Rethinking Your Fraud Approach: Retraining and Continuous Optimization
We’re almost to the end of our mini-series on “Rethinking Your Fraud Approach.” Last time we described how to activate your new enhanced fraud approach —actually deploying it, getting all the data sources connected, and running the rules based on the model’s outputs. We also recommended combining machine learning and a decision management service like Enova Decisions Cloud to augment—not replace—your fraud ops team. However, there’s one more topic that’s necessary to make sure your model—and your decisions—are always performing as expected: retraining and continuous optimization.
No matter how carefully you designed, trained, and tested your decision model, there will always be times when you need to fine-tune its performance. That’s because we don’t live in a static world, and the economic conditions and behavioral assumptions on which we based and trained our model change over time. As a result, we can’t assume that data distributions in the future will reflect the same patterns as when we originally trained and deployed our model. This phenomenon—known as model drift—may occur gradually or swiftly, but either way, it means our models must adapt if we are to maintain the expected results.
Contrary to popular belief, however, continuous optimization is not inherent in machine learning, and retraining must be designed into your process. How often you retrain your model with new data depends on your internal process. Some organizations retrain on a set time interval—say, every six months—where others base it on a number of events (such as having processed 10,000 applications). Still others monitor thresholds on the error rates. Regardless, the best practice is to set the interval or threshold parameters so you have sufficient new outcome data to retrain the model—something you can do automatically through Enova Decisions Cloud.
Besides retraining to mitigate model drift, you should also determine whether your rules need to be adjusted based on the new outcome data and current business goals. For example, if the business needs to further reduce its fraud risk, you could reduce the auto-approval threshold, thereby increasing the number of applications that must go through manual review. (Enova Decisions Cloud allows you to easily make such rules changes and even run scenario testing in the background.)
Finally, we’ve already talked about how leveraging third-party data from a variety of sources can reduce your fraud risk, but you should look into incorporating additional data sources at least annually. While doing this on your own can be both time-consuming and costly, Enova Decisions can save you the time and trouble. We’re constantly testing new sources to incorporate into the Enova Decisions Cloud, and we’ll let you know when there is something new (like behavioral biometrics data) you should consider adding.
We hope you’ve found our series on Rethinking your Fraud Approach enlightening, and even more so that it has put you on path to effectively fighting fraud now and the future. And Enova Decisions is here to help you get there faster. We’ve packaged up our 15+ years of fraud detection and prevention know-how into fraud solution accelerators. Contact us today to learn how get better fraud detection and prevention up-and-running quickly.