Today's email. Questions and proposed response in ...
# support
t
Today's email. Questions and proposed response in thread. MIGUEL ANGEL CANTERA CUARTANGO <macantera@ubu.es> 1:49 AM (8 hours ago) Hello! I have been using TimeGTP recently and I really like it, good work! I have a couple of questions to ask you: 1. Is it possible to feed the API with no datetime format? It is the case of my application where I use the data coming forma sensor every 2 seconds. 2. How the algorithm knows if the horizon is too high to make a good forecast? Does it check only the relation between the frequency set in the forecasting and the number of values (before doing the prediction), or does it check the performance of the forecasting (after the prediction)? Best regards, Miguel Cantera - Profesor Asociado Área de Tecnología Electrónica
1
Is it possible to feed the API with no datetime format? It is the case of my application where I use the data coming forma sensor every 2 seconds. You do need to use some frequency parameter that indicates time unit between consecutive data points. However you should be able to set it to use the frequency from your sensor. Here's some more information on the
freq
parameter. https://docs.nixtla.io/docs/timegpt_quickstart#using-datetime-index-to-infer-frequency
How the algorithm knows if the horizon is too high to make a good forecast? Does it check only the relation between the frequency set in the forecasting and the number of values (before doing the prediction), or does it check the performance of the forecasting (after the prediction)? I don't think the algorithm checks after forecasting? But I could use some help with this one. Thanks!
c
For the first one, you can add that they can use freq as 2S. For the second one, we do not check the relation between freq and h. Suggest them to use the long horizon model. I would add the caveat that we are currently improving the high frequency model.
👍 1