Fraud and payment risk management is incredibly complicated. It’s made up of a set of processes often requiring large teams to remain effective. Indeed, implementing a set of rules and machine learning models is a good start, but the work really begins after this stage, as those rules and models require constant monitoring to ensure they continue to perform as required.
Fraud trends change regularly, so this is a bigger task than most organizations anticipate when starting their journey into risk management. The main reason this is such a large undertaking is the enormous amount of data involved – it is simply impractical in most cases for even the largest teams to inspect it all. This is mainly manually undertaken by fraud specialists and only a small fraction of the data can be investigated; meaning some fraud will go by undetected, until a customer notices and requests a charge back.
Another problem with payments ‘big data’ is with building effective fraud risk models. Fraud only constitutes a tiny fraction of the overall number of payments, which makes it extremely difficult to detect effectively, even with the use of machine learning. Modelling software has come a long way and today can produce some truly outstanding results – provided the data is good and the problem is well posed. The process of determining the best data on which to train a model is largely manual, and again, requires a lot of effort for the top results.
These processes need to be performed for each separate customer on a fraud risk company’s roster and quickly becomes a problem, as the customer base grows, and the data outgrows what the current team can manage. The classic approach to this problem is to hire more staff to cope with the increased workload. With team sizes exceeding 50 people in many cases – providing initial short-term growth, it is unsustainable, as eventually staffing costs will consume all profit.
The answer: autopilot ML – process automation powered by machine learning
The machine powered components fall into two parts: the pure ML element for building fraud detection models and the automated process management component.
ML fraud modelling technology will continue to advance, by incorporating more advanced techniques and additional data not yet collected, as of today. The auto-pilot end-to-end process will become more and more sophisticated by removing the manual effort of the following processes:
- Ensuring the best performance is constantly achieved, as models tend to degrade in performance over time due to shifting fraud patterns. This process involves continual monitoring of the implemented fraud strategy, comprised of manual rules and machine learning based models, to ensure none of these algorithms are generating excessive numbers of fraud alerts. Badly performing models are evaluated against the latest data to discover the reason behind the decrease in performance such that a suitable replacement may be found.
- Curation – removing old rules and models that are no longer suitable. This can be difficult as older rules/models are often put in place to stop a very specific fraud pattern and there is a worry that removing it would open this up to fraudsters again.
- Fraud pattern discovery – A big part of a fraud analysts time is consumed with finding ‘the needle in the haystack’; identifying where new frauds are happening and the detail of how they are performed.
- Model/rule creation. Once a fraud pattern is defined, a model or set of rules needs to be created such that the fraud pattern can be defended against. Traditionally this was performed by fraud analysts, however this is today being offloaded to data scientists to create models – itself another process increasingly tackled by machine.
- Implementation of newly developed models/rules. Once the fraud pattern defence has been developed it is important to understand how it will affect the strategy. There is no use implementing a model which will flood the fraud analysts with alerts. By using a machine to automate the process of creating and testing a new set of candidate models or manual rules against a particular (machine discovered) fraud problem, the human component need only set the experiment up, receive results and make suggestions to the fraud manager for which to implement.
It is not too much of a stretch to image most of the fraud risk strategy process becoming automated. Instead of the expanding teams of today performing the same manual task continually, those same staff members could be used to spot enhancements in customer insight. This would enable analysts to thoroughly investigate complex fraud patterns the machine has not picked up on, or to assist in other tasks outside of risk management which provide added business value.
Process automation is continuing to innovate and provide increased efficiency and profit gains in the places it’s implemented. The automation revolution isn’t coming, it’s here, so prepare your business for streamlining, more effective, engaged staff and increased profit.
In summary the questions you should ask are simple:
- Are technology solutions/providers allowing you to scale with ease or creating more bottlenecks?
- Are your end to end fraud management roadmaps based around autopilot ML?