This question came up while I'm working on a related issue in mlflow.
Right now I use the time aware modeling/python api in the format of providing a dataframe
date,sales,temperature,pressure,output
01-09-2020,100,101,5000,10
01-10-2020,120,91,4000,24
01-11-2020,50,110,6000,30
and then I get a return dataframe like
date,prediction
01-12-2022,81
01-13-2022,60
01-14-2022,111
...
let's say it's a window of 60 days (so I provide 60 records), and a prediction interval of 14 days (so I get returned 14 records)
is the time aware platform meant to be used like this (making a prediction for a single subject/single prediction at a time)? It feels so jarring, considering that the not-time-aware seems conducive to making millions of predictions at a time.
follow up question is what kind of big-data/large amount of prediction use cases there are for time series that can be done on this platform?
Hi, Jonathan!
While I'm consulting over your mlflow question, let me try to help you with understanding time aware models.
You are correct about your estimations in example given, but there are few things are missing:
1) Series ID in multi-series setup. Your example described only one series, while you can do predictions to many in parallel. Also there is segmented modeling to scale up this parameter to really high values.
2) You can provide more than 60 values of historical data for prediction, let's say 60+n, then you can have forecast range predictions which will be able to give 14 predictions for up to n forecast points, depending on size of your range.
Additionally: