My goal is to create a time series model with

  1. Multiple Entities - I have multiple products with pre orders and they all have the a similar bell shaped curve peeking at the release date of the product but different orders of magnitude in unit salles OR I can use their cumulative slaes what is an "S" shaped curve. But I only have about 100 products 1 year of daily data to do the training on.
  2. Multivariate - I have a wide variety of data on these indie movies for each day: A.) number of times people added them to the Wishlist, B.) page views, C.) time spent on the page AVG, -> Y.) Target value is the number of products people payed for (it is the same pre order before release and normal purchase after release date)
  3. Multi-step - predicting 60 days ahead would be the goal
  4. Every day refreshing the predictions for every product - Does this requires me to retrain the modell on the whole dataset?

Already Read

  • I have found algorithms that can do prediction on 1 variate maybe even Multivariate. Multi-step is already problematic and I don't know how to add the Multiple Entities part at all. So I cant fine a project or guide that would contain all these 3 parts that I nd
  • I have tried LSTM (13 different models with different datasets) but on longer "Multi-step" it is not working so more than 1 or 2 days. I also cant make the LSTM to accept Multiple Entities so I just chained each products data after each other historically, I do understand that it is not an optimal practice for sure.
  • Python package non popular so I cant find projects to it - https://stats.stackexchange.com/a/412355/256200
  • I always see this R guide but I don't use R. I need help with Python - https://otexts.com/fpp2/hierarchical.html
  • Not multiple variable and not Multi-step - https://stats.stackexchange.com/questions/356008/multiple-time-series-prediction-python
  • 1
    $\begingroup$ Have you tried XGBoost? It seems difficult at first view to predict 60 days with only 365 days of training, but it could be possible with some data pre processing. I had to deal with similar problems but with stock markets, and I had to find the right time range sampling (ex: grouping the mean values for one week) to make better predictions. $\endgroup$ Commented Jul 2, 2021 at 13:57
  • $\begingroup$ I have used XGBOOST before for Regression and Classification. I am not sure how to push the target value historically back to predict the future. I am not sure about his whole turning time series problem to regression problem. I would love to see a guide on it. $\endgroup$
    – sogu
    Commented Jul 2, 2021 at 14:21
  • $\begingroup$ I've found this code but I don't know if it could be helpful or not: programmersought.com/article/96993038054 $\endgroup$ Commented Jul 3, 2021 at 7:43


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.