Insights

Singapore Pools uses AI to enhance player safety

By Matthew Spinks, I.T., Data Analysis & Research Consultant and Li Jing, Data Scientist, Singapore Pools

Author
Matthew Spinks & Li Jing
Published

Online betting and gaming come with a wealth of player analytics that can be used to understand player behavior and engagement. The data scientists at Singapore Pools have harnessed the power of player analytics, and artificial intelligence (AI) to develop a program that promotes responsible gaming by gently ‘nudging’ players in the right direction when short-term deviations from normal gaming behavior like anomalous spend patterns are detected. The Nudge Model application won the ‘Best Flagship Award’ for Responsible Gaming at the 2022 World Lottery Summit.

The emergence of online gaming

It is widely recognized that people around the world enjoy various forms of gaming, including lotteries, sports betting, and casino-style games. In recent years, the advent of the Internet and Internet-enabled technologies such as tablets, smartphones, and the World Wide Web has led to a proliferation of online gaming platforms, whereby people can play games for wagers and bet on sporting events at their convenience and from the privacy of their homes. This has resulted in an entirely new digital delivery channel, which has enjoyed significant growth over the past decade.

As a channel, online gaming is differentiated from in-person gaming – viz. gaming in-person at a physical outlet like a convenience store (where a customer buys lottery products) or casino – and has its own distinct characteristics. In particular, the ease and convenience of access to online gaming, together with the ability to place bets in rapid succession with little to no effort, mean that online gaming has the potential for problematic play. It is well known that, for a minority of people, gaming can be problematic to the point of addiction.  

Responsible Gaming (RG) occurs in a regulated environment where the potential for harm associated with gambling is minimized, typically by enabling people to make informed decisions about their participation in betting and wagering. Preventing problem gaming through RG is the most critical element of the lottery and sports betting sector's commitment to Corporate Social Responsibility (CSR). By facilitating responsible play, WLA lottery and sports betting members grow their sales – and hence returns to good causes – in an ethical and principled way. In addition, by channeling the natural impulse for gaming into regulated gaming markets, they play a crucial role in fostering the continuity of public order and the fight against illegal gaming and match fixing.

Ensuring responsible play in Singapore

Singapore Pools is Singapore’s sole legal operator of lottery, sports betting, and horse race wagering, whose surplus from gaming turnover is returned to the community and public good. Formed in 1968 to combat illegal gaming activities and to channel surplus earnings to benefit the community responsible gaming has been a pillar of Singapore Pools’ regulated gaming offer since its inception. As such, the company institutes a comprehensive range of measures to encourage responsible play, including restricting play to adults, prohibiting betting on credit, and providing safe (regulated and alcohol- and smoke-free) environments for wagering. Other RG measures actively promoted by Singapore Pools include safer play reminders such as “Play Responsibly” messages in all the company’s communications, and self-control mechanisms such as setting deposit and betting limits, as well as voluntary self-exclusion.

For the small cadre of problem gamblers, voluntary self-exclusion offers gamblers a choice to ban themselves from particular land-based gaming venues, or online gaming platforms. This individualized (as opposed to a population-wide) harm reduction measure aimed at preventing gamblers from further financial, social, and psychological distress, has some measure of success. Nonetheless, self-exclusion is a ‘blunt instrument’ that comes with a number of documented inadequacies. For example, because the decision to self-exclude is an individual choice, customers that should self-exclude may not. Equally, self-excluded customers may continue to gamble at providers that are not covered by the exclusion.  

In Singapore, the issues associated with self-exclusion are seen in practice. Individuals may apply to exclude themselves from gambling activities through the National Council on Problem Gambling (NCPG). However, such customers make up a very small subset of the customer base. Typically, individuals do not recognize that they are at risk of spending above their means; or, if they do, they often lack the self-control to act upon that knowledge. This situation gives rise to a challenge when it comes to identifying potentially problematic gaming behavior automatically, as it creates a so-called class imbalance problem that hinders the accuracy and effectiveness of traditional supervised machine learning approaches.

Since 2016, Singapore Pools’ players have been able to play the lottery remotely and bet on sports and horse racing digitally, via individual online accounts. Singapore Pools was aware of the potential impact of problem gaming arising from the digital channel, and the limited efficacy of self-exclusion in particular, as a mechanism to support problem gamers. In 2018, it set out to find an alternative way of implementing harm reduction measures for assisting gamers at risk from online gambling. Exploiting player analytics gathered through its online gaming portal, Singapore Pools developed an application, called the ‘Nudge Model’, which applies AI techniques to player analytics to generate insights into player behavior. These insights are fed back to players as a series of prompts that ‘nudge’ them into adopting good habits of responsible play.

