Breaking Through the Noise
Regulators are requiring gaming companies to implement systems that predict problem gambling behavior. But how effective are they? Kahlil Simeon-Rose asks have we hit a ceiling on the usefulness of responsible gambling algorithms?
Regulators are increasingly expecting gambling operators to implement risk-detection systems that support interventions for high-risk players. Germany’s 2021 Interstate Treaty on Gambling, for instance, mandated an automated system for the early detection of at-risk players. Similarly, the UK Gambling Commission has long required licensees to monitor customer activity for “markers of harm,” and in New Jersey, the Division of Gaming Enforcement’s best practices require online operators to use automated triggers to flag high-risk behavior. While these rules are forward-looking, the limitations of the underlying systems are not well understood by many stakeholders.
Consider for example that the world’s best gambling risk-detection algorithms are about 80 percent accurate in predicting which players are at higher risk. However, these models are often trained using data categorized by tools like the Problem Gambling Severity Index, which is itself a very noisy measure, often only about 80 percent accurate in predicting the results of a formal clinical assessment. Assuming these probabilities are independent, the best risk-detection models are only about 64 percent accurate in identifying truly high-risk individuals, a surprisingly underwhelming result.
One might argue that a “noisy” model is acceptable; it’s better to contact more players with false positives than to miss someone genuinely at risk. But this view misses the point. The entire value proposition of these models is accurate prediction that enables the efficient provision of support. If the goal were simply to contact everyone potentially at risk, we could discard the system and target every customer with every tool. A bad risk-detection model, at its limit, does exactly that.
Can Shifting to Session-Level Feedback Make a Difference?
We need to re-evaluate what we are trying to accomplish with risk-detection algorithms. Predicting a clinical level of harm years in the future seems infeasible. The best signals tend to emerge when an individual is already experiencing harms. For example, one of the strongest indicators is multiple deposits in the same day, which clearly overlays onto ‘chasing losses’ criteria that appears in assessment tools.
The fundamental challenge is that we lack a clear understanding of what separates a positive gambling experience from a negative one, especially early on. Two people can have nearly identical play patterns at the start of their gambling journeys, and identifying the key distinctions that lead one toward harm remains an open question.
A breakthrough for the field would be to develop an (even imprecise) target variable that could be measured after every single session. Current models rely on outcome variables like self-exclusion, which may happen only once in a lifetime and thus provide very little signal on a given behavior. In contrast, a session-based indicator would generate a training dataset that is multiple orders of magnitude larger.
The analogy here is the “next-word prediction” that powers Large Language Models. These algorithms work so well because they learn from the immense scale of written text. Looking at a high-frequency outcome (the next word) allows the model to learn effectively from a constant stream of data. With each gambling session providing a new data point, a responsible gambling model could iterate and refine its understanding of harmful patterns at a vastly accelerated rate. This approach shifts the focus from a long-term, elusive prediction to an immediate, actionable assessment of in-the-moment impact, which could lead to more timely and nuanced interventions.
How to Build It?
To make such an approach work, we need to define new, high-frequency outcome variables. An obvious place to start would be a metric like “session satisfaction,” captured through a simple post-session prompt, much like a ride-sharing app asks for a rating after every trip that seeks to understand whether expectations were met. While any single rating is a noisy signal, gathering hundreds or even thousands of data points per active user, rather than just one or two over their lifetime, would allow a machine learning model to learn at an unprecedented pace.
This system could also provide richer insights. To torture the ride-sharing analogy further, just as a user who gives less than five stars is prompted for specific reasons, a player reporting a negative experience could be asked to select from a list of contributing factors. This could help distinguish between simple disappointment over a financial loss and potential early markers of harm, like a player projecting myths about game outcomes. The Pathways Model of Problem and Pathological Gambling suggests that the development of a gambling addiction is a long and multi-step process for players, implying that there must be an approach that changes an individual’s future risk profile. Developing an even imprecise target variable that could be measured every session would be a breakthrough.
Ultimately, these are projects that will need to be pursued collaboratively. Regulators, operators, and independent researchers each can create value. Operators have the behavioral data and the platforms to test these novel, high-frequency feedback systems. Researchers can provide the scientific rigor needed to design valid experiments and analyze the resulting data, ensuring that the new metrics are meaningful and effective. Regulators, in turn, can foster this innovation by creating frameworks, such as regulatory sandboxes or secure data aggregation, that allow for experimentation. This tripartite partnership is the most promising path to breaking the current impasse and developing a new generation of responsible gambling tools that are timely, nuanced, and more effective at preventing harm.
Kahlil Simeon-Rose is an Associate Professor in the Carson College of Business at Washington State University. His research focuses on public policy and consumer behavior in gambling. His background includes industry roles as a professional economist and in responsible gambling management.
