Tackling Time-Series Forecasting Generalization via Mitigating Concept Drift

Time-series forecasting finds broad applications in real-world scenarios. Due to the dynamic nature of time series data, it is important for time-series forecasting models to handle potential distribution shifts over time. In this paper, we initially identify two types of distribution shifts in time series: concept drift and temporal shift. We acknowledge that while existing studies primarily focus on addressing temporal shift issues in time series forecasting, designing proper concept drift methods for time series forecasting has received comparatively less attention. Motivated by the need to address potential concept drift, while conventional concept drift methods via invariant learning face certain challenges in time-series forecasting, we propose a soft attention mechanism that finds invariant patterns from both lookback and horizon time series. Additionally, we emphasize the critical importance of mitigating temporal shifts as a preliminary to addressing concept drift. In this context, we introduce ShifTS, a method-agnostic framework designed to tackle temporal shift first and then concept drift within a unified approach. Extensive experiments demonstrate the efficacy of ShifTS in consistently enhancing the forecasting accuracy of agnostic models across multiple datasets, and outperforming existing concept drift, temporal shift, and combined baselines.

💡 Research Summary

Time‑series forecasting is indispensable in domains such as finance, transportation, and epidemiology, yet the dynamic nature of temporal data inevitably leads to distribution shifts. The authors distinguish two fundamental types of shift: Temporal Shift (changes in marginal distributions while conditional distributions remain stable) and Concept Drift (changes in the conditional distribution P(Y | X) over time). While prior work has largely focused on mitigating Temporal Shift through normalization or adaptive preprocessing, systematic treatment of Concept Drift in standard (offline) forecasting settings has been largely absent.

The paper introduces a two‑stage, model‑agnostic framework called ShifTS (Shift‑then‑Concept‑drift). The first stage addresses Temporal Shift by normalizing both input and target series to a standard Gaussian (zero mean, unit variance) before feeding them to any forecasting model, and denormalizing the predictions afterward. This ensures that marginal distributions are fixed, allowing subsequent components to focus on conditional relationships.

The core contribution for Concept Drift mitigation is Soft Attention Masking (SAM). The authors argue that relying solely on the look‑back window X_L to predict the horizon Y_H can be insufficient because future exogenous variables X_H may contain causal cues that stabilize the conditional distribution. Directly modeling P(Y_H | X_L, X_H) is infeasible at test time because X_H is unknown. SAM circumvents this by concatenating the full series