Is it always the case that more variables are necessary better? After some sweet spot, with more variables false correlations increase exponentially.
However, you can't know which features carry information until you collect and analyze them. For a problem like fraud -- where "expert" input probably would not allow you to figure out which features you need ahead of time -- it was almost certainly more reasonable to gather all the data and then, after the fact, perform feature selection[1].
- Collect as much as possible
- Figure out what features are worthwhile
- Focus on those features
Our competitors have an extremely narrow lens into all the data around a transaction. We've found things that they'll never find or even have access to in the first place. Blog posts to come here as well.