Deep Learning with PyTorch

Time sensitivity

Let's take the case of predicting stock prices. We have data from January to December. In this case, if we do a shuffle or stratified sampling then we end up with an information leak, as the prices could be sensitive to time. So, create the validation dataset in such a way that there is no information leak. In this case, choosing the December data as the validation dataset could make more sense. In the case of stock prices it is more complex than this, so domain-specific knowledge also comes into play when choosing the validation split.