Variance in gambling: why short sessions lie and long ones settle

Expected value is the centre of the dartboard. Variance is how far the darts land from that centre. In gambling the darts are session results. A 93.33% RTP colour has a centre at −6.67% of turnover. A two-hour sitting can land at +40% or −80% of turnover and still be a legal dart.

Statistically, variance is the expected squared deviation from the mean. Players meet it as volatility, swing, downswing, variance. The useful facts are the same: short n is noisy; high-payout rares are noisier; raising the stake scales the dollar noise.

This sits in the Games and odds topic. Price the centre with expected value and house edge first, then use this page for the width. PVPspinArena is 18+.

A useful translation: if EV is the slope of a long road, variance is the potholes. Driving further does not fill the potholes; it applies the slope to more kilometres. That is why “I’ll play until it evens out” is how a minus-EV night becomes a minus-EV week.

Law of large numbers needs a large number. A session of 20–80 bets is a small number for any game that is not almost deterministic.

Mean grows faster than noise, but only later

For independent bets, expected total loss scales with n (n times EV per bet). The standard deviation of the total scales with √n. The ratio |mean| / sd therefore grows like √n. That is why casinos are calm about millions of spins and you are not calm about forty.

• At n = 25, √n = 5. Noise is large relative to a 6.67% edge.
• At n = 100, √n = 10. Still plenty of room to finish a session up on a minus-EV colour.
• At n = 10,000, √n = 100. The edge has had time to show in the books. Almost no hobby session lives here.

“I’m up, so the model is wrong”

Being up after 40 Purple bets is compatible with −6.67% EV. The model predicted a distribution, not a single point. Treating a lucky sample as a new p is how people double the next stake. RTP explained covers the same idea from the payback-percentage side: 96% is not a 96% win chance, and it is not tonight’s result either.

Downswing length, not just session P/L

Variance also shows up as the longest losing run inside a session. On a 46.67% hit, the chance of 6 losses in a row is (8/15)^6 ≈ 2.3%, about 1 in 43 isolated starts. In 80 bets you have many overlapping windows. Seeing a six-loss clump is normal weather. Seeing it after you raised the unit is how a “small session” becomes a large hole. Track the clump size you can survive, not only the closing balance.

Two knobs dominate the feel of a session when RTP is held fixed.

Bet size

Stake $1 or $10 on the same colour and the percentage edge is identical. Dollar variance scales with the stake (and with stake squared for variance itself). A $10 unit makes a five-loss run look like a crisis; a $1 unit makes it look like a blip. If the session budget is $40, $10 units can end the night in four misses. That is not bad luck in a mysterious sense. It is a budget with no room for the left tail.

Hit rate

Purple wins about 46.67% of the time and pays 2x. Green wins about 6.67% of the time and pays 14x. Same EV, different variance. Green sessions are long flatlines broken by a spike. Purple sessions tick up and down every other spin. Slots with rare bonuses are Green-like. Even-money table bets are Purple-like.

You can watch both colours on Roulette in the same minute. The wheel does not make Green “due” after a quiet stretch. It only makes Green rare.

People mash these together because both make you lose. They are not the same lever.

• House edge is the expected leak per dollar wagered. It is in the paytable. Systems cannot remove it.
• Variance is how far a finite path wanders around that leak. Stake size, hit rate and n set it.

A low-edge, high-variance game (rare huge jackpot, small published edge) can bust a small bankroll faster than a higher-edge, low-variance even-money bet, because you never survive to the mean. A high-edge, low-variance game grinds you in a straightish line. Neither is a strategy. Both are reasons to pick a stake that matches the budget’s stomach.

PvP Coinflip with a 0% fee has ~0 edge and still has full-stake variance: one flip is ±100% of the stake. Edge and variance parted company there. You can be in a fair game and have a brutal hour.

Suppose 50 independent $2 Purple bets. Turnover = $100. Expected result ≈ −$6.67.

Model each bet as +$2 profit with p = 7/15 and −$2 with q = 8/15 (you lose the $2 stake). Per-bet EV ≈ −$0.267. Per-bet variance is large relative to that mean: you either win $2 or lose $2, roughly a coin with a thumb on one side.

A rough sd for the 50-bet total is on the order of $14. That is not a precise lab figure; it is enough to see the shape. Outcomes between about −$35 and +$20 are unsurprising. An outcome of +$40 is a tail, not a new RTP. An outcome of −$100 (a wipe) is also a tail, and it is how sessions die when the unit is too large for the remaining balance.

Jackpot share

A $5 ticket in a $100 Jackpot pot (0% fee) has p = 5% at $100, EV ≈ $0, and a huge variance: usually −$5, rarely +$95. Ten such tickets in a night are not “due” a hit. The distribution is still 10 independent long shots (or fewer if you re-enter the same style of pot). Short sessions of jackpot tickets lie even harder than Purple, because hits are scarce.

Chance of zero hits in ten independent 5% tickets: 0.95^10 ≈ 60%. Most ten-ticket nights are a clean −$50. That is not a broken generator. That is a 5% shot refusing to appear on cue. People then write “variance” when they mean “I bought a rare event ten times”. Both descriptions are true; only the second one sizes the next ticket honestly.

Independent rounds have no memory. A verified fair generator does not get “hot” in a causal sense. It produces a clump of wins that was always allowed by the binomial (or multinomial) tail.

The gambler’s fallacy is the belief that a clump of losses makes a win more likely. The hot-hand error is the belief that a clump of wins makes another win more likely. Both treat variance as a signal about p. Both are wrong on a memoryless wheel.

What you can check

You can check that the last streak was honest: open fairness and recompute the seeds. That confirms the darts were real. It does not raise p on the next dart.

What you should not do

Do not raise the unit because the session is “running”. That converts a lucky sample into extra turnover at the same negative EV. Do not raise the unit because the session is “cold” and must bounce. That is chase. If the ride is no longer entertainment, stop. Variance is not a debt the table owes you.

You cannot delete variance. You can choose a unit and a stop so the left tail does not take money you need.

• Small unit versus session budget. If you cannot absorb ten losses in a row, the unit is too big for even-money colours. For Green, think in dozens of misses.
• Time and money stops. Fast rounds manufacture n. n manufactures both the mean leak and the chance of a loud swing.
• Do not resize from the streak. Resize from the remaining budget, downward if needed, never upward to recover.
• Match the game to the stomach. Want a longer, tickier session? Purple or a fee-free flip. Want a lottery feel? Green or a small jackpot share. Same or similar EV per dollar is possible; the ride is not.

Variance in gambling is why one friend cashed out smiling and another swore the site was broken, on the same edge. Long books settle. Short nights argue. Believe the paytable, size the unit, and treat a streak as weather.

If you want a single pre-session sentence: “My expected cost is e times turnover; my plausible range is a few session sds around that.” Write both numbers. If the left edge of the range would hurt rent, shrink the unit until it would not. That is variance management. It is not an edge. Short sessions will still lie. You just stop letting the lie set the next stake. Write the range first.