Multiple rolling windows prediction
#14
by
totallytong
- opened
The blog post says that the test set can be multiple prediction_length longer than the training set for testing on multiple rolling windows. How can this be done using the test sampler and instance splitter magic as in the blog post?
In particular, suppose I have the data of 100 time series, each with 500 daily values, with prediction_length being 10. I also have a 5-dimensional real dynamic time feature for each time series (so together they are of shape (100, 500, 5)). I would like to use the last 50 days as the testing window. How can I arrange the instance splitters to achieve this?