A better approach to writing this article would be to gather a wide array of customer features, fit a model using training data from actual fraudulent/non-fraud

A better approach to writing this article would be to gather a wide array of customer features, fit a model using training data from actual fraudulent/non-fraudulent orders, and then interpret the model to explain the features actually connected to fraud. My guess would be that given there's no reason to believe a particular functional form or additivity of effects, a random forest would likely be the most effective classifier, but ultimately I'd just go with whatever empirically does best on the test set. As-is the article is basically a pretty naive approach to feature engineering a few features that may or may not ultimately be useful in the real data. It's a cute anecdote, but hire a data scientist.