Whitepaper: The new role of the fraud analyst in machine learning

The scale of the fraud problem

The numbers are stark. According to a Nilson report, card fraud was over $24 billion globally in 2016. According to a recent study by Javelin, financial fraud is costing retailers around 7.5% of their annual revenue, with digital retailers feeling the greatest impact. While the size of this number will raise some eyebrows (it includes the cost of combatting fraud) at its size what is not in question is that the volume, variety and complexity of the fraud attacks are all growing. The pressures therefore on the shoulders of the fraud analyst grow ever weightier.

The fraud analyst is the central player in this world. They are the saviors of those of us who have had our identity stolen, only to find that a large loan has been taken out in our name. The fraud analyst’s basic role is to spot the bad guys - which transactions are real and which fraudulent. Fraud prevention is vital for online merchants because if fraud happens, it's the merchant who pays.

Fraud analysis is multi-disciplinary, using know-how from behavioral and forensics specialisms. A fraud analyst can wear multiple hats too, partly identifying fraudulent transactions from data, partly doing research to identify trends in fraud. They often interact with other areas of the business to offer insight and help to optimize processes. In today’s cyber crime ridden world of ecommerce, the fraud analyst is our super hero.

But to produce those heroics the analysts need support. Ecommerce reached $2 trillion in 2016. That’s a lot of transactions to create rules for, to manually review for,  and to create future strategies for. To manage this scale, it’s clear that machine learning has an important role to play. But what does the introduction of these number-crunching technologies mean for day-to-day working life of an analyst?

The changing word of fraud analysis

The last decade has seen some astonishing changes in the way we work. Not least is the coming of automation. Automation has entered industries as diverse as manufacturing, banking, and healthcare. A study by Accenture found that in the finance industry ‘Finance Automation’ reduced processing costs by 80%. In healthcare, automation has been used in places like pharmacies to perform repetitive tasks like medication counting. In medicine we are starting to see the role of the radiologist change with machines scanning x-rays much more efficiently and quickly with radiologists interpreting the results.

Fraud analysis is now entering that period of automation. Machine learning-based systems such as Ravelin are being used to learn trends and spot patterns in transactions which point to fraud.

The machine (computer) starts off with some initial variables, then as more data comes in the program adjusts the output - it improves over time, giving more accurate results. Humans do this too of course and are infinitely more adaptable. The difference of course with machines is the scale at which machines can perform that analysis and the breadth of data they can consume.

Is ML a friend or foe?

One of the worries about automation and techniques like machine learning is that they will ‘take our jobs’.  Whether it is a failing or a success of the technique we have rarely seen this happen in any other industry. The digital spreadsheet for instance certainly removed a great many manual data entry roles in accountancy. But it gave wing to many more value-adding accountancy roles in its wake with accountants able to charge more for their services.

Humans create good strategies

While good data gives rise to good strategy; data itself cannot create a fraud strategy. Machine learning might well be a very important part of that strategy but it does not replace it and it never should. The higher value work of determining the they type of fraud threat, the tactics and resources required to tackle it, the policies, the SLAs, the reporting to the business - these requirements are not going away. Machine learning products like Ravelin can underpin these strategies, but it is not a substitute for determining what is right for a business.

Keeping the machines honest

Machines are far from perfect. They can overweight signals and block good transactions. They can misinterpret information and wrongly allow fraudulent payments. They can take time to adapt to a new threat.

Any reputable vendor will be completely frank about this. Obviously these errors should be kept within agreed limits but the interpretation of and feedback about these missed and bad calls are the key to improving performance.

Working with your vendor, it’s important to establish productive feedback loops. For instance within Ravelin it is possible to confirm and reject decisions that are made by the machines. These are strong signals to the machine learning models that will enforce or weaken a decision matrix, ensuring the models continually provide improved precision and recall.

Further to that we recommend weekly meetings with a vendor to discuss trends and suggest areas where performance can be improved. Detailed analysis of customers or transactions that have been wrongly determined are invaluable feedback for a data science team.  It helps improve the models to improve future performance.

Screenshot 2018 11 09 At 13 25 42

Fraud investigation for profit and for customer service

A perhaps surprising outcome of the adoption of machine learning is the increased relevance of the fraud team within the customer’s experience. Whereas in the past the sheer volume of manual review would often restrict time available for investigations, there is now time available to any customer queries and investigate decisions.

For instance, a perfectly legitimate customer can be blocked for fraud. It’s possible to rapidly check  a fraud decision and if it appears a genuine error,  overturn it and try to save the customer. However, increasingly fraudsters are more than happy to call or email in to try to overturn a decision.

Having the fraudster’s complete history, the reason for the decision, and their link analysis network straight in a single dashboard means an analyst can confirm the decision with 100% confidence. Machine learning tools are often misconstrued as a black box environment that are impossible to query. Ravelin provides rich detailed information, sensibly displayed to allow fraud analysts to confirm decisions one way or the other in seconds.

The role of the fraud analyst therefore will move closer in scope to the customer service team and should be seen more as much a part of the customer experience as part of revenue protection. Some recruitment of people with the correct qualities may be necessary to manage this tweak in responsibilities.

Giving The Fraud Analyst Wings

Machine learning should be seen as a useful tool for the fraud analyst. Rather than replacing the analyst, it works to enhance the role, taking the leg work out of fraud analytics. The analyst team will be much more concerned with strategy, with the accuracy of the results and with improving the decisions. There will be a significant shift in the interpretation of the role, seeing it more clearly as a value-adding intelligence role that is critical to revenue generation as the business gets better at allowing more, good customers, while stopping more fraudsters at less cost and with less interruption to the business.