Walkforward cross validation is recommended for a time series model to walk through the dataset by expanding the window size (timestep) for training and testing and NOT the conventional Kfold and leave one out validation due to the reasons to preserve the time series sequential property and avoid risk of data leakages.
However, when I validated the data against all of the above cross validation methods, I observed that the bias and variance for the conventional validation is found to be better (lower) than the walkfoward cross validation, which I believe should be other way round.
I will appreciate if anyone can shed some light on bias and variances of a time series model please.
Thank You
In other words, you observe a higher bias applying the walk forward optimisation instead of the cross-validation, don't you? Well, it is difficult to say much without knowing more about the problem of interest and the parameters of choice but I would not be surprised since you probably use way more training data in your cross-validation - which potentially makes the problem considerably easier.
Khandu Om :
See if these articles can help you to find your answer:
Dear Khandu,
the results can be significantly different because in the right validaton method for the time series you should not use future data to calculate the forecast. Typical cross validation use future data, so it is not good.
Regards,
Grzegorz
- Badges
- Science topic