about normalization and sliding windows about autoformer HOT 2 CLOSED

Thanks for the interest.

1. As stated in the previous issue, we keep this pre-processing setting as previous baselines. Here are the reasons:
   (1) We adopt this for a fair comparison with the previous models. Since the train set is the same among different models, I think our evaluation is fair.
   (2) I think comparing within the real-data scale is also accepted. However, I have to point out, for multivariate data, if one dimension's scale is [0,1] and another is [0,100], the metric can be dominated by the dimension with the larger value scale. Or you can adopt other metrics, such as the MAPE or MSPE in https://github.com/thuml/Autoformer/blob/main/utils/metrics.py.

2. Quickly over the entire set is ok. Both methods can obtain a fair comparison. But for real-world applications, the rule is to use the real-time data as much as possible. Thus, if you have enough computation resources, updating the forecasting results as quickly as possible is better.

from autoformer.

Wow, thank you for your prompt reply. That's very clear now.

from autoformer.