Hi All, currently I am working on project where I am forecasting demand for the next two weeks. We do have ~1000 Outlets with ~1000 products that are separted in around 50 groups. Three years of data are around 200M rows. Currently, we are using two different strategies (1) gradient boosting with all data and (2) gradient boosting per group. Both approaches work reasonably well. However, now we would also like to explore if we can gain anything by using deep learning models and thats where I found your neuralforecast library. Would you have any pointers in which directions we should explore? One challenge for us is, that not all articles are offered all the time (some only for like 3-4 weeks per year. However, there might be similar articles that were offered in the other times) and also not in all outlets. Therefore we have many gaps in the data and consequently different lenghts of the timeseries. Additionally, we do have much additional data about promotions, weather, prices etc. Thanks a lot for any input that you might be able to share!

I think you could try an Nhits model. Here is a getting started guide. https://colab.research.google.com/github/Nixtla/neuralforecast/blob/main/examples/mqnhits.ipynb

thanks. we will try it out and see how it goes 🙂