• Home
  • News
  • Promoted
  • PIN-UP Global on the implementation of ML and AI in combating fraud in sports betting
igamingnext photo
In the PIN-UP ecosystem they create their own ML/AI models to recognise fraud actions. Volodymyr Todurov, chief analytics officer, told NEXT.io about how artificial intelligence helps optimise anti-fraud processes and what benefits businesses receive.

NEXT.io: Volodymyr, can you tell us what risk management is in sports betting?

VT: All risk management in a sportsbook is based on individual limits when placing a bet. Basically, limits are set for a certain hierarchy – type of sport, tournament, event, markets.

Both limits on the maximum bet amount that a client can make, and on the acceptance delay time. This also includes margin management.

There are also individual limits for different customer segments – beginners, regular users, VIP players, high-risk players, and so on.

Limits may vary for different marketing activities. This is a dynamic system that allows you to adjust the risk appetite of the sports book flexibly and receive predictable profitability while satisfying client demand.

NEXT.io: How are fraudulent game styles determined without AI? Tell us about the classic approach, which was relevant both before the era of AI and now.

VT: The first one is the methodology for handling the anomalies. First, the norm is determined, and anything that differs significantly from the norm is already regarded as an anomaly.

A sharp shift in turnover in a specific tournament, a sharp skew in the live/pre-match distribution in a specific customer segment, a jump in the average bet, an increase in the share of bets placed under maximum limits. This is a signal for you that you need to conduct a more detailed analysis of the segment in which the anomaly occurred.

Secondly, this is an analysis of user preferences. What amounts, at what point in time and in what markets does the client bet. In this context, it is quite effective to find clients who play on errors in the bookmaker’s odds or use the product as an intermediate link in another fraud combination (I’m talking about related products – financial fraud, casino fraud, bonus abuse, etc.).

Third, this is special software that is used by both bookmakers and arbitrageurs. This is probably the most widespread pattern of fraudulent clients, since it is the simplest one. Up to process automation. The good thing is that bookmakers also see all arbitrage situations in their odds and react to clients who play exclusively in these markets at that moment in time.

NEXT.io: What are the advantages of using AI in this process?

VT: When we deal with huge volumes of data, the advantage of ML/AI models is obvious. This is a more advanced ability to generalize information and identify patterns of customer behavior that is not available with methods such as manual transaction analysis or the use of linear rules.

AI is an additional set of tools that helps the operations team work. It reduces the errors rate and decision-making time in each specific case, and opens up new knowledge about our clients.

In some ways, AI even deepens our expertise and broadens our horizons, since the models take a slightly different approach to decision making. Many tasks can already be given to AI autonomously.

Thus, AI helps to scale your business volumes without scaling the risk-management team, removes boring routine from daily operations and allows you to concentrate on a deeper analysis of other processes.

NEXT.io: What is your AI model and how does it work?

VT: AI-based decision models analyze a huge number of parameters of each transaction, such as the frequency of bets, their sizes, the selected events, as well as their history.  The model analyzes these parameters, choose those with the highest predictive power and uses them in the forecast.

For our task, machine learning models were used, the main one being the extreme gradient boosting model, which is one of the best for structured data.

The essence of the method is to build an ensemble of models that consistently refine each other, thereby increasing the accuracy of the forecast.

At the input of the model we give a detailed history of bets, at the output we get a probability in the range of 0..1 that the user is a fraud.

NEXT.io: How was the workflow for creating this product built? Tell us about the important stages.

VT: Probably the most important thing for models is the quality of the training data. So we started by hiring an employee with a serious background in sportsbook risk management to join the team. His main task for about three months was to validate the correctness of the segments of fraudulent players.

Every day he reviewed de-identified information from hundreds of accounts, preparing data for training the model. This helped us a lot at the start, since the very first models at the proof of concept stage immediately showed us an adequate level of type 1 and type 2 errors for operational work.

Next, an important step was to conduct high-quality onboarding of Data Science specialists who worked on the models. All team members must clearly understand the domain area and use terminology to maximize the results.

We trained models in several iterations. After each iteration we rub the resulting model on data on different time periods. Thus on ongoing basis clarifying the user marking and improving data quality for the next version of the model.

NEXT.io: Do you have plans for future models? What problems will they solve?

VT: The next stage is the creation of a decision-making system based on deep neural networks and generative artificial intelligence. We will reduce the decision-making time, bringing closer the goal of increasing the accuracy and speed of determining the client’s fraudulent game style.

This approach, thanks to its ability to learn and adapt, will be able to study in detail the history of each client, using various scenarios to identify a particular group, including fraud.

These new tools, brought to a sufficient level of accuracy, are very helpful in the routine manual work of the anti-fraud team.

For large businesses, this is an excellent auxiliary tool for retrospective cross-checks, which allows you to find out whether we have correctly labeled all clients as “good/bad”, so that managers can have a peaceful sleep – whether we have missed any dangerous players.

For small businesses, this is an excellent optimization point. Having such a tool in your arsenal, you will have no need to immediately recruit a large staff of analysts; the bulk of manual work can be given to AI, increasing the efficiency of the anti-fraud department.


Volodymyr Todurov has eight years of experience in gambling, having held the positions of bookmaker, risk manager, and head of antifraud & analytics. He currently holds the position of chief analytics officer at PIN-UP Global.

He is an expert in risk management operations of sportsbook, casino, affiliate programs, marketing, payment systems. Volodymyr specializes in data-driven decision-making approach, economic modeling, and forecasting.

Similar posts