The testing data will sequentially become training data after each prediction has been made. Adequacy of historical data is detected each day of testing and the training set is then updated accordingly.
To detect the convergence time, we use a method described by Phithakkitnukoon and Dantu with threshold of 0.02.