The paper presents an approach to time series forecasting, namely the Adversarial Sparse Transformer (AST), which adopts Transformer for generating a sparse attention map over time series signals and a generative model (GAN) for adversarial regularization in forecasting. Extensive experiments on several real-world datasets show the effectiveness and efficiency of AST. Pros: - The problem area is important and interesting. - The approach is technically sounding. Cons: - The novelty is limited, or not well justified, as sparse Transformer and GAN for time series farecasting has been already studied by some recent work (e.g., timeGAN and more). More discussions and empirical comparison with those competing methods would make this paper stronger. - The code is provided, making it difficult to replicate the work by others. - The clarify of some technical descriptions and the English writing have rooms for improvement. Overall, the work is not outstanding paper but OK to accept.