(Y|X)$).
• Figure 2: A research roadmap of representative methods for TS-OOG. The methods are categorized along three dimensions: data distribution assumptions (indicated by border color), representation learning strategies (indicated by line style), and OOD evaluation protocols (indicated by block color). Methods within the roadmap are presented in chronological order. For a more comprehensive summary, please refer to the main text.
• Figure 3: A comprehensive taxonomy of TS-OOG methods.
• Figure 4: A taxonomy examining methods for TS-OOG from three perspectives: data distribution, representation learning, and OOD evaluation. This includes various approaches like covariate shift adaptation, invariant learning, adaptive mechanism-based methods, and large time series models (LTSMs). Note that OOD evaluation specifically for LTSMs is not summarized here due to the current lack of standardized evaluation protocols.
• Figure 5: Each circle represents a data instance, with different colors indicating the class to which the instance belongs. (a). Original data: The distribution of instances and the classification boundary (dashed line) remain stable. (b). Covariate shift: The distribution of instances changes (the feature distribution $\mathbb{P} (X)$ changes), but the class decision boundary $\mathbb{P} (Y|X)$ remains unchanged. (c). Concept drift: The class decision boundary $\mathbb{P} (Y|X)$ changes, but the distribution of instances (the feature distribution $\mathbb{P} (X)$) remains unchanged.