Standardization and prediction on new data As far as I know it is common practice to do standardization of variables before shrinkage or PCA, which are methods I intend to use on my model selection for a predictive model. But the problem is, how do I use the model with standardize coefs on future data. Do I simply assume that future data has the same distribution as my current or are there work arounds?
 A: First you should always make sure your training data and future data have very similar distributions. Otherwise, the model you developed will not working in production.
There are many different ways to normalize a feature, one simple way is subtract mean and divided by standard deviation (i.e., scale function in R). You can always do such scaling/normalizing for the feature you want to normalize in training data (using mean and variance in training data), and future data separately (use empirical mean an variance in production data you have seen so far). Again, based on our assumption, these two data have very similar distributions, at least their mean and standard deviation will not be way off. If they are close enough, the normalization should be fair.
Finally, it is very often to see "outliers" in production environment. For example, your training data is always positive but it is possible to see very few negative instances in future data. If that happens, you should be brave enough to say, "I am seeing strange things and this is different from the training data. So I need to stop to make predictions." Instead of making the predictions anyway.
