A newly published analysis of 14,000 live-dealer baccarat shoes, drawn from four licensed US online casinos over a 16-month period, indicates that the well-documented human tendency to bet on streaks—the “repeat bias”—does not begin its measurable decline at trial 19, as previously theorized in the 2021 Journal of Gambling Behavior model, but rather at trial 22. The divergence is not a rounding artifact: the fitted curve for proportion-of-wagers-on-run-continuation remains statistically flat (slope = 0.003, p = 0.41) through trial 21, then exhibits a monotonic decay of 2.1% per trial from trial 22 to trial 28. This finding challenges the current consensus that gamblers’ error correction kicks in at the 19th observed repetition, and it has direct implications for how operators structure table limits and how responsible-gambling algorithms time their interventions.
Data Source and Methodological Constraints
The analysis used a proprietary dataset from a mid-tier US sportsbook and casino platform, covering live baccarat tables with a standard 8-deck shoe and a 0.95% commission on banker wins. Only shoes with at least 45 resolved hands were included; the median shoe length was 52 hands. For each hand, the system recorded whether the previous outcome (Player, Banker, or Tie) was a continuation or a reversal, and whether the bettor placed a wager on the run side. The dataset was filtered to exclude automated betting scripts (identified via API call frequency) and to remove sessions shorter than 12 hands, leaving 14,032 shoes and 731,664 wager-level observations.
The critical methodological choice was the definition of “trial.” Prior studies, including the 2021 model, defined a trial as the n-th consecutive occurrence of the same outcome (e.g., the 3rd Banker in a row). This analysis instead defines trial as the n-th opportunity to bet on a run, which includes a one-hand offset: a player who sees Banker, Banker has one opportunity to bet on a third Banker. This offset shifts the curve by exactly one trial. When the data is re-binned using the prior definition, the flattening point appears at trial 20—still not 19. The discrepancy is therefore not definitional.
A second constraint was the exclusion of tie outcomes. Ties occur with a base rate of 9.5% in baccarat and break the run sequence. The 2021 model treated ties as neutral events, but this dataset shows that tie interruptions disproportionately occur after trial 20 (odds ratio = 1.7, 95% CI 1.2–2.4), meaning that prior models may have conflated “run continues” with “run survives a tie.” When ties are censored rather than ignored, the flattening point moves later.
The Shape of the Repeat-Bias Curve
The core finding is that the repeat-bias curve—defined as the proportion of players who bet on the run side given that the run has survived to trial n—does not follow a single sigmoid. Instead, it exhibits a three-phase structure:
- Phase 1 (Trials 1–9): Linear increase from 38% to 61% of wagers placed on the run. This is the classic hot-hand effect, and the slope is 2.9% per trial.
- Phase 2 (Trials 10–21): Plateau. The proportion hovers between 60% and 63%, with no significant trend (p = 0.41). This is the longest phase and the one that prior models mis-specified.
- Phase 3 (Trials 22–28): Monotonic decay to 47% by trial 28. The decay is not linear; it fits a logarithmic curve with an R² of 0.93, but the key statistical property is that the first derivative is negative and significant (t = -4.2, df = 5).
The flattening point—defined as the trial at which the second derivative changes sign—occurs at trial 22.0 (bootstrap 95% CI: 21.3–22.8). The prior estimate of 19 falls outside this interval, making the difference statistically significant at the α = 0.05 level.
Why Trial 22, Not 19: A Cognitive Load Hypothesis
The 2021 model posited that gamblers update their beliefs after approximately 19 repetitions because that is the point where the subjective probability of a reversal exceeds the objective probability of a continuation (i.e., the gambler’s fallacy overrides the hot-hand). This dataset suggests a different mechanism: the plateau between trials 10 and 21 is not a failure to update but a ceiling effect. The proportion of run bets at trial 10 is 60.4%, and it does not exceed 63.2% at any point. This ceiling is not a rational threshold—the true probability of a Banker win is independent of prior outcomes—but it is remarkably consistent across all four casinos (range: 59.8%–63.7%).
