Our results show that DLinear outperforms existing complex Transformer-based models in most cases by a large margin. In particular, for the Exchange-Rate dataset that does not have obvious periodicity, the prediction errors of the state-of-the-art method zhou2022fedformer are more than twice larger than those of DLinear.
Because DLinear is a linear model, its weights can directly reveal how DLinear works. Figure 4(a) and (b) visualize the weights of the trend and the remainder layers on the Exchange-Rate dataset. Due to the lack of periodicity and seasonality in financial data, it is hard to observe clear patterns, but the trend layer reveals greater weights of information closer to the outputs, representing their larger contributions to the predicted values.
https://www.arxiv-vanity.com/papers/2205.13504/