Behind the scenes of anti-fraud processes at PIN-UP.TECH: traditional tools and innovative technologies

Volodymyr Todurov, chief analytics officer at PIN-UP.
Volodymyr Todurov, chief analytics officer at PIN-UP.

Volodymyr Todurov, chief analytics officer at PIN-UP, details its data-driven approach to fraud prevention, combining traditional tools with AI and machine learning to combat everything from fake wins to bonus abuse.

Opinion.- igaming is a challenging industry with a variety of factors within billions of transactions and processes inside. That’s why DDDM (data-driven decision-making) approach is the only and irreplaceable way to move through this storm, especially for worldwide global companies. 

In terms of fraud-fighting there are two pillars to be efficient: business expertise and data analysis. I will talk about the second one and how we use data analysis to develop anti-fraud tools.

As always, fraud-fighting starts with “norm” and “anomaly” definitions for each process or business metric. It is a merge of business expertise and data science. 

As I mentioned, igaming industry is characterized by continuous changes. You expand by new geos, new payment flows, you always try new acquisition and retention strategies. Thus, your “norm” and “anomaly” bounds are changing. 

In this stage, technologies are coming to help optimize your monitoring process. In PIN-UP we combine both traditional tools e.g., rule engines/scoring models with AI/ML-models.

I believe that ML/AI is an extremely powerful technology in terms of fraud-fighting. However, it’s not more than just another tool. There are always tons of human work behind each single model. 

The key to success is the quality of markup data for training models and ongoing efficiency monitoring. Models are degrading in production and the challenge here is to adopt them with minimum delay not to make a wrong business decision or not to flee your operational risk-management team with false positive alerts.

Our current strategy is to predict each fraud scenario with both regular tools and ML-models. I will give a couple of examples below.

First, the games section. The most common workflow in case of an extremely big win is to check the legitimacy of the win with your game provider. The fraud manager makes a request, it takes time to check on the provider’s side and we receive a reply

This delay affects lucky user and brings not as smooth user experience as we want it to be. So, we started to develop our own tool with the internal name “games healthchecker service” to analyze automatically if the one single win looking abnormal fits the normal winning distribution matrix of the game. 

This tool dramatically reduces time-to-check the legitimacy of the win and brings smoother experience to the user not to wait extra time to keep playing or withdrawing the win. 

On the other hand, the games healthchecker service allows us to react to hacked/broken games mostly immediately. But, this is the only case of analyzing a single win. What should we do to cover the continuous fraudulent gamestyle e.g. bonus engine abusing? This is the challenge for ML-models. 

Continuous analysis of user’s preference on placing bets with real/bonus purses fits perfectly to the segmentation ML-model showing the segment of users with disbalance in turnover within the 24h time-frame on the gameID level of aggregation. With the first iteration in production, we disabled few dozens of games from bonus balance use for the amount over $500,000 of potential losses monthly.

Second, sports betting. We were happy with the regular tools and internal workflows in general. But there was a challenge for us: how to reduce the time to detect the fraudulent gamestyle. We kept all the good-working monitoring flows and focused on real-time predictive models. Using the gradient boosting model with the 5-bets window calculation we got very interesting results. 

So, with a balance of 0.8 per cent alerting users we marked over 80 per cent of fraudulent accounts. As a result, fraud managers could react faster on such accounts. Mostly, it takes 5-6 bets per fraud account to run the alert. 

But again, the model is just another way of detecting that helped us to make double-checks and better alerting for changing gamestyles. We are still using the classic risk-management approach of bets monitoring on the provider’s side combined with cutting-edge predictive technologies on our side. 

ML/AI will definitely replace the big scope of current tools in the industry, will lead to more business insights and increase the team’s efficiency. At the same time, anti-fraud teams will adjust their workflows to train models with fraud patterns instead of working the alerts generated by linear scoring engines and rule engines. 

Wisely selecting the tool for each task is the main challenge for high-level managers in the context of this article. Often, you just need to reconsider trigger weights in your linear scoring engine and that’s enough. No need to use a sledge-hammer to crack a nut.

At the same time, we conduct experiments with integrating new technologies to increase the anti-fraud efficiency wherever it is possible.

Volodymyr Todurov has eight years of experience in igaming, having held the positions of risk manager, and head of anti-fraud & 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 modelling, and forecasting.

In this article: