Forecast validation with apply forecast

Barclaeys

hi,

我用一个预测验证odel to validate a Holt-Winters model (Sep/2015 - May/2020) and afterwards I want to add an apply forecast operator to still make a prediction of future values (Jun/2020 - Oct/2020). However the predicted values occur for dates on which I have actual values (from Nov/2019 - Apr/2020)? Does anyone know how this is possible? Could it be because my last window that is used to validate the model has not used these 6 months?

Thank you for providing some insights.
Bart

MartinLiebig

Hi@Barclaeys,

This is likely an effect of the definition of windows, which is changed in an upcoming release.

I suspect you used a window size bigger than 1 (likely >7). We define our validation windows from the start. So it can be that the last start of a training window is not the last date in your data set. The model which is returned is the last model.

In order to get the "final model" you need to retrain your model on the last window_size data points.

Best,

Martin