CO-AUTHORED BY NOAH ACRES AND STEPHEN MOORE
In this second in a series of articles regarding the role of AI in the gaming industry, Differential Labs focuses on the building block of AI: data. More specifically it discusses the shortcomings of current casino data and the promise of new data sources. Differential Labs Managing Director Clayton Peister is joined by Noah Acres of Acres and Stephen Moore of Walker Digital Table Systems.
When we speak to operators about AI adoption, we often start with a simple refrain: casino data is always wrong, but it is typically consistently wrong which makes it useful. While it is true that current casino data is useful, there is much we do not understand about patrons. Start with valuation: the current lingua franca used to describe a patron’s value – ADT – is typically wrong 80% of the time.
ADW, a measure which combines actual win-loss and theorical win, is hardly any better. As a consequence, the way we (as operators) think that players interact with the property is wildly different from the patron’s personal experience. By extension, the most meaningful impacts other industries are gaining from AI have largely missed the casino industry. This need not be the case anymore: new source data, in particular the bet-by-bet transaction exhaust from smart-tables, slot systems with real-time data collection and online gaming hold the promise of deeper patron intimacy. We believe these new data are the missing puzzle pieces required to personalize at scale and unlock AI’s most ambitious promises.
Specifically, these data sources power an understanding of player preferences, more accurate view of player behavior and more real-time and relevant results. Predictive models built on top of these new data are not only more accurate but can also capture player preferences in ways difficult to imagine in the past. Savvy operators will embed the resulting models on top of streaming bet transactions unlocking new operational optimization, fraud detection and triggering real-time actions.
Recalling that we see bet-by-bet transactions enabling better patron engagement along the dimensions of accuracy, preferences and more relevant real-time data, table data improves along all three of these dimensions while slot data will principally improve in terms of preference and relevancy. For instance, in slots today we only know total values per session while variation inside that slot session, like linage and volume across the pay-table, is unknown; bet-by-bet slot data illuminates these behaviors.
Take accuracy
Of every wager made – be it on slots, online or at a table game – the house expects to win the edge of the game. In Asia, most operators assume that the edge averages around 1.25% while we can see in bet-by-bet data that the real edge is around 1.7% and the median or typical win rate is 2.6%. In fact, less than 10% of baccarat patrons have a win rate between 1.1% and 1.3%. In addition, properties typically also understate games played and average wager dollars for table action. Next time a player complains about being under-rated, maybe believe them.
Building models on top of inaccurate ratings is fraught and sometimes a fool’s errand. We recently ran an experiment with an operator adopting smart table technology who was still running manual ratings. We trained three models to predict a patron’s daily spend (average daily win/loss) over the long run. We only considered patrons with multiple days in order to lessen the impact of volatility.
The first model we developed (let’s call it the baseline model) was a state-of-the-art model with all the knowledge we have gathered over the last 10 years regarding increasing accuracy. The second model was a simple moving-average but using bet-by-bet data from the smart tables, while the third model was a heavily tuned state-of-the-art predictive model built on top of bet-by-bet transactions. The results are summarized below:
Model 1
For ratings, the first model learned the inaccuracies incumbent in the data. For table dominant players, the model built on top of ratings was 30% to 40% wrong over the long-term. On slot ratings the “average” had a low error rate, but volatility still weighed on the accuracy for the typical (median) player where the error rate was at least 20%. Perhaps more importantly there was a weak relationship between the ratings’ theoretical and the players actual daily spend and the correlation weakened further when the player had a high average bet or frequently took side-bets.
Model 2
The simple model built on top of the bet-by-bet transactions handily beat the ratings model, reducing the error from 30% to 40% to closer to a 10% error.
Model 3
A state-of-the-art model built on bet-by-bet transactions reduces this error further but not materially; more importantly the accuracy was far more nuanced capturing concepts like player potential.
The takeaway from this experiment is to emphasize the importance of the more pedestrian subject of data quality which underpins the more glamorous predictive algorithms. As AI practitioners, we are often a technical bunch enthralled by elegant math, and we tend to focus on the more interesting elements of data science while real gains in accuracy come from data quality and feature engineering. Perhaps more importantly, the error in ratings for typical patrons makes the concept of AI driven personalization out of reach.
While the subject of a forthcoming article, accuracy plus granularity has the promise of revolutionizing both anti-money laundering and responsible gaming. For instance, with RFID- enabled chips we can see chip transferring activity which can illuminate money laundering, loan sharking and even promo abuse. On slots we can see sudden variation in play relative to a feature occurring: indicative of abusive advantage play.
Preferences
A good host knows their player’s gaming preferences almost instinctively; often characterizing players as ping-pong bettors or describing risk preferences in a rainbow of different terms. Bet-by-bet transactions enable us to draw a more complete picture of all players across the floor. For instance, baccarat players are notoriously trend-oriented but allegiance to the trend varies distinctively across ethnic lines. Continuing with the baccarat example, we can see within smart table data that when four “Banker” coups occur in a row that players from sino-sphere countries will bet “Banker” 80% of the time. Southeast Asian players, by contrast, will only bet “Banker” ~65% of the time. Likewise on slots, previously we never knew how or if a player changes their behavior after hitting a feature or getting into a bonus round with a small to no payoff. Incorporating these many splendid preferences into marketing initiatives is an obvious application.
Less obvious but perhaps just as powerful are the implications for fraud and risk. For instance, with the smart table data we began to see the baccarat “side bet advantage play” risk back in 2020. Side bet advantage play has since proliferated across Asia and is accelerating in conjunction with new bet options. With smart tables we are counting the shoe and looking for unusual play variation relative to the count all in real-time. This issue is even more pronounced online where operators often offer rebates based on volume, making even base bets susceptible to arbitrage play; a weakness exploited by bots. The Russian hacker scam is the slot version of an abusive player which can be addressed only by granular data.
Relevance
The typical player has five to 10 rating sessions per day across a couple hours of play, making responsiveness to the player’s current situation diffuse and retroactive. Bet-by-bet transactions hold great promise for situational actions like updating leaderboards, marketing challenges or fraud detection. By extension, real-time data enables operators to determine the best time for making interventions based on the patron’s current situation.
Resolution
Pundits are positioning AI as a crystal ball capable of solving any flavor of problem, but until data fundamentals catch-up, results will continue to be confined to specific and narrow use cases. New data sources – most prominently bet-by-bet transactions – could be the missing piece in the casino AI puzzle.
