Continuous Evolution Pool: Taming Recurring Concept Drift in Online Time Series Forecasting

TL;DR

The paper introduces Continuous Evolution Pool (CEP), a privacy-preserving framework for online time series forecasting under recurring concept drift and delayed feedback. CEP maintains a dynamic pool of specialized forecasters represented by lightweight statistical genes, using nearest retrieval, concept-driven evolution, and systematic elimination to manage knowledge while never storing raw historical samples. By decoupling concept identification from forecasting, CEP converts a non-stationary learning problem into multiple stationary subproblems, yielding sublinear regret and strong empirical gains (often over 20% reduction in forecasting error) across real-world datasets and backbones. The approach demonstrates robust performance with tight memory constraints, suitability for edge deployment, and principled hyperparameter guidance, offering a practical solution for privacy-conscious, drift-prone time-series applications.

Abstract

Recurring concept drift poses a dual challenge in online time series forecasting: mitigating catastrophic forgetting while adhering to strict privacy constraints that prevent retaining historical data. Existing approaches predominantly rely on parameter updates or experience replay, which inevitably suffer from knowledge overwriting or privacy risks. To address this, we propose the Continuous Evolution Pool (CEP), a privacy-preserving framework that maintains a dynamic pool of specialized forecasters. Instead of storing raw samples, CEP utilizes lightweight statistical genes to decouple concept identification from forecasting. Specifically, it employs a Retrieval mechanism to identify the nearest concept based on gene similarity, an Evolution strategy to spawn new forecasters upon detecting distribution shifts, and an Elimination policy to prune obsolete models under memory constraints. Experiments on real-world datasets demonstrate that CEP significantly outperforms state-of-the-art baselines, reducing forecasting error by over 20% without accessing historical ground truth.

Figures (10)

• Figure 1: Concept recurrence is a notable phenomenon in time series data. For instance, in (a), the second concept recurs following the emergence of the fourth concept. During online updates and concept drifts, forecasters often suffer from forgetting previously learned concepts, as in (b). To address this challenge and effectively track recurring concepts, the Continuous Evolution Pool employs an evolving and retrieval strategy, assigning distinct forecasters to manage different concepts.
• Figure 2: Illustration of the proposed CEP mechanism: (a) Adaptive Updating: Continuously update the position of the evolved forecaster in the gene space as new instances arrive. (b) Nearest Evolution: If the instance surpasses the forecaster's evolution threshold, evolve and replicate the closest forecaster at the instance. (c) Optimizer Adjustment: Adjust the learning rate for the shifted concept to ensure accurate adaptation. (d) Forecaster Elimination: The forecaster $f_4$ associated with the rarely-occurring noise concept may be removed due to prolonged inactivity. (e) Nearest Retrieval: Identify the nearest forecaster in the gene space when an input instance is encountered.
• Figure 3: Delayed feedback setting. The blue region represents the input, while the yellow region denotes the corresponding ground truth. In the previous online forecasting setting, the model knows part of the ground truth for the next sample beforehand. This situation is considered unfair. In the real world, the model must predict the future values of the ground truth. Therefore, a delay setting is used here.
• Figure 4: Visualization of forecasting results
• Figure 5: Visualization of Metric