Pavel Zapolskii is a seasoned mathematician with over five years of experience in top-tier product companies at the intersection of technology and finance.
He holds a Master’s degree in Mathematics, specializing in Probability Theory, and has extensive experience integrating sophisticated technology with practical solutions.
Pavel’s extensive machine learning experience covers a wide range of applications, from predictive analytics to fraud detection. He has implemented and optimized complex algorithms to produce dependable, data-driven solutions that improve operational efficiency and security. As a Senior Data Scientist at Exness, Pavel continues to use his vast knowledge to advance the field of B2C fintech.
What are the requirements for anti-fraud protection in the fintech industry? What distinguishes fintech from other industries in terms of anti-fraud protection?
I would divide the response to this question into two business streams: product and technical.
In technical terms, the anti-fraud solution is oddly similar to that of any large B2C business: the accuracy of our actions is very important to us, and the cost of error is high in both directions – missing the fraudster and punishing the innocent. It is almost impossible to fully automate anti-fraud; you must either deal with errors before making a decision or support users afterward.
The key distinctive feature of antifraud in fintech is the biggest impact on the company’s overall EBITDA. There is no more dangerous type of fraud than payment and FX conversion fraud. Unlike in other areas, investing in a team of both operational anti-fraud and machine learning specialists is critical to a fintech company’s survival. We can say that ML improves the recall of the entire pipeline and operational activity increases precision.
How can anti-fraud systems help maintain the sustained growth of B2C fintech companies?
Starting with a certain cash flow level, any fintech company that does not implement anti-fraud measures will experience an increasing loss of net profit. Furthermore, it is not only about the company’s finances, but also about customer churn caused by a loss of credibility and reputation in the marketplace.
In the illegal market, there are entire companies that “rob” newly formed companies in the field of finance, preventing them from exceeding a specific profit margin. For example, if your fintech product’s monthly FX conversion turnover exceeds $10 million or the total loan issue exceeds $100,000, even a 1% leakage of the company’s effective profit may be appealing to scammers, who will pay attention to you.
What typical mistakes do fintech companies make when implementing anti-fraud systems?
One of the most common mistakes made when developing an anti-fraud system is attempting to place the entire burden of responsibilities solely on the operational or software teams.
In the first case, it is impossible to cover the entire fraud flow without automation, and it will be extremely costly from the standpoint of the wage fund. On the other hand, attempting to detect all cases of fraud using analytical and ML systems results in a significant decrease in algorithm accuracy and serious monetary and reputational losses for the company.
It is important to understand that the software-ML team is primarily a development team whose direct job responsibilities do not include extensive investigation or a thorough understanding of the specifics of the most dangerous fraud for your company. As a result, even the most complex and effective model can be inferior to a simple heuristic logic that has been reviewed and refined.
What is more important is that many fraud cases are extremely difficult to detect. For example, payment systems in any fintech require enhanced monitoring: your company must be sufficiently data-driven to not only restore the history of the incident, but also simply notice that something occurred.
Can you provide recommendations on how to avoid common mistakes?
I believe the secret sauce is in the layering of all traffic checks: an easy math problem shows that with a detection accuracy of only 60%, which is slightly more random than a coin flip, a system of three ‘independent’ filters will already achieve nearly 95% accuracy! These filters can be:
- An analytical cross-section of products, such as dangerous regions or areas of traffic purchase.
- A machine learning algorithm (e.g., boosting) predicts a user’s likelihood to cheat based on their profile.
- An operational and support team investigates controversial cases.
Here’s a simplified example of an anti-fraud system for a payment department:
- We discovered that poor quality traffic frequently originates from India and is associated with Indian rupiah deposits/withdrawals. Using analytical methods, we discovered that a surge in fraudulent activity occurs on weekends and frequently involves a large number of transactions in a row (> 5 in 10 minutes). That is our product analytical filter, and the better this step works, the fewer false negatives we have.
- We define the cost metric for our payments and train machine learning models to predict spikes and consistent negative values based on profile/activity behaviour. Then we work with the score to define “easy cases” and “complex cases.”
- We transfer complex cases to the operational team, who work on each case individually to understand the real situation, and then experts return to stage 2 with errors/tricky cases for iterative model improvement. The more expertise we have here, the fewer false positives we have.
Can you give an example of a modern B2C fintech company with excellent anti-fraud systems in your view? What makes it effective and successful?
One modern B2C fintech company recognized for its excellent anti-fraud systems is Revolut. It has made significant strides in incorporating advanced technology and strategies to combat fraud effectively. They have built an impressive machine learning system based on real-time streaming data, using robust models and algorithms to reduce false positives. Another major point of success is their partnerships with top fraud prevention networks. Sharing expertise helps them stay updated on emerging threats and leverage shared intelligence for enhanced security.
What trends and innovations will affect the development of anti-fraud systems for B2C fintech in the next 5 years, and how should companies prepare to integrate these advancements?
As history has shown, scammers devise new ways to steal money as anti-fraud systems evolve. And, as AI LLM models’ capabilities improve, operational people are playing a smaller role in the anti-fraud effects pipeline. Computer vision and chatbots are two examples of products that can help fully automate manual work.
However, keep in mind that scammers will have access to all of these technologies, resulting in new, more complex schemes.
