Dynamic Multi-period Experts for Online Time Series Forecasting

Authors: Seungha Hong, Sukang Chae, Suyeon Kim, Sanghwan Jang, Hwanjo Yu Date: 2026-03-10 Paper ID: openalex:2603.09062

Summary

This paper addresses the limitation of existing Online Time Series Forecasting (OTSF) methods by categorizing concept drift into Recurring Drift and Emergent Drift. The authors introduce DynaME (Dynamic Multi-period Experts), a hybrid framework specifically designed to manage this dual nature of drift. DynaME employs a committee of specialized experts that adaptively fit to recurring historical periodic patterns, while simultaneously utilizing a stable, general expert when high uncertainty signals the emergence of novel patterns. Experiments on standard benchmarks confirm that DynaME effectively adapts to complex drift scenarios and establishes a new state-of-the-art performance.

Key Contributions

• Redefined concept drift in Online Time Series Forecasting (OTSF) into two types: Recurring Drift and Emergent Drift.
• Proposed DynaME (Dynamic Multi-period Experts), a novel hybrid framework to handle the dual nature of concept drift in OTSF.
• For Recurring Drift, DynaME uses a committee of experts dynamically fitted to relevant historical periodic patterns.
• For Emergent Drift, DynaME detects high uncertainty and switches reliance to a stable, general expert.
• Demonstrated significant performance improvement over existing baselines across multiple benchmark datasets and backbones in OTSF tasks.

Limitations

The abstract does not explicitly detail limitations, but a potential area for future work is characterizing the complexity of managing the expert committee under extreme drift scenarios.

Open Questions & Future Work

• extending-dynamme-to-foundation-models

Key Concepts

• Dynamic Multi-period Experts (DynaME): A hybrid online time series forecasting framework that uses specialized experts for recurring drift and a general expert for emergent drift.

Datasets

• benchmark datasets

Limitations

The abstract does not explicitly detail limitations, but a potential area for future work is characterizing the complexity of managing the expert committee under extreme drift scenarios.

Links

• Abstract
• PDF