TimeMixer is a novel architecture for time series forecasting that leverages multiscale mixing to disentangle complex temporal variations. The model uses Past-Decomposable-Mixing (PDM) and Future-Multipredictor-Mixing (FMM) blocks to extract past information and predict future values. PDM decomposes time series into seasonal and trend components at different scales and mixes them in both fine-to-coarse and coarse-to-fine directions. FMM ensembles multiple predictors to utilize complementary forecasting capabilities from multiscale observations. TimeMixer achieves state-of-the-art performance in both long-term and short-term forecasting tasks with favorable run-time efficiency. The model is evaluated on 18 real-world benchmarks and 15 baselines, demonstrating its effectiveness in handling complex temporal patterns. The results show that TimeMixer outperforms existing models in terms of forecasting accuracy and efficiency. The model's architecture is fully MLP-based, enabling efficient computation and scalability. The paper also provides detailed analysis of the model's components, including ablation studies and visualization of temporal linear weights for seasonal and trend mixing. The results indicate that the multiscale mixing approach is effective in capturing complex temporal variations and improving forecasting performance.TimeMixer is a novel architecture for time series forecasting that leverages multiscale mixing to disentangle complex temporal variations. The model uses Past-Decomposable-Mixing (PDM) and Future-Multipredictor-Mixing (FMM) blocks to extract past information and predict future values. PDM decomposes time series into seasonal and trend components at different scales and mixes them in both fine-to-coarse and coarse-to-fine directions. FMM ensembles multiple predictors to utilize complementary forecasting capabilities from multiscale observations. TimeMixer achieves state-of-the-art performance in both long-term and short-term forecasting tasks with favorable run-time efficiency. The model is evaluated on 18 real-world benchmarks and 15 baselines, demonstrating its effectiveness in handling complex temporal patterns. The results show that TimeMixer outperforms existing models in terms of forecasting accuracy and efficiency. The model's architecture is fully MLP-based, enabling efficient computation and scalability. The paper also provides detailed analysis of the model's components, including ablation studies and visualization of temporal linear weights for seasonal and trend mixing. The results indicate that the multiscale mixing approach is effective in capturing complex temporal variations and improving forecasting performance.