Responsible Gambling Popups Fire 2.3s After the Third Loss

A Zagreb-based behavioural research team that supplies harm-reduction tooling to three licensed Croatian operators has published timing data showing that a specific class of responsible gambling popup fires a median of 2.3 seconds after a player's third consecutive losing spin, rather than at session start, at deposit, or after a fixed number of minutes. The figure comes from 41,000 anonymised sessions logged across 2024 and the first half of 2025, and it is the first Croatian dataset to measure intervention latency rather than intervention frequency. The finding matters because almost every compliance conversation in this market has been about whether the message appears at all, not about when.

The 2.3-second figure is not arbitrary. It is the point at which the underlying model decides a player has crossed from variance into something the operator's own risk engine flags as a pattern. Three losses in a row is unremarkable on a 96% RTP slot — the probability of three consecutive losses on a game with a 96.2% return and flat betting sits somewhere near 0.6%, which happens thousands of times a day across a mid-sized Croatian-facing site. The popup is not triggered by the losses. It is triggered by what the losses do to bet sizing, spin speed, or a combination of the two, and the 2.3 seconds is simply how long the system takes to confirm the pattern after the third loss resolves.

What the timing actually measures

The distinction between trigger and latency is where most of the public discussion has gone wrong. When the research team says 2.3 seconds, they mean the interval between the settlement of the third losing round and the render of the popup on the player's screen. That interval includes the model inference, the decision to intervene, and the client-side delivery. It does not include the decision to classify the session at all, which happens earlier, often at the second loss.

This matters for operators evaluating their own tooling. A vendor promising "real-time intervention" is usually describing the delivery layer, not the classification layer. If the classification is slow — some legacy systems batch risk scoring every 30 or 60 seconds — then a 200ms popup render is meaningless. The Croatian research found that the fastest quartile of operators in the sample classified at 1.1 seconds after the second loss and delivered at 2.0 seconds after the third. The slowest quartile classified at 18 seconds after the third loss and delivered at 22 seconds. Both groups describe themselves as offering real-time intervention.

The 2.3-second median also hides a long tail. Roughly 9% of interventions in the dataset fired more than 12 seconds after the third loss, and in 340 sessions the popup never fired at all despite the session later being flagged by the operator's own retrospective review. Those 340 sessions are the interesting ones, because they suggest the classification layer is not deterministic — it depends on features that were present in some sessions and absent in others, and the research team has not published which features those are.

The three-loss window is narrower than it sounds

Three consecutive losses sounds like a meaningful threshold until you look at the distribution. On a typical Croatian-facing slot portfolio, the median session contains between four and seven losing streaks of three or more. A player spinning 200 times on a 96% RTP game will hit three-in-a-row losses roughly 30 to 40 times. If the popup fired on every one of those, it would be wallpaper. The reason it fires on 2.3 seconds after the third loss in only a small fraction of sessions is that the model is not counting losses — it is watching the derivative of behaviour around them.

The clearest signal in the dataset is spin interval compression. Players who reduce their average spin interval by more than 40% across the three losing spins are 11 times more likely to receive the popup than players whose spin interval is stable. Bet escalation is a weaker signal on its own — a 2x increase in stake after three losses is common and often rational, since a player chasing a bonus round may simply be increasing stake to hit a threshold. The combination of compressed spin interval and any stake increase is what pushes the model over its threshold in most of the flagged sessions.

Croatia's regulatory context makes timing unusually consequential

Croatia's gambling framework, administered through the Ministry of Finance's gambling division, has since the 2020 amendments to the Law on Games of Chance required licensed operators to maintain "responsible gambling measures" without specifying timing. The 2023 secondary legislation tightened reporting requirements — operators must now log interventions and their outcomes, and the logs are subject to inspection — but it still does not mandate a latency threshold. That gap is why the 2.3-second figure is emerging from private research rather than from a regulator's rulebook.

The practical consequence is that Croatian operators have latitude to define their own intervention windows, and most have chosen windows that are defensible in an inspection rather than effective in a session. A popup that fires at minute 30 of a session is easy to log and easy to defend. A popup that fires 2.3 seconds after a specific behavioural pattern is harder to explain to an inspector who is not a data scientist, which is part of why the faster systems tend to be at operators with in-house modelling teams rather than at operators using white-label platforms.

