We recently had an interesting client experience, developing a ML based transaction fraud model for a major bank. This customer was using a Fraud Management Solution to monitor frauds for its internet banking and mobile banking portfolio. However, it had noticed that it was facing more frauds which were increasing steadily.
The client ask—and the answer
The client asked for help to develop an additional solution that would fill some of those gaps. We developed a machine-learning based fraud detection model, and integrated it into the bank’s systems. The new integrated systems now support predictive models and strengthen fraud detection capability. High-risk transactions are triggered for review or declined. This means that investigators are able to target suspicious cases more efficiently, and reduce the time they need to spend on each case. This has greatly improved the productivity of the fraud monitoring team.
All transactions are now scored in real-time, enabling faster and more informed risk-based decisions. This gives much better accuracy in fraud detection. Overall, we were able to help the bank to capture more frauds, more efficiently with lower false positives. Happy customers are always good—but what did we learn from this process?
Learning from experience
Here are my key learning points from the process of developing this fraud management framework.
- There is an ideal length of time for a development process, and there is also what is possible
One of the biggest lessons to emerge from this process for me was the sheer art of the possible. Normally, developing a machine learning model like this would take three to four months. However, we were able to complete this in three to four weeks, because of some timeline challenges at customer’s end. The model that was delivered was not exactly the same as what would have been delivered after several months, and that required careful communication. However, the customer obtained something useful to them within a short period—and it can now be improved over time.
- Even a very good generic fraud management solution can be improved with some bespoke additions
SAS Fraud Management is very much our bread and butter in fraud management. We’re proud of it as a product, and it works very well—as a generic solution. It can scan every transaction in milliseconds, and uses rules to flag up potentially fraudulent transactions, and it’s good. In this customer, it has flagged up over 1,000 potential frauds. However, generic systems can only take you so far. This machine learning fraud model development was designed as a bespoke add-on to SAS Fraud Management. It was therefore specifically designed to address particular problems experienced by that customer. The results speak for themselves: the additional model gave an uplift of around 100 times. There is a huge value-add to that, but the new model was very much an add-on to the existing solution.
- A hybrid approach gives you far more options
I think we probably all understand that in principle, a hybrid approach—one that combines several different fraud detection techniques—has the greatest chance of success. However, it is still good to see this demonstrated in practice with a first model iteration. The model that we developed combined business rules, anomaly detection, lookups and more, and the results were spectacular. It was almost unbelievable to see the size of the uplift in detection, and the additional value that this generated for the bank.
- There is a process that you need to go through in analytics development
There are only so many corners that you can cut in developing an analytical model even when speed is a priority. You still need to go through the right process. For example, when developing the model pipeline, you need univariate analysis before multivariate analysis. This allows you to identify the most important variables, and understand the relationships between variables. You then have to develop candidate models, and compare them to identify the champion and challenger models. You could miss out one or more steps—but in the end, it could have serious consequences. The key is understanding which steps you can abbreviate to shorten the process.
Interested in the topic? Read the white paper – Balancing Fraud Detection and the Customer Experience