Extensive analysis and evaluation of the Nudge Model application has demonstrated its efficacy since its production launch at the end of 2020.  

The Nudge Model

The idea of the Nudge Model is simplicity itself. Players are assigned a gaming profile, based on observed transaction data and analysis. The gaming profile is updated every month using a ‘traffic light’ system featuring a green, amber, or red risk indicator bar and corresponding nudge messages, encouraging safe gambling habits:

Customers can view their zone and the corresponding nudge message in both a pop-up message and layer dashboard when they login to their account on any platform.

The Singapore Pools’ online player account dashboard front-end showing a Nudge Model Yellow Zone message. (Parts of the dashboard are redacted to preserve commercial-in-confidence.) The Nudge Model message is conveyed to end users in their preferred language as per their personal account settings.

Internationally recognized, the Nudge Model won ‘Best Flagship Award’ for Responsible Gaming at the World Lottery Summit 2022. It was hailed by the judging panel as a ‘great submission, using data technology to better understand player behavior as part of their RG commitments’, and further praised Singapore Pools’ core commitment to CSR and RG for their ‘very dedicated [RG] program’ and the Nudge Model as ‘a great example for the use of advanced technology’ to support the playing public.

Origins behind the application

Besides ease and convenience of play, online gaming also distinguishes itself from play through the bricks and mortar retail network by enabling the collection of player analytics, such as information about player engagement. With physical gaming, such data tends to be either largely anonymous or simply unavailable. Following the decision by Singapore Pools to initiate the Nudge Model project, the company’s data scientists, at the time, all led by Leemon, Senior Director, Corporate Strategy and Customers, were tasked with leveraging the power of player analytics to implement new harm reduction measures for players who could be at risk from online gambling.

The team began by exploring existing uses of AI driven by player analytics in promoting RG best practices. They found that most AI models exploiting player analytics are trained with customer betting behavior data derived mainly or exclusively from Western markets.

This was potentially problematic, because it did not reflect local market conditions. For instance, there could be differences in the betting behavior in Western markets compared to that in Singapore. A more serious shortcoming in existing AI applications was that, generally speaking, they did not sufficiently consider seasonal variations in player behavior. For example, gamblers’ betting behavior can change over time as their financial situation and the prize pool, matches or races being offered change; the FIFA World Cup soccer tournament held every four years is the exemplar of an external event driving significant changes in seasonal play.

What is unsupervised learning?

In unsupervised learning, aka unsupervised machine learning, a computer program uses unlabeled data to learn by itself without any human supervision or intervention. Roughly speaking, unlabeled data is data that has not been classified, characterized, or otherwise identified. It does this by analyzing and clustering the unlabeled data sets to identify or observe hidden patterns or groupings in the pool of data. The ability of unsupervised learning to identify similarities and differences in information makes it the ideal solution for applications in (for example) image and pattern recognition, for customer segmentation applications, and for exploratory data analysis.

Following this review, Singapore Pools decided to build its own in-house RG model. Additionally, it concluded, as a result of a literature survey, that unsupervised learning was the best approach to capture the subtleties in player behavior, leading to the Nudge Model.

How the Nudge Model works

The Nudge Model uses unsupervised learning methods to train the transactional data of Singapore Pools’ online customers and form dynamic probabilistic cohorts. If any players have a spending trend vastly different from their designated cohort, a nudge message is sent to encourage them to moderate their spending.

The Nudge Model selects those customers who have: (i) had their player account with Singapore Pools for at least 12 weeks; and (ii) who have placed at least one bet in the past 12 weeks. It then groups the resulting customer base into several probabilistic cohorts using an unsupervised clustering algorithm. In particular, the algorithm forms two layers of cohorts: a first layer cohort, based on customers’ product preferences; and then a second layer cohort for all the first layer cohorts based on customer spend in each of the past 12 weeks.

The model then computes the customer’s and cohort’s spending slope separately, by fitting a regression line to the last six weeks’ weekly spending. It then further computes the deviation between the customer and the cohort, which is the difference between their spending slopes. Finally, it computes the deviation score for each customer.

The essence of the algorithm.

A schematic illustration of the Nudge Model architecture, the core of which is contained within the indicated bounding box.

Based on the customer deviation score, customers are classified into the three ‘traffic light zones’ identified above, viz.:

  • Green zone. Player behavior is normal compared to cohort.
  • Amber zone. Player behavior is flagged as potentially concerning compared to cohort.
  • Red zone. Player behavior is flagged as potentially problematic compared to cohort.