There is also a market-structure issue. A significant share of Croatian players access internationally licensed sites, which sit outside the Ministry of Finance's reporting regime entirely. Those sites are under no obligation to log interventions, and the research team explicitly excluded them from the dataset because the data was not available. So the 2.3-second figure describes the regulated segment only. The unregulated segment, which by most estimates carries a larger share of high-intensity play, is unmeasured.

What the logs show about outcomes

The dataset includes a follow-up window of 24 hours after each intervention, which allows a rough read on whether the popup changes anything. The headline result is modest: players who received the popup at the 2.3-second median reduced their session length by an average of 14 minutes, or about 19% of the median remaining session length at the point of intervention. Players who received the popup in the slow tail — more than 12 seconds — reduced session length by 4 minutes, or about 6%.

That difference is consistent with the intuition that earlier intervention catches players before the decision to continue has consolidated, but it is not clean evidence. The fast-intervention group and the slow-intervention group are not randomly assigned; they differ by operator, by game portfolio, and probably by player demographics. The research team is careful to describe the 14-minute figure as an association, not a causal estimate. A proper test would randomise latency within a single operator's population, and as far as is publicly known, no Croatian operator has run that test.

The design problem nobody has solved

A popup that fires 2.3 seconds after the third loss is a popup that fires while the reels are still spinning on the fourth. That is the design tension the timing data exposes. The faster the intervention, the more it interrupts active play, and the more likely it is to be dismissed reflexively. The dataset shows a dismissal rate of 71% for popups that fire within 5 seconds of a loss, against 54% for popups that fire after a 30-second pause in play. Fast intervention catches the player, but it catches them in a state where they are least likely to read anything.

This is not a new problem in behavioural intervention design, but it is unusually sharp in iGaming because the product is engineered to maintain flow. A slot session has a rhythm, and any interruption to that rhythm is felt as friction. The 71% dismissal rate is the cost of the 2.3-second median, and it is not obvious that it is worth paying. An operator could reasonably argue that a 54% dismissal rate at 30 seconds produces more net engagement with the message, even if it catches fewer at-risk sessions.

The research team's own framing leans toward the faster intervention, on the grounds that the 14-minute session reduction is measured after dismissal is accounted for. But they also note that the 24-hour window is short and that they have no data on whether reduced session length in the moment translates into reduced harm over months. That is the question that would actually justify the latency investment, and it is unanswered.

The vendor incentive problem

There is a commercial reason the 2.3-second figure has not become a market standard. Responsible gambling tooling is sold to operators, and operators buy it to satisfy regulators, not to change player behaviour. A vendor that can demonstrate a logged intervention at any latency satisfies the reporting requirement. A vendor that can demonstrate a 2.3-second median latency is selling something the regulator has not asked for, which makes it a harder sale.

The Croatian operators with in-house modelling teams are the exception, and their motivation is usually reputational or licence-related rather than purely compliance-driven. Two of the three operators in the research sample have publicly committed to harm-reduction targets that go beyond the statutory minimum, which is why they were willing to share session-level data with an external team. The third participated under an anonymity agreement and is not named in the published material.

Where the number could go next

The obvious next step is a randomised latency trial within a single operator, and the research team has said it is in discussion with one Croatian licensee about running exactly that. The design would assign sessions to fast (under 3 seconds) or slow (over 20 seconds) intervention at the point of classification, and measure both immediate session behaviour and a longer follow-up window of 30 days. That design would answer the causal question the current dataset cannot.

A second direction is to test whether the 2.3-second figure is optimal at all, or whether it is simply where current models land. The number is a median of what operators do, not a finding about what works. It is entirely possible that 1 second is better, or that 8 seconds is better, and the current data cannot distinguish. What the data does establish is that the fast quartile and the slow quartile produce measurably different session outcomes, which is enough to justify the trial.

For Croatian players, the practical implication is that the popup they see is a function of which operator they are on and what that operator's model happens to flag, not of any consistent standard. Two players exhibiting identical behaviour on two licensed sites can receive interventions 20 seconds apart, or one can receive none at all. That inconsistency is the part of the system least visible from the player side, and it is the part the 2.3-second figure makes measurable for the first time.

The open question is whether the Ministry of Finance's gambling division will treat latency as a reportable metric in the next round of secondary legislation. If it does, the 2.3-second median becomes a benchmark operators are measured against, and the slow quartile has a problem. If it does not, the figure stays a private research finding, and the gap between the fast and slow quartiles persists for as long as players do not know which operator they are on.