The increasingly massive volume of requests, the need to provide an almost immediate response, often via the Internet, and the rapid evolution of technologies all work against the detection of fraud in financial companies. Added to this is the worrying increase in fraud attempts, catalyzed by ever-present inflation and a complex economic situation that seems to be going on for a long time.
As a fact, in 2019, fraud claims accounted for only 6.2% of the total, a figure that in 2020 escalated to 10.6%, in 2021 to 14.4% and in 2022 to 30.3% of the total received, doubling the number of frauds identified in a single year.
And it’s not just the quantity of fraud attempts that is increasing, but also the quality. It’s a world in which concepts such as the FaaS(Fraud as a Service), which refers to the wide range of services offered in hacker forums to lower the cost of entry for those who wish to dedicate themselves professionally to financial fraud.
On top of this, fraud detection must be as passive as possible. Fraud prevention systems need to be integrated at the appropriate points in the sales funnel as transparently as possible, minimizing the friction of the process of acquiring a new customer. To this end, a response must be provided as early as possible.
But there are advanced technological solutions for defense, with proven results, which we will discuss in this article. We will focus on the detection of unwillingness to pay, not on unwillingness to pay, acceptance scoring or anti-laundering mechanisms.
Any defense begins with a study of possible attacks, separating cases whose treatment is necessarily different.
Attack #1: Impersonation
The most basic type of fraud is the one in which a person provides real data, but of another person. Here the solution is to use a robust, real-time digital identification system of the potential customer, ensuring that he is who he claims to be.
Biometric identification and multifactor authentication have become benchmark technologies in this area. The former requires data such as fingerprint or facial recognition, and the latter requires a double identification system from the customer, for example, a password and a code sent to a verified cell phone. However, they do not solve the challenge of new customers who need to be responded to before they move on to another provider.
In conclusion, these methods are necessary to ensure a secure entry point, but to optimize the identification/authentication experience in an adequate timeframe, they may not be sufficient.
Attack #2: Misrepresentation of data
The second type of fraud is that of false data or documents, which is particularly dangerous given the greater professionalization of those who commit fraud, who can intuit how acceptance scores operate. Data such as income, length of service, home ownership, and other less intuitive data such as marital status or property regime can be modified to make the models believe that there is a payment capacity that is not real.
The first solution is sectoral data sharing, using shared records where customer data can be verified at several points. If the customer had previous credits at another point, or at least applications, the information has been verified as there is no longer the urgency of response that exists in the acquisition of a new customer.
This is where artificial intelligence solutions begin to show their strength. Deep Learning models have proven to be exceptionally good, thanks to their ability to process and learn from unstructured data, such as text and images, opening new avenues in the identification of forged documents.
Recent technological advances thus add value to the defense system, mainly due to the speed and precision with which they work without relying on the human factor.
Attack #3: Fraud of intent
Finally, the third type of fraud is the most difficult to detect. In this case, the customer is who he says he is and the information he gives is true. And he passes the acceptance process by having the ability to pay. But for whatever reason he has no intention of doing so. There may be something on the customer’s financial horizon that has not been taken into account in the acceptance score, and that has led the customer to a desperate situation.
Worse still, the client may be well aware of the existence of loopholes through which to escape legal proceedings: something as simple as paying the first installments makes it possible that at the end of a long legal process there is no “bad faith”, or fraud, with the only result being a refinancing of the contract. Of course, this can lead to new defaults.
The only way to detect this last type of fraud, which has also proven to be very effective for the other types, is the use of AI models specifically trained to detect anomalies in the data being entered. The discourse formed by data from the customer’s identity on the one hand, and data from the specific specifications of the desired product on the other hand, can contain discrepancies that these models quickly spot.
There are two approaches here, applying Deep Learning or Machine Learning methods. Deep Learning includes unsupervised methods such as autoencoderswhich detect discrepancies and do not require labeled data to work. This comes in handy in the case that we do not have a manual labeling of all contracts, which identifies them as frauds or not.
However, given the importance of fraud prevention, it is relatively normal to have labels for all records. After all, all successful loans or loans with at least two or three payments are automatically discarded. The case of unsuccessful loans is more problematic, since it is unknown what would have happened in the case of applications rejected for not passing an acceptance score. The same applies to the much more common case where the client ends up not signing the contract.
Even with that, supervised Machine Learning methods work perfectly here, detecting irregular behaviors in real time and obtaining excellent results. Too expensive or too cheap options, optional details outside the customer profile, strange conditions… these models ask themselves the same questions that a human expert would ask, giving the ability to increase the speed of the acceptance process while keeping the assumed fraud low. They can also incorporate methods to provide explainability and interpretability, and can serve as an aid to the analyst.
The difficulty here is usually the existence of hugely unbalanced samples, where the proportion of the variable we seek to detect is enormously low. This can be solved by artificially increasing the proportion of that variable (oversampling), using models that handle this type of situation well, or with the more recent technique of adding synthetic data.
This is really where the full power of today’s technology adds value to the traditional fraud prevention process.
Integrated solution
The huge advantage of these models is that they also detect the other two types of fraud well, being enormously useful on occasions when the other technologies fail. After all, these models are lie detectors, which in a sense includes all three types of fraud. Combined with the other solutions, it is possible to concentrate more than 90% of the fraud in 8-9% of the requests, which can receive a more personalized treatment by an expert.
It can be seen as follows: a liar who is an expert in technology, but not very knowledgeable in business, will be detected by these models, and a liar who is an expert in business but not so good in technology will be detected by the aforementioned technological solutions. Exceptional liars in both fields will continue to pass both controls, which is an incentive to improve our models day by day, incorporating the most recent advances in an area in continuous expansion.
If you want to take financial security to the next level, contact us.
