TimeXer is a novel approach designed to enhance time series forecasting with exogenous variables using deep learning models. The paper addresses the challenge of incorporating external information to improve the accuracy of endogenous variable forecasts, which are often influenced by various external factors. Unlike existing methods that either treat all variables equally or ignore exogenous information, TimeXer leverages the canonical Transformer architecture to simultaneously capture both endogenous and exogenous temporal dependencies. The key contributions of TimeXer include: 1. **Model Design**: TimeXer employs patch-wise self-attention and variate-wise cross-attention to capture temporal and multivariate correlations. It introduces learnable global tokens to bridge the causal information from exogenous series to endogenous temporal patches. 2. **Performance**: Extensive experiments on twelve real-world datasets demonstrate that TimeXer achieves state-of-the-art performance in both univariate and multivariate forecasting tasks. It effectively utilizes exogenous information to enhance endogenous forecasts, showing notable generality and scalability. 3. **Scalability**: TimeXer is capable of handling large-scale time series data, as demonstrated by its performance on a weather dataset with 3,850 stations and 36 exogenous meteorological variables. 4. **Ablation Study**: Detailed ablation studies validate the effectiveness of TimeXer's design, showing that the proposed embedding and attention mechanisms are crucial for achieving superior performance. 5. **Generality**: TimeXer demonstrates robustness in handling scenarios with missing values and varying look-back lengths, indicating its adaptability to real-world challenges. Overall, TimeXer provides a robust and efficient solution for time series forecasting with exogenous variables, making it a valuable tool for practical applications in various domains.TimeXer is a novel approach designed to enhance time series forecasting with exogenous variables using deep learning models. The paper addresses the challenge of incorporating external information to improve the accuracy of endogenous variable forecasts, which are often influenced by various external factors. Unlike existing methods that either treat all variables equally or ignore exogenous information, TimeXer leverages the canonical Transformer architecture to simultaneously capture both endogenous and exogenous temporal dependencies. The key contributions of TimeXer include: 1. **Model Design**: TimeXer employs patch-wise self-attention and variate-wise cross-attention to capture temporal and multivariate correlations. It introduces learnable global tokens to bridge the causal information from exogenous series to endogenous temporal patches. 2. **Performance**: Extensive experiments on twelve real-world datasets demonstrate that TimeXer achieves state-of-the-art performance in both univariate and multivariate forecasting tasks. It effectively utilizes exogenous information to enhance endogenous forecasts, showing notable generality and scalability. 3. **Scalability**: TimeXer is capable of handling large-scale time series data, as demonstrated by its performance on a weather dataset with 3,850 stations and 36 exogenous meteorological variables. 4. **Ablation Study**: Detailed ablation studies validate the effectiveness of TimeXer's design, showing that the proposed embedding and attention mechanisms are crucial for achieving superior performance. 5. **Generality**: TimeXer demonstrates robustness in handling scenarios with missing values and varying look-back lengths, indicating its adaptability to real-world challenges. Overall, TimeXer provides a robust and efficient solution for time series forecasting with exogenous variables, making it a valuable tool for practical applications in various domains.