Forecasting time series with ML models

Hello again,

I was looking at the workflow “04_Machine_Learning” from the course “L4-TS Introduction to Time Series Analysis” and I was wondering why the target is residual and not the signal.
Another workflow from @Iris “Simple auto-regressive model to predict a time series” for example uses the signal.
Is it important to use the residual as target in ML models? Am I missing something here?

Thanks in advance and best Regards,

Hi @Franziska_W,

In the L4-TS course we walk through how to decompose a time series into a trend, seasonal pattern, and residual. Some forecasting models require data be stationary and this is why we go through this process.

I believe the workflow you’re referring to uses a Random Forest model. In this case stationarity is not required and you’re safe to train the model on the original time series if you would like. It was trained on the residual in the exercise for consistency with the rest of the course.

Hope this helps! If not feel free to ask more questions here.


Hi @Corey

thank you so much for your detailed answer. So far all of my questions have been answered, but maybe I will come back to your offer. Thank you!


This topic was automatically closed 182 days after the last reply. New replies are no longer allowed.