Could AI prevent problem gambling?
Many gambling companies use AI/machine learning to entice customers to buy more. But it is also possible to use the new technology for the opposite purpose.
Companies such as Amazon, Netflix and Facebook collect data on their customers and create models that identify each individual customer/user’s preferences and needs. This enables them to know which books, films, posts or news articles that are likely to appeal to each individual customer. Commercially speaking, this means increased sales, repeat sales, clicks and likes, and it enhances the customer experience and boosts their loyalty.
Machine learning is the learning engine under the bonnet of AI. In simple terms, machine learning means the computer can independently search for patterns in large quantities of data and develop models of analysis without being specifically told what to look out for. For example, that can encompass finding customers with the same preferences, like in the case above.
Machine learning has flourished in recent years. Digitisation has led to increased access to data, and at the same time there has been a big improvement in the computing power that is needed to make use of the method.
Possibilities and risks of new technology
Like other tools, technology can be used in different ways. Unfortunately, in the international gambling industry there are examples of AI being used cynically to exploit vulnerable individuals.
Two articles published in The Guardian newspaper address the unethical use of this new technology. The first article reveals how gambling firms buy lists of potential customers from third parties, where the specification shows they are individuals with low income, own at least one credit card, and even indicates those who have self-excluded from gambling. The second article looks at inside information on how gambling companies create models to determine exactly when and where customers will be triggered into gambling, so that they can draw as much money as possible from each individual customer. They know, for example, when it is payday, and which day and times to call with offers on free games or bonuses to entice customers to play again.
AI can be used to prevent problem gambling
All 2.2 million Norsk Tipping customers have passed identity checks and therefore have a unique gaming account. Whenever customers play a game, receive winnings, set gaming limits or take a break, this activity is saved on their customer profile. This wealthy supply of data provides unique possibilities to exploit AI to prevent problem gambling.
Norsk Tipping’s mandate is clear: the company shall act to prevent the negative consequences of gambling. There are two areas in particular where machine learning can improve Norsk Tipping’s ability to prevent problem gambling: exposing risky gaming patterns and personalizing preventive measures to reduce risky gambling.
Exposing risky gambling patterns
Norsk Tipping utilises the analysis tool Playscan to expose risk-filled gaming patterns. Playscan also provides the customer with feedback which explains the reasoning behind the risk assessment and suggests preventive measures where appropriate.
The first generation of Playscan was created by experts on responsible gambling, who used research on behavioural patterns associated with risky or problematic gambling behaviour as their basis for an algorithm. In addition, the tool features a self-assessment in which customers answer 16 questions on their own gambling habits to detect the level of risk associated with gaming.
The challenge of creating a model to detect problem gambling is identifying what the model should look out for. How can we know that one customer has a good relationship with gambling, whilst another has gambling problems? Naturally, we have neither access to the customer’s inner thoughts nor to the registry of those with diagnosed problems. The latter would not help either, seeing as it is only a minority of problem gamblers who seek help from their GP.
After being on the market for several years, the first Playscan model had collected a large number of self-assessments, which could then be used to develop the second-generation model. A red self-assessment means the customer personally acknowledged they had a gambling problem. By using the customer’s data for the period prior to their response to the self-assessment questionnaire, it was possible to train the model. More than 60,000 completed self-assessments were used to develop the second generation of the Playscan model.
The development of Playscan over time emphasises an essential requirement for succeeding with machine learning: the significance of investing in data. By requesting customers to take the self-assessment, enough responses were collected to create a database large enough to train the model against. It can also be a challenge to develop the model against self-assessments, i.e. recognising under-reporting by participants. However, the new model is clearly better than the previous one. It has a higher level of accuracy (ROC curve/AUC) and the customer feedback shows a greater degree of agreement with the risk assessment.
Personalised initiatives to prevent risky gambling
The area of responsible gaming also offers the potential to create personalised initiatives aimed at players with risky or problematic gambling behaviour. It is not a given that a casino gambler and a sports gambler, or a 20-year-old and a 50-year-old customer, perceive the same initiative as relevant and meaningful.
What is required for personalization like this? Again, in order to succeed, one must ‘invest’ in data. Different initiatives must be tested on a representative sample of customers before it is possible to use machine learning to identify which customers make use of which initiatives.
Case study: Selecting customers for a proactive call
One of Norsk Tipping’s initiatives targeted at customers with risky/problematic gambling habits is a so-called proactive call. During a phone call, the customer is given facts on their gambling spending and the need for changes is discussed. In the majority of the calls, at least one initiative to reduce their gaming is agreed. The most common measure is the reduction of their gaming limit.
The annual capacity for these proactive calls is much lower than the potential target group. The question in this situation is: Given the limited capacity, which customers should be prioritised?
Some of the customers we phoned were gaming on behalf of a group of players. Other customers gave us the feedback that they were in control, had a good income, and that it was an important hobby for them. If we could identify the customers who would benefit from the call, we would be able to reach more customers over the course of the year who needed the nudge towards making a change. We could increase the efficacy of the initiative.
This became the starting point for a study on machine learning that was completed as a project assignment as part of a Master of Management degrees at the BI Norwegian Business School during the spring of 2018.
We used a sample data set of 1,400 customers who had received a proactive call. The customers had been randomly picked from among the 10,000 customers who had lost the most money through gaming in the last year. This meant we had a representative data set with which we could create a model for this specific group of customers. We estimated the effect of the discussion for each customer by comparing their spending (theoretical loss) 12 weeks before the proactive call with their spending in the 12 weeks after.
(1) Customers whose spending had decreased following the discussion.
(2) Customers whose spending had increased or had only slightly decreased.
A review of research informed our selection of relevant data which was to be used for developing the model. For example, we included aggregates of gaming consumption, data on the use of responsible gaming tools, the customer’s gender, age and duration of customer relationship, as well as the number of occasions they had made contact with the customer service team.
Next, the standard procedure for machine learning was employed; the model development and the data evaluation were carried out in DataRobot, which is an automated machine learning tool.
The evaluation showed that we managed to develop several models with an ability to provide rather highly accurate predictions. To a great extent the models were able to identify the customers who made use of the proactive calls.
In the autumn of 2018, we used the model to select customers for the proactive calls and made calls to 300 customers as a result. The evaluation showed that the model worked even when applied to new customers. The result was roughly as expected.
In this case,the practical implications were that more customers who had need of, and benefited from a proactive call, were reached.
Want to know more?
If you want to read more about clinical psychologist Jakob Jonssons research on the effects of receiving a proactive call, you can read his blog here: Big Losers: Long-term effects from a proactive call.
The complete report from the first study is published in the journal Psychology of Addictive Behaviors.
If you are interested in further details relating to this case study, you can read more in this article in the journal Magma (in Norwegian).
About Norsk Tipping:
Norsk Tipping AS (Norsk Tipping) is Norway’s state-owned gaming company with sole rights as operator of several gaming activities in the Norwegian domestic market. The mandate is to offer gaming activities in a safe and secure environment under public control with the aim of preventing negative consequences of gambling.
Norsk Tipping has approximately 2 million players from the adult population of 4 million every year. In 2017 the company had a turnover of approximately 355 million Euro.
Product portfolio:Lottery Games, sports games, scratchcards, online casino and bingo, interactive video terminals.
The Norwegian Gaming Authority supervises the company’s games operation.