I have a doubt regarding usage of external datasets like gdp rate, unemployment rate... etc., in test set for time series prediction. These datasets are historical and can be used along with train set, but how can it be used along with test set ? Do we need to predict it first and then use it along test set or are there any better ways to do it ?
The general idea is that you should use the same data on your train, test and prediction sets. For economical data this can be tricky :
In your train set, you need to get what data was available at a given time. This can be tricky because most of the time such economical data can be revised a long time after being first published. For example in my country, I think GDP can be revised up to 3 years after initial publication. Sor for exemple, if you have an instance in 2015 in your train / test set, you need to get what data was available in 2015, not revised numbers that could happen up to 2018.
In your test set, and for general prediction, the actual value can be an estimation of what is currently happening (nowcasting), so you have to make sure that you data source are consistent. Nowcasting techniques are sometimes used to get the current values (for exemple to get an idea of the activity of the hotel sector you can look at booking.com # of reservation ... etc.). Those techniques are entirely different from what is generaly used to get final numbers (agglomeration of fiscal declarations). Ideally you have to make sure the methods are consistent between, train, test and prediction sets. Using other methods to get data for prediction should not be intirely excluded, but should come with very strong warnings.