This paper proposes a Latent Diffusion Transformer (LDT) for probabilistic time series forecasting. The LDT model is designed to address the challenges of high-dimensional multivariate time series forecasting by transforming the problem into a latent space time series generation task. The model consists of two main components: a symmetric statistics-aware autoencoder and a diffusion-based conditional generator. The autoencoder compresses multivariate time series into a latent representation by considering dynamic statistics, while the diffusion-based generator efficiently generates realistic time series values in a continuous latent space using a self-conditioning guidance mechanism. The LDT model is trained on various real-world datasets, including energy, traffic, and taxi data, and achieves state-of-the-art performance in high-dimensional multivariate time series forecasting. The model's key contributions include the introduction of the LDT framework, the development of a practical LDT structure with a self-conditioning mechanism and a non-autoregressive transformer, and extensive experiments demonstrating LDT's superior performance compared to existing methods in probabilistic time series forecasting. The LDT model uses a non-autoregressive denoising network to improve forecasting efficiency and accuracy. The model's self-conditioning guidance mechanism allows for more accurate and stable generation of time series data by incorporating relevant covariates. The model's performance is evaluated using metrics such as CRPS-sum and MSE, and it outperforms existing methods on multiple datasets. The LDT model is also shown to be effective in both deterministic and uncertainty estimation tasks, demonstrating its versatility in different forecasting scenarios. The model's ability to adapt to different forecasting scenarios by adjusting guidance strength is also highlighted. Overall, the LDT model provides a novel and effective approach to high-dimensional multivariate time series forecasting.This paper proposes a Latent Diffusion Transformer (LDT) for probabilistic time series forecasting. The LDT model is designed to address the challenges of high-dimensional multivariate time series forecasting by transforming the problem into a latent space time series generation task. The model consists of two main components: a symmetric statistics-aware autoencoder and a diffusion-based conditional generator. The autoencoder compresses multivariate time series into a latent representation by considering dynamic statistics, while the diffusion-based generator efficiently generates realistic time series values in a continuous latent space using a self-conditioning guidance mechanism. The LDT model is trained on various real-world datasets, including energy, traffic, and taxi data, and achieves state-of-the-art performance in high-dimensional multivariate time series forecasting. The model's key contributions include the introduction of the LDT framework, the development of a practical LDT structure with a self-conditioning mechanism and a non-autoregressive transformer, and extensive experiments demonstrating LDT's superior performance compared to existing methods in probabilistic time series forecasting. The LDT model uses a non-autoregressive denoising network to improve forecasting efficiency and accuracy. The model's self-conditioning guidance mechanism allows for more accurate and stable generation of time series data by incorporating relevant covariates. The model's performance is evaluated using metrics such as CRPS-sum and MSE, and it outperforms existing methods on multiple datasets. The LDT model is also shown to be effective in both deterministic and uncertainty estimation tasks, demonstrating its versatility in different forecasting scenarios. The model's ability to adapt to different forecasting scenarios by adjusting guidance strength is also highlighted. Overall, the LDT model provides a novel and effective approach to high-dimensional multivariate time series forecasting.