Time-Series-Library

I've been working by my own on some personal projects on value prediction of multivariate mixed-frequency data, that is, thousands of values recorded about power generation and many exogenous variables for high frequencies used to get an estimation of a set of target variables some hours in advance. And so I've stumbled upon the world of time series forecasting with deep learning [1]. First of all: do enterprises (if they actually use deep learning, because it kind of works in my experience) retrain the neural network at each step with the new row of data? And, if so, do they take all data as the training set, and validate nothing? Second: do they use some kernel for the loss in order to penalize errors of the model in more recent data? That would seem extremely logical to me. Third: which arquitecture is it often used? I've been using LS

>Compared to MLPs, transformers save on parameter count by skimping on the number of parameters That is only correct if you look at models with equal parameter count from a purely theoretical perspective. In practice, it is possible to train transformers to orders of magnitude bigger scales than MLPs because they are so much more efficient. That's why I said a modern transformer will easily beat these puny modern MLPs, but only in cases where data and compute budgets allow it. That is not even a question. If you look at recent time series forecasting leaderboard entries, you'll almost always see transformers playing along at the top of it: https://github.com/thuml/Time-Series-Library

Cool! I find it has been implemented in the tslib (https://github.com/thuml/Time-Series-Library), the results seem promising when I reproduce the experiments.