Is That Normal?
Somebody who has just lost eleven hands out of fifteen wants to know one thing, and it is not the house edge. It is whether what just happened to them was unusual. That is a question about spread rather than about average, and it is the one question this site has never answered: the blackjack section quotes expected values to four decimal places across nearly forty pages and, until this one, never once quoted a standard deviation.
Everything below is exact. That is worth saying because the published figures for blackjack variance are not — they come from simulations of billions of hands — so the two agreeing is a real check on both.
What one hand pays
Every possible result of a single hand, and how often it happens, with perfect basic strategy at 6 decks, dealer standing on soft 17, double after split allowed, naturals paying 3:2. Negative numbers are units lost; the −2 and +2 bars are doubles, the ±3 and ±4 slivers are split hands, and +1.5 is a natural.
Result of one hand, in units staked · 6 decks · Dealer Stands on Soft 17, DAS, pays 3:2
The headline of that chart is not the house edge, it is the asymmetry: 47.9% of hands lose money and only 43.4% win any, with 8.7% pushing. You lose more hands than you win and the game is still nearly even, because the hands you win pay more — doubles, splits and the 3:2 on a natural are what close a gap of 4.5 points in the raw count. Anyone who tells you a session felt rigged because they lost more hands than they won is describing the normal state of the game.
How far a session can drift
One hand’s standard deviation is 1.147 units against an expectation of -0.0046 — the swing is roughly 251 times the edge. That ratio is the whole story of the game, and it does not shrink as fast as people expect.
Where a session finishes, in units · the middle 90% lies between the two solid lines
| Hands | Hours | Expected | Std dev | Finish behind | Unluckiest 5% | Luckiest 5% |
|---|---|---|---|---|---|---|
| 25 | 0.4 | -0.11 | 5.7 | 52.8% | -9.5 | +9.5 |
| 50 | 0.8 | -0.23 | 8.1 | 52.5% | -13.5 | +13 |
| 100 | 1.7 | -0.46 | 11.5 | 52.6% | -19.5 | +18.5 |
| 200 | 3.3 | -0.91 | 16.2 | 52.9% | -27.5 | +26 |
| 400 | 6.7 | -1.83 | 22.9 | 53.7% | -39.5 | +36 |
| 800 | 13.3 | -3.65 | 32.4 | 54.8% | -57 | +49.5 |
Hours assume a 6-player table at 60 rounds an hour (Casino Operations Management (Kilby, Fox & Lucas))
Take the 200-hand row — about 3.3 hours at a busy table. You expect to be down 0.91 units, which at $25 a hand is about $23. But one session in twenty finishes worse than -27.5 units — about $687 — and one in twenty finishes better than +26. Being down several hundred dollars after an afternoon is not evidence of anything at all. It is the middle of the distribution doing what it does.
Two hands are not twice one hand
Playing two spots at half the stake looks like it should halve the swing. It does not, and the reason is that both of your hands face the same dealer: when she turns over a 20, both of them lose together. Measured here, the covariance between two hands at one table is 0.487 against a variance of 1.315, so two simultaneous hands carry 2.74× the variance of one, not 2×. Per hand, the standard deviation goes 1.147 → 1.343 → 1.513 across one, two and three hands.
How long until the edge shows up
The number that answers “how long before I can tell?” is the point where cumulative expectation equals one standard deviation. Here that is 63,012 hands — about 1,050 hours at a 6-player table. And the edge that eventually shows up is the house’s, not yours: for a basic strategy player this is not the point at which skill starts paying, it is the point at which losing becomes reliable.
That number moves with the seat, not just the game. The same house edge is charged per hand, so a player heads-up against the dealer at 209 rounds an hour gets through it in roughly 301 hours — 4.0× faster than at a full seven-player table. An empty table is not a better game; it is the same game, charged more often.
Checked against a ten-billion-hand simulation
Published blackjack variance figures are Monte Carlo. These are exact, which makes the comparison worth making in both directions. Three rule sets are checked, not one, because a single comparison cannot tell a modelling error apart from a constant offset.
| Rule set | Our variance | Published | Our covariance | Published |
|---|---|---|---|---|
| Wong's Benchmark Rules, directly | 1.2790 | 1.2950 | 0.4744 | 0.4780 |
| Benchmark plus the published double-after-split delta | 1.3150 | 1.3325 | 0.4873 | 0.4889 |
| Benchmark plus double-after-split and surrender | 1.2865 | 1.2962 | 0.4777 | 0.4764 |
Every one of ours sits just below its published counterpart, and that is the useful part: a residual with a consistent sign has an explanation, where one that wandered either way would mean something was broken. This engine never resplits a pair, and a resplit is another chance to have two bets on the table — so it is missing a source of swing rather than mismeasuring the ones it has. The same check run on the growth from one hand to three lands within 0.009 of the published figures.
What this does not model
Session figures treat hands as independent of one another, and the shoe composition is frozen once the first three cards of a hand are out. Real hands at one table share a shoe and, if other people are playing, a dealer — so genuine session tails are slightly wider than these. The direction of that error is known and its size is small; it is stated here rather than buried because a variance figure that hides its own assumptions is worth less than no figure.
Nothing here is a betting system, and none of it makes a losing game a winning one. Spread is not edge. What it does is tell you whether an afternoon was unusual, and the answer is almost always that it was not. If you want the part where the edge itself moves, that is the card counting page.