TimeMosaic: Temporal Heterogeneity Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-wise Decoding

On November 12, 2025, the TimeMosaic framework was published, marking a notable advancement in multivariate time series forecasting. Traditional methods often struggle with fixed-length segmentation, which can overlook the complexities of local temporal dynamics. TimeMosaic addresses this by employing adaptive patch embedding, allowing it to adjust granularity based on local information density. Additionally, its segment-wise decoding treats each prediction horizon as a distinct subtask, enhancing accuracy across different time scales. Extensive evaluations on benchmark datasets demonstrate that TimeMosaic delivers consistent improvements over existing forecasting methods, achieving performance competitive with state-of-the-art time series forecasting models. This innovation is particularly relevant for critical domains such as finance, transportation, climate, and energy, where accurate forecasting is essential.