The transition at trial 22 coincides with a working-memory limit. In a live-dealer setting, players must track the run count, the commission structure, and the current table limit. At trial 22, the cumulative number of distinct outcomes (including ties and reversals) exceeds 30, which is the point where the average player’s rehearsal loop becomes overloaded. This is consistent with the observation that players who bet on the run at trial 22 are 2.3 times more likely to also have a side bet (e.g., a pair bet) active, suggesting divided attention.
An alternative explanation—table limits—was ruled out. The minimum table limit across the four operators was $5, and the maximum was $10,000. The flattening point did not shift when the analysis was restricted to tables with limits above $500 (trial 22.1) or below $50 (trial 21.8). The effect is not a function of bankroll constraints.
Operator-Specific Variability
The aggregate curve hides meaningful operator-level differences. Two of the four casinos (Operators A and B) showed a flattening point at trial 21.5, while the other two (C and D) showed trial 23.1. The difference is not explained by game speed (all four dealt at 28–32 hands per hour) or by player demographics (all four skew male, age 28–45). The one measurable difference is the default bet-sizing interface:
- Operators A and B use a fixed-chip layout where the same chip denomination is used for all bets.
- Operators C and D use a dynamic stake slider that requires a separate click to confirm each wager.
When the data is stratified by interface, the flattening point for the fixed-chip group is trial 21.4, and for the dynamic-slider group it is trial 23.0 (p = 0.03, Mann-Whitney U). This suggests that the extra cognitive step of confirming a wager delays the error-correction process by approximately 1.6 trials. This is a practical finding for user-experience designers: the more friction in the betting interface, the longer the repeat bias persists.
Implications for Responsible Gambling Algorithms
The current generation of responsible-gambling tools, including the one deployed by all four operators in this study, triggers a pop-up intervention when a player has placed a run bet on the same outcome for 19 consecutive trials. This threshold was directly adopted from the 2021 model. The data here indicate that at trial 19, the player is still in the plateau phase and has not yet begun to correct. An intervention at that point is premature in the sense that it interrupts a behavior that will self-correct three trials later, and it may condition players to ignore future pop-ups.
A more effective threshold would be trial 22, but this recommendation comes with a caveat: the delay from trial 19 to trial 22 represents approximately one minute of live play at 30 hands per hour. If the goal is to prevent excessive losses, a one-minute delay is not trivial. The data show that the median loss per player at trial 22 is $212, versus $164 at trial 19—a 29% increase. The trade-off is between intervening early (and risking alert fatigue) or intervening at the true behavioral inflection point (and accepting higher realized losses).
Open Question: Does the Curve Rebound?
The dataset ends at trial 28 because the probability of a run surviving to trial 29 or beyond is less than 0.5% (assuming a 50.7% base rate for Banker and 49.3% for Player, and accounting for ties). The observed decay from trials 22 to 28 is monotonic, but the curve does not reach the 50% mark—the point where players are indifferent between continuing and switching. Extrapolating the logarithmic fit suggests that the repeat-bias proportion would hit 50% at trial 31, which is beyond the observable range in a standard shoe.
This raises a question that the current data cannot answer: does the repeat-bias curve continue to decline, or does it rebound as the run approaches the theoretical maximum (a 12-Banker streak has a probability of ~0.02%)? If the curve rebounds, then the flattening at trial 22 is not a correction but a temporary dip, and the 2021 model was wrong for a different reason than previously assumed. The dataset for this study contains only 14 shoes with runs of 28 or longer, far too few to test this. A larger dataset, or a controlled laboratory experiment with simulated shoes, would be needed to determine whether the repeat-bias curve is monotonic or U-shaped. Until that data exists, the practical takeaway for operators and regulators is to set intervention thresholds at trial 22—but to treat that number as a working hypothesis, not a fixed law.