In fraud detection and prevention business the biggest thing is to catch thieves. That’s far from rocket science but agencies still hire world class expects to employ state of the art (rocket-science-alike) techniques to catch thieves. You may think that state-of-the-art techniques are what’s most important. We don’t. Long before we even started asking ourselves what is the most important in fraud analytics business we had the opportunity to take part in (or simply watch) many fraud analytics projects. Not all of them were phenomenal success but the ones which were shared something in common.
First. Successful organizations understood well that the battle can only be won on the ground – when dealing with customers and making small ticket decisions. Not-so-successful ones thought that most important decisions are made on the executive floor.
We are not saying that major decision such as a merger or acquisition is not important, but we can certainly argue that small decisions (many of them) are equally important for an organization.
In our practice we promote micro risk management. We try to improve decisions made daily (in large quantities) by making use of data and advanced analytics.
Daily decisions can be recorded to create large data sets over time and improve model performance. We don’t have that luxury in major events such as mergers/acquisitions.
Second. Successful projects are art as much they are science. Take fraud detection models for example. The obvious target variable, the one which can be justified to a rational person, is fraud tag. A case which was suspected to be a fraud and which was confirmed to be fraud in court. What’s wrong with that? Noting? Except the fact that that is not what we want to model. Insurers don’t need an exact fraud model. They want something which will help them identify claims which are more likely to be fraudulent. Here is where art/craft kicks in.
If you were a rigorous scientist you wouldn’t even consider using claims processing agent’s suspicions as target variable. That would be too subjective – you could say. You may, but this is not a good practice. Fraud cases may take years before they can legally be confirmed as fraud. Some fraud cases may never be flagged as such because of legal loopholes, small enough to be ignored, due to lack of resources or for some other reason. The fact that you are dealing with unconfirmed fraud doesn’t mean that you should ignore this risk.
You could use staff hunches instead. This is where your fancy predictive analytics meet craft. Remember, models are for humans not for robots.
By doing this you will have bigger data set to work with (a lot of hunches). Not all hunches will prove to be correct. Be warned, it’s a delicate balance. How you decide the threshold can make or break the model.
More on fraud analytics next week. Stay tuned.