0
$\begingroup$

I have a time series dataset containing daily data like below. Let's assume that I would like to make some forecasts of my temporal serie (x) and use it as a second feature feature (f) to predict the label (y). How to properly use forecasts as input features without risking overfitting ?

Let's assume that I have two datasets (train and test sets) such as the training set is :

date x y
2000-01-01 1 0
2000-01-02 2 1
2000-01-03 4 1
... ... 0
2014-12-31 512 0

And the test set is :

date x y
2015-01-01 1024 0
2015-01-02 2048 0
2015-01-03 4096 1
2015-01-04 8192 1
... ... ...
2021-08-21 65536 0

For the test set, I can use forecasts of (x) as an input feature (f) such as :

date x f y
2015-01-01 1024 2046 0
2015-01-02 2048 4099.5 0
2015-01-03 4096 8190.6 1
2015-01-04 8192 16381 1
... ... ... ...
2021-08-21 65536 131072 0

However, for the training set, is is better to use the real value (f1) or the forecast (f2) in order to have a proper validation and avoid risking overfitting ? What would you choose and why ?

date x f1 f2 y
2000-01-01 1 2 1.5 0
2000-01-02 2 4 2.1 1
2000-01-03 4 8 8.3 1
... ... ... ... ...
2014-12-31 512 1024 2049.2 0
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.