Forecasting multiple time series that share underlying dynamics but also have unique behaviors is a known difficulty, especially when each series comes with many covariates. A new paper on arXiv introduces Hopformer, a transformer-based approach designed to tackle this problem directly.

The model's name reflects its core idea: homogeneity pursuit. Rather than forcing all series into one pattern or treating each in isolation, Hopformer aims to extract what is common across series while preserving the meaningful differences that matter for accurate forecasts.

The abstract is brief, so details on the architecture and experiments are not yet available from this source. Still, the framing points to a practical trade-off between global and local information, a central issue in multi-series forecasting.