Time series forecasting is a widely used technique in various applications such as traffic planning and weather forecasting. However, real-world time series often exhibit intricate temporal variations, making forecasting challenging. This paper introduces *TimeMixer*, a novel architecture that addresses this issue by leveraging a multiscale-mixing approach. Unlike traditional methods that focus on plain decomposition or multiperiodicity analysis, TimeMixer decomposes time series into distinct patterns at different sampling scales, capturing both fine-grained seasonal and coarse-grained trend components. The architecture consists of *Past-Decomposable-Mixing* (PDM) and *Future-Multipredictor-Mixing* (FMM) blocks. PDM mixes the decomposed seasonal and trend components in a fine-to-coarse and coarse-to-fine direction, respectively, while FMM ensemble multiple predictors to utilize complementary forecasting capabilities from multiscale observations. This approach enables TimeMixer to achieve state-of-the-art performance in both long-term and short-term forecasting tasks with superior efficiency. The paper also includes extensive experiments on various benchmarks, demonstrating the effectiveness and efficiency of TimeMixer.Time series forecasting is a widely used technique in various applications such as traffic planning and weather forecasting. However, real-world time series often exhibit intricate temporal variations, making forecasting challenging. This paper introduces *TimeMixer*, a novel architecture that addresses this issue by leveraging a multiscale-mixing approach. Unlike traditional methods that focus on plain decomposition or multiperiodicity analysis, TimeMixer decomposes time series into distinct patterns at different sampling scales, capturing both fine-grained seasonal and coarse-grained trend components. The architecture consists of *Past-Decomposable-Mixing* (PDM) and *Future-Multipredictor-Mixing* (FMM) blocks. PDM mixes the decomposed seasonal and trend components in a fine-to-coarse and coarse-to-fine direction, respectively, while FMM ensemble multiple predictors to utilize complementary forecasting capabilities from multiscale observations. This approach enables TimeMixer to achieve state-of-the-art performance in both long-term and short-term forecasting tasks with superior efficiency. The paper also includes extensive experiments on various benchmarks, demonstrating the effectiveness and efficiency of TimeMixer.