Why two identical games feel different
Take two slots, both certified at 96% RTP, both spinning at the same speed, both costing the same per spin. One produces a steady dribble of small wins. The other produces almost nothing for two hundred spins and then pays 400 times stake.
Over an infinite number of spins they are identical. Over the hundred spins a real person plays, they are different products with different audiences, different bankroll requirements and different regulatory profiles.
The number that distinguishes them is volatility, and it is the second most important figure in gambling mathematics after expected loss per hour.
Defining it properly
Volatility is the standard deviation of the return per unit staked.
For a single round, compute the expected return, then compute the expected squared deviation from that expectation across all outcomes, and take the square root. The result is in units of stake.
Slot standard deviations are several times those of table games. In Harrigan and Dixon's 2009 analysis of Ontario design documents (PAR sheets), the two mechanical reel slots carry volatility indices of 10.5 to 17.4, which at the 90% z-score the paper uses correspond to standard deviations of roughly 6 to 10.5 per unit staked, and the Wizard of Odds quotes 8.74 for one slot machine. Titles with very large maximum wins go considerably higher, because a rare top prize adds heavily to the variance: as a worked illustration, a 50,000x prize that supplies one percentage point of the RTP adds about 500 to the variance on its own, which by itself is a standard deviation above 22. Double-zero roulette, for comparison, has a standard deviation of 0.998614 on any even-money bet, close to 1, which is why table games feel so much steadier than slots even before speed is considered.
Because the raw figure is unintuitive, it is usually published in transformed form. On the Ontario PAR sheets the volatility index is the standard deviation multiplied by the z-score for a chosen confidence interval, 1.65 for 90%, so that it can be used to put bounds on the payback percentage. Online studios more often publish a band or star rating: in one UK study the supplier's measure was the standard deviation of the return on a single bet, simulated across typically a billion bets and expressed on a scale of 1 to 8. These scales are not standardised between studios, so a four-star game from one supplier is not comparable with a four-star game from another. If volatility matters to a decision, ask for the standard deviation.
The arithmetic of dispersion
Standard deviation for a single round is not directly useful. What matters is the spread over a session, and that follows a rule worth memorising.
Over n independent rounds, the expected loss scales with n, and the standard deviation scales with the square root of n.
Take a game with 4% house edge and a standard deviation of 10 per unit staked, at 1 unit per spin, over 1,000 spins.
- Expected loss = 1,000 x 0.04 = 40 units
- Standard deviation over the session = 10 x square root of 1,000 = 10 x 31.6 = 316 units
The expected loss is 40 and the typical deviation from it is 316, nearly eight times larger. That is why a thousand spins tells a player essentially nothing about a game's RTP, and why a session that ends 200 units up is entirely unremarkable.
Now run the same game for 1,000,000 spins.
- Expected loss = 40,000 units
- Standard deviation = 10 x 1,000 = 10,000 units
The expected loss is now four times the standard deviation. The result has converged enough that the theoretical edge dominates. This is the long run, and the arithmetic shows exactly how long it takes to arrive: the ratio of expectation to deviation improves only with the square root of the number of rounds, so getting twice as close to theory requires four times as many spins.
A published PAR sheet shows the same thing on a real game. For a 92.5% version of Double Diamond Deluxe, with a standard deviation of 6.35 per unit staked, the 90% confidence interval for the payback percentage runs from 59.45% to 125.71% after 1,000 plays, and narrows to 91.54% to 93.63% only after 1,000,000 plays.
What this means for the operator
The same arithmetic runs on the operator's side, which is the part most often missed.
An operator running a single high-volatility title on a small number of positions faces exactly the dispersion above, in reverse. A month of data on such a game is a sample far too small to estimate its actual performance, and the observed hold can sit well above or below theory for extended periods.
Three practical consequences follow.
Do not reallocate on short samples. Removing a high-volatility game after a month of poor hold is frequently removing a game that was performing to specification and got unlucky. Compute the expected dispersion before concluding anything.
Aggregate before judging. Total wagered across all positions and all sessions converges much faster than any single position. Portfolio-level hold is meaningful long before individual-title hold is. Regulators build the same arithmetic into fault detection: Britain's Gambling Commission expects licensees to compare live RTP with advertised RTP and says volatility is a key parameter when setting the allowable tolerance for each game, while warning that monitoring must not be so aggregated that it hides errors at a lower level.
Understand the liability tail. A game with a 50,000x maximum win creates a single-event liability that can exceed a small operator's monthly margin on that title. This is why maximum win caps exist commercially as well as mathematically, and why some operators restrict stake levels on the highest-volatility titles.
What it means for the player
For the player the consequence is bankroll. A given bankroll buys a very different amount of play depending on volatility.
Continuing the illustrative figures: a player with 500 units on a low-volatility 96% game with a standard deviation of 4 will very likely play for a long session and lose something close to the expected amount. The same 500 units on a game with a standard deviation of 18 will frequently be gone in a fraction of the time, and will occasionally turn into several thousand.
This is why volatility preference is real and strong. Players self-select, and a portfolio that is all one profile will alienate half the audience.
It also has a responsible gambling dimension that is worth stating plainly. High volatility widens the gap between what a player experiences and what the theoretical return describes. A player told the game returns 96% and who has lost twelve consecutive sessions has not been misled by the number, but the number has not described their experience either. Disclosing volatility alongside RTP is a meaningful improvement in that respect. Britain does not require it: its technical standards require the RTP, the house edge or the probability of winning events to be available before the customer commits to gamble, but say nothing about volatility, so a volatility rating, where one is shown, is the studio's own measure. Research on UK online slots players has also found the link between volatility and behaviour to be complex, with players at two operators responding differently to changes in volatility.
Measuring volatility in practice
For a base game that can be enumerated, the standard deviation is computed exactly alongside the RTP: the same enumeration that produces the mean produces the second moment.
For games with features that cannot be enumerated, volatility comes from simulation. Run the game many million times, record the return per spin, and compute the standard deviation of the distribution. Test houses rely on the same relationship: under the Gambling Commission's testing strategy, simulation testing plays a game automatically for a number of games that depends on the volatility of the game as per the game maths, to check that the actual RTP is within an acceptable range of the expected RTP.
Simulation introduces its own statistical question, and the discipline that separates a rigorous model from a sloppy one is stating the confidence interval. A simulation of ten million spins on a very high-volatility game with a rare, enormous top prize may not have hit that prize enough times to estimate its contribution reliably. A common mitigation is to enumerate the rare high-value outcomes analytically and simulate only the rest, and a maths sheet that does not say which parts were simulated and to what precision is not finished.
Volatility as a design decision
Volatility is not a byproduct. It is chosen, and the tools for choosing it are well understood.
To raise volatility: concentrate more of the RTP in rarer, larger outcomes. Increase the top payouts and reduce their frequency. Move return from base game into the feature. Add multipliers, particularly uncapped or slowly capped ones, since a multiplier lengthens the right tail without touching frequency. Raise the maximum win.
To lower volatility: spread return across more frequent, smaller wins. Increase hit frequency above stake. Shift return from feature to base game. Cap multipliers tightly.
Much of the commercial pressure of the last decade has pushed towards higher volatility: streaming culture, social sharing of large wins and bonus-buy mechanics that sell direct access to the high-variance part of the game. In January 2020 Britain's Gambling Commission found six operators offering feature buy-in on slots, one game charging more than £3,000 to enter the bonus feature, and all six removed it. That pressure is commercial rather than mathematical, and the mathematical consequence is that the more play moves to high-volatility titles, the wider the gap between the advertised RTP and any individual player's experience.