Revolutionary approaches to data analysis are enabling companies to understand individual customer behaviour. For the Gaming industry, this presented an opportunity to conduct the world’s largest research project into problem gambling. Over 10 billion gaming machine events were analysed to discover whether it is possible to minimise risk of harm through understanding customer data. David Excell, Featurespace CTO, explains how the research was carried out and how big data analysis could help to make better policy.
A challenge for compliance and healthy gaming
Complying with licensing conditions regarding responsible gambling is a challenge for the Gaming industry. Betting machines in bookmakers, in particular, continually attract political and media attention over concerns about their impact on gambling problems.
To combat this problem, the Responsible Gambling Trust wanted to investigate whether it is possible to identify harmful patterns of play from gaming machines in bookmakers. Additionally, they wanted to explore whether it is possible to create a list of markers of harm that indicate players at risk, and draw implications for responsible gambling interventions to minimise harm.
To tackle this complex issue, which involved co-ordinating and analysing vast volumes of data, the Responsible Gambling Trust chose to partner with Featurespace and NatCen (National Centre for Public Research) for their respective predictive analytics and research expertise.
A series of ground-breaking firsts
This ambitious research project was the first time that five of the largest UK bookmakers (Betfred, Coral, Ladbrokes, Paddy Power, and William Hill) made their industry data available for analysis. It was impressive seeing the Gaming sector come together to try and find ways to minimise the risk of harm to their players.
It was also the first time in the UK that an independent, large scale research project had tackled the issue of gambling-related harm.
Tackling responsible gambling with customer analytics
The Gaming sector has always been ahead of the curve in using data and analytics to understand their customers, embracing machine learning approaches to combat fraud by understanding each player’s behaviour in real-time. The same principles applied to tackle the identification of harmful play in gaming – looking for changes in behavioural data that indicate when an individual player is at risk of harm.
A machine learning approach was used to measure the predictability of each individual player’s interactions. Over 10 billion gaming machine events were analysed to identify key markers of harm that indicate which players are at risk of unhealthy play. The data set included over 6.5 billion bets placed and 333,000 uniquely identifiable customers, covering a 10-month period.
Using an Adaptive Behavioural Analytics approach, the system used proprietary algorithms to analyse these 10 billion gaming machine events and build statistical profiles of ‘normal’ patterns of behaviour for each anonymous customer, spotting the anomalies that indicate the exact moment when behaviour is changing. This enables the automated system to understand, in real-time, whether or not the change is a significant indicator of a player being at risk of harmful play.
Linking to data from customer loyalty cards
Surveys of loyalty card customers were undertaken and linked to the transactional data. Loyalty card holders tend to be highly-engaged gamblers, and traditionally Gaming operators have to rely on customers self-excluding if they think they are at risk of addictive play. During the research project, a physiological screening process (known as the Problem Gambling Severity Index) was used to identify problem gamblers from a random sample of loyalty card holders who had gambled on machines in a bookmakers. The PGSI score was used as a proxy for harmful play. NatCen collected survey responses for 4,001 of this sample. The responses were then linked to loyalty card data to understand whether patterns of harmful play could be identified from individual player’s data.
Results: identifying players at risk of harm
So is it possible to minimise player harm by analysing customer data?
The short answer is yes – the research identified players at risk of harmful play, where the operator could potentially send personalised individual interventions to reduce the effects of gambling-related harm. 19 potential markers of harm were identified following review of the evidence, such as time markers (e.g. frequency of gambling), and contextual markers (e.g. how the person behaves while gambling on machines). The model built during the research showed a 66% improvement in detecting players at risk of harm, compared to the industry standard.
The analysis showed that modern machine learning systems are capable of spotting the risk of harm for each individual player. However, the research also identified a need for further work to define the thresholds of where harm lies for each individual, and demonstrated that to be able to adequately distinguish between problem and non-problem gamblers, a combination of variables needs to be considered. It is not possible to accurately identify problem gamblers through one variable alone.
A holistic approach to gambling-related harm is needed which considers both the individual player, and the way that different types of gaming play interact. It’s not enough to apply thresholds across all players – what is harmful for one player may be perfectly healthy play for another.
Implications for demonstrating policy compliance
The data analysis showed that machine learning systems capable of cutting-edge pattern recognition can have a crucial impact on identifying players at risk of gambling harm. By automatically spotting the early warning signs of unhealthy behaviour in player data, Gaming organisations have the opportunity to intervene with individual players, minimising the risk of further harm.
Another key finding of the research was that there was not a single model for identifying problem gamblers – multiple models produced a similar level of accuracy. This demonstrated that there was not a homogenous pattern that identified all problem gamblers, indicating that there are different forms of problem gambling that exhibit different behaviours. This has interesting implications for policy, because it indicates that it is not just one type of behaviour that needs to be identified to spot all players at risk of harm, but a range of behaviours to identify the different forms of problem gambling amongst the population.
Currently, policy frameworks are not always set up well for identifying this type of situation, where multiple and varying factors need consideration. However, Gaming organisations have an opportunity to embrace new technological approaches to be able to understand each of their players in real-time. In this way, they can monitor multiple types of behaviour, demonstrating to regulators that they have control over meeting responsible gambling compliance to protect their players from harm.
A globally-used analytical approach
There are of course, much wider compliance implications for an Adaptive Behavioural Analytics approach beyond the Gaming industry – from banks fighting the increasing velocity of fraud (while meeting FCA mandates to treat customers fairly), to monitoring insider trading, and creating modern fraud protection for new mobile payment platforms. There have even been investigations into applications in the Health sector, for example monitoring skin lesion changes for early warning signs of cancer.
Now that individual behaviours can be understood, it’s up to organisations to make use of them, proactively indicating their ability to meet policy requirements to the government, and offering a better service to their customers.
David Excell
Co-founder and CTO, Featurespace
David is Co-founder of Featurespace. He has over seventeen years of experience transferring technology into practical business applications, and under his leadership Featurespace has grown from a concept to a commercial success with many blue-chip customers. David has been awarded more than 11 prizes and scholarships for his academic and commercial achievements, including the 2011 ITC Enterprise Award for Young Entrepreneur.
1 comment
Comment by Pankaj posted on
Interesting article, David!
I am wondering how would you define Problem Gambling in data terms (in a simple yet robust way). Could something like High Spenders - High Frequency players be classified as Problem Gamblers? If yes, how would you go about defining thresh holds for 'High'?
Many thanks!