[dil8]

> If you want to predict (multiple) time series using multiple series as input/predictors, that's a whole new level of difficulty. I don't know of a good automatic/fast/scalable approach that properly guards against overfitting

Have you had a look at algorithms contained in pytorch forcasting? https://pytorch-forecasting.readthedocs.io/en/latest/

[em500]

We only did a few internal NN and LSTM implementations in the past, we should probably evaluate the new pytorch stuff soon. But as you can imagine a lot of our time was consumed by modelling pandemic-induced dynamics (which is especially at longer forecast horizons are much more driven by assumptions rather than by data/models).

[microprediction]

I'll try to add notebook examples at https://www.microprediction.com/blog/popular-timeseries-pack... and get some pytorch-forecasting Elo ratings going. As an aside, anyone who wants to see a particular approach get rated is welcome to help! It amounts to creating a "skater" which is a simple functional form of forecasting. https://github.com/microprediction/timemachines/tree/main/ti...