The assigned zones are refreshed each month. Customers then receive the same nudge message until the next month refresh.  

Evaluation of results

In each month, customers are clustered into 30 to 35 cohorts in total. On average, 83% of them are in the Green Zone, 15% are in the Amber Zone, and 2% are in the Red Zone. Commenting on the efficacy of the ‘traffic light’ system for nudge messages, Singapore Pools data scientist and project lead Li Jing said, “Based on an analysis we have conducted, receiving a Red or Amber Zone nudge message led the customers to drop their spending in the same month by respectively 37% and 20% on average. Besides the total spending, both the bet frequency and the frequency with which customers attend a draw, a match, or a race also dropped significantly.”

Empirical evidence shows that nudge messages work best for sports betting and the least well for lottery. Interpretation of the data suggests that this is because the Red Zone average spend on sports tends to be much higher than the average spend on lottery. Receiving a Red Zone or Amber Zone nudge message led customers to drop their sports spend by respectively 39% and 28% on average respectively and to drop their lottery spend by 31% and 15% on average respectively in the same month. These results are statistically significant.

The data scientists at Singapore Pools have invested heavily in evaluating and validating the Nudge Model application. For instance, to validate the analysis result, one month an A/B testing was conducted by randomly choosing a small sample of customers classified as Red Zone customers and placing them in the Green Zone instead; by randomly serving end users two versions of the player portal that differ only in the selection of Nudge Model Zone, the relative efficacy of the Zones and hence the Nudge Model messages can be measured. The team found that the group that received the Red Zone nudge message dropped its spending 26.6 percentage points more than customers in the sample group that received the Green Zone nudge message.

Five percent of customers fall into the Red Zone repeatedly. This indicates that nudge messages not only have the potential to assist in moderating customer spending in the short term, but also to modify behavior in the longer term through continuous reinforcement of good behaviors via nudge message reminders. Commenting on the potential for longer term application, Li Jing said, “When we first presented the model internally at Singapore Pools, one concern raised was the hypothesis that customers could increase their spending, receive the nudge message once and then migrate to a high spending cohort. They might do this repeatedly, migrating to a higher spending cohort again, and thus ‘normalize’ their increased spending pattern in the long term. However, when we analyzed the data, we could see that this phenomenon did not happen to our repeated Red Zone customers. The compounded monthly growth rate among those customers is statically significantly lower than customers who have never been in the Red Zone or in the Red Zone exactly once.”

The data science team at Singapore Pools has also monitored the model for illegal operator cannibalization. Account closure rates among customers that are in the Red Zone are tracked closely. To date there has been no sign showing that nudge messages encourage customers to close their Singapore Pools account and bet with illegal operators.

A survey has also been undertaken asking Singapore Pools’ customers whether nudge messages influence behavior. More than 2,000 customers participated in the survey and 74% agreed that the nudge messages encouraged them to bet responsibly.

Outcomes and next steps

The Nudge Model represents a significant improvement over conventional methods to address problem gaming:

  • Self-exclusion is a key tool for controlling problem gaming using conventional approaches to responsible play. However, individuals may not recognize that they are at risk. The Nudge Model addresses this by inferring the trajectory of player gaming from exhibited behavior.
  • Self-excluded customers traditionally make up a very small subset of a gaming entity’s player base, creating a class imbalance problem. The two-layer clustering approach of the Nudge Model addresses this by grouping customers who place bets on similar products and with similar spending patterns, thereby identifying low, medium, or high deviations relative to the cohorts in question.
  • Conventional approaches to problem gaming are limited in their ability to reflect seasonal behavior. In contrast, the Nudge Model recognizes that the intensity of play by seasonal customers varies considerably and may even be quite high during specific periods like the FIFA World Cup. As customers are compared against their cohort, this prevents seasonal customers from being categorized as potential problem gamblers.
  • By the time players reach the stage of requesting self-exclusion, they may already have suffered significant negative personal, family, or other social consequences. The Nudge Model aims to pick up early signs of high-risk betting behavior before said customers suffer distress and ‘nudge’ them back to a safer level of gaming. If customers continue to exhibit at-risk behavior (i.e. stay in the Red Zone), direct interventions can also be made.

To further enhance the responsible gambling initiatives, the data scientists have planned additional analyses and interventions targeting customers who have consistently been in the Red zone for multiple months.

Another current limitation of the Nudge Model is that all customers within the same zone receive the same nudge message. One possibility is to customize nudge messages even further based on the customer’s demographics, assigned segment, product preferences, or any other factors. This could also be useful in the future to change messages from time-to-time to reduce message saturation.

About

Media inquiries

Legal notice