I am reading a book about keras and the author introduces new deep learning techniques and choose the problem of estimating the temperature from local data to compare performance of those techniques. The baseline is predicting tomorrow temperature by today temperature and no wonder deep learning cannot do any better. So how are you going to compare deep learning methods solving a problem that cannot be solved or approximated in any way? I am really disappointed, the author seems to be a very bright person, so I cannot understand why he didn't choose better applications to illustrate the power of the different methods. There must be something that I don't get, perhaps in a rush the author decides to finish the book no matter what way?