Once I used the test set to assess the model performance is good why can’t I retrain from scratch the model on train and test sets toghether?
Take as example forecasting: I have a time series of length L, I train until T (T < L) and then test from T to L. I get a good performance so I know the model is good. Then at inference time I have to forecast the L+1 element, so why can’t I train the model until L and then do inference on L+1?

I would expect the model to perform better once is trained on a larger dataset (train+test) so what am I missing?