In recent blog posts I’ve discussed the nuances of first party fraud, third party fraud, and the importance of continuously monitoring customer relationships well after initiation to ensure that legitimate accounts have not been compromised by bad actors. Machine learning can be a powerful weapon in minimizing fraud of all types- if it is deployed properly.
Too often, machine learning is regarded as a magic potion rather than a central pillar of a comprehensive fraud prevention program. Regulators have been known to push back against machine learning algorithms in use at some financial institutions, to the frustration of compliance officers. Both sides have valid points- let’s look at some best practices to help ensure these tools are leveraged effectively.
Beware the Black Box
The biggest complication hindering broader use of machine learning is that its underlying models are often not transparent. The last explanation an examiner or auditor wants to hear for why a given decision was made is “because that’s what the model told me to do.” Auditors must be able to verify that a model is deterministic- that its logic can be traced and, well, audited.
This demand is only reasonable- whether a loan approval, an account opening or the decision to green light a card transaction, banks and credit unions need to be equipped to explain the basis for their actions, for a variety of reasons. The use of certain criteria is illegal, of course, and regulators must be able to confirm they have not entered the equation. In the “self-learning” aspects of the most sophisticated models, it’s even possible for such prohibited factors to unwittingly seep into the logic, without scrutiny and understanding.
Yes, it’s called machine learning- but the idea has never been for the machines to hold that newfound knowledge in strict confidence. Humans should be the primary beneficiaries of such learning.
A Complement, not a Replacement
These models are increasingly probing larger data sets, taking a more holistic view in pursuit of unearthing unexpected correlations that may prove effective in predicting behavior. This involves sifting through more information on geo-location, account types, channel context, and the features, attributes, and sequencing of transactions in order to draw conclusions.
Obviously this cannot be done manually- which is why it hasn’t been part of the process until now. It requires intelligent mathematical models, powerful computing capacity- and a workflow that gets the relevant output of that analysis into an industry professional’s hands for a considered, real-time decision.
This is the other major misperception of machine learning- that it somehow wrests control of operating decisions from the hands of skilled staff. When implemented properly, the opposite is true. Machine learning should execute the data crunching- at degrees of difficulty beyond the capabilities of the most talented spreadsheet jockey- and serve up more robust information with which risk analysts, loan officers, etc. can make better decisions.
A blend between machine learning recommendations, rules and heuristics based controls that frame the machine learning, and skilled analysts are the solution.
Auditors and examiners demand a comfort level that the benefits of machine learning are being leveraged properly. That’s precisely the sort of comfort bank and credit union leadership should demand as well.