Generating Time Series by Using Latent Space

Time series forecasting is a crucial aspect of analyzing time series data, enabling predictions about future trends. Deep learning methods, particularly the transformer model, have become popular in time series forecasting. However, most existing models are discriminative, focusing on the relationship between past and future values. In contrast, time series data is generated from a high-dimensional latent space. This paper introduces LaTrans, a novel transformer-based time series forecasting model. Unlike previous models, LaTrans leverages the concept of latent space, where future time series can be generated. The model combines the power of latent space and transformer architectures, using attention layers to extract probability distributions and compress them into the latent space. Furthermore, it demonstrates the importance of the latent space in time series forecasting and shows that future time series can be generated within this space. The paper compares LaTrans with other transformer-based methods and investigates the influence of the KL divergence weight on forecasting results. These findings contribute to advancing the field of time series forecasting, highlighting the benefits of incorporating latent space and providing a new model that outperforms existing transformer-based approaches.