0
$\begingroup$

Let's say we have a forecasting model that was trained on any data before 2021 and now we need to make a prediction on data in 2023, for an accurate prediction we need to either give the data of 2022 as input and extrapolate the prediction off of this recent data and in case we don't have this data of 2022, we will ask the model to make the predictions on 2022 first then extend it to 2023, which I believe will be less accurate as the errors in year 2022's prediction will eventually influence the inference of 2023.

I want to know what are the common methods to deal with such a problem in production, should I maintain a database with most recent data or expect the user to give it to me or simply predict everything that is missing

Also, should there be a presence of all the missing data while making an inference or would the most recent most data do .

I want to know a general approach in forecasting algorithms, so that I could extend this approach to all the algorithms.

Thanks

$\endgroup$

1 Answer 1

0
$\begingroup$

Asnwer:

You should "maintain a database with most recent data".

  1. The first approach is to retrain your model regularly with recent data, so in this case, you would retrain your model with 2022 data before predicting 2023. So you should be retaining data for this future need.

  2. The second approach is not something you want to do, which is predict further into the future than your model can accurately do within an acceptable level of error. If you're outside of whatever your acceptable level of error, your predictions are not trustworthy and you should not use them to base any decisions on.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.