How is data science reinventing fraud detection?

This article introduces data science and machine learning for fraud detection, analyzing past data and identifying new patterns. Companies must analyze threats, collect and analyze data to build effective fraud detection models.

According to the Association of Accounting and Management Professionals and Directors (APDC), internal fraud costs French companies 5% of their revenue.

Just like with external fraud, such as cyberattacks, companies are still poorly equipped to deal with the risk of internal fraud: 6 out of 10 companies have not allocated a specific budget to fight fraud (2020 Euler Hermes and DFCG study).

How can companies protect themselves from this threat?

Let’s start at the beginning,

Where does fraud come from? Or, why does an individual commit fraud?

In the 1960s, American criminologist Donald Cressey developed the theory of the "  Fraud Triangle ." According to his theory, an employee, manager, or director may be driven to commit fraud for three reasons:

• pressure : often financial, such as an overly high standard of living or addictions, but sometimes stemming from the company, for example through unattainable goals.
• opportunity : gaps in the company's internal controls.
• rationalization : a line of reasoning that justifies and makes the act of fraud acceptable ("everyone does it," "the amount is negligible compared to the company's revenue," etc.)  

Faced with this, the first reaction of companies is prevention

To deal with the risk of fraud, the first line of defense is, of course, prevention.

The primary preventive measure is to map out threats and draw up a clear action plan to be communicated throughout the company. The key to a successful prevention policy is to secure widespread participation and weave these practices into the company culture.

However, that is not enough

Prevention must be accompanied by internal controls and audits as well as alert systems. Yet, manual checks are time-consuming and often ineffective : an approving manager will tend to skim over expense report approvals or small transactions, for example, as these tasks are far from their daily priorities. As for audits, analyzing thousands of lines of data by audit teams is very time-consuming and, above all, costly.

This is where data science comes in.

Data science, and particularly machine learning, are already widely used today for issues such as spam detection, medical diagnosis, and media content recommendation. Machine learning consists of analyzing data to understand, predict, and classify information. This technology is therefore ideal for fraud detection.

In fact, it allows for the analysis of past fraud, identifying its origin and the environment in which it was committed. This way, teams can take action and implement new automated controls or points of vigilance to remind employees of as part of the prevention plan.

In practical terms, how does it work?

There are two main families of machine learning algorithms: supervised models and unsupervised models. Before going further, it is worth noting that applying either of these model families does not provide a definitive answer as to whether or not fraud is present, but rather provides a probability of encountering fraud (e.g., "This transaction has an 86% chance of being fraudulent").

The Supervised models learn based on so-called "labeled" data, meaning data that contains the expected result. For example, a dataset of accounting transactions with a "fraudulent" or "non-fraudulent" tag for each line. The model then learns to recognize fraudulent situations similar to those it has been shown. This method is effective for detecting fraud identical to past situations.

The limitation of this solution is the fraudster creativity. Indeed, when it comes to fraud, adapting and updating techniques is paramount.

We then use unsupervised models. In this case, the model is trained with a dataset without labels. It must then search on its own for similarities or anomalous behaviors within the data. Thus, the unsupervised model will make it possible to discover new types of fraud by decrypting large volumes of information that are impossible to analyze manually.

By combining the two, it is then possible to continuously monitor and escalate risks of known fraud while detecting new sources of threats within the organization.

Other techniques can of course complement these methods: automatic transaction checks based on business rules, a risk scoring system for teams or departments to prioritize prevention and control actions, data sampling to limit manual checks to the highest-risk operations, and more.

Where should my company start?

1. Analyze threats and map risks.
2. This will allow you to identify blind spots and develop a prevention plan.
3. Your data: COLLECT, STRUCTURE, and ANALYZE IT

From there, you will be able to build effective models to:

• Detect fraud quickly
• Identify threat sources
• Anticipate and take appropriate action