C1_W2 - Feature Scaling for Unbounded Data Points


While I try to understand the use cases of feature scaling, Andrew Ng says it is almost a good idea to use scaling to normalize our data. However, I wonder what if our data has no bound such as stock values. In that case, the lower bound of data will be zero, but upper bound theoretically can go to infinity. If we use scaling in training, we can probably go beyond that scaled dataset in prediction. But if we don’t use it, this time we may face a slow descent. What should we do in these cases?

Thank you

The data set is normalized based on what you have in the training set. Not based on any theoretical future values.

Does this affect our prediction accuracies since weights are set based on the normalized dataset?