Variance, Drawdown and Blackjack Bankroll Planning

Variance describes the range around an average

Blackjack results arrive in individual hands, while expected value describes an average over a very large collection of comparable hands. A negative house edge can coexist with a winning session, and a favourable model can coexist with a losing session. Neither result changes the underlying expectation by itself.

Variance measures how widely individual and short-run outcomes spread around that average. A blackjack hand can settle for one unit, lose one unit, pay a natural at a different rate, or include an extra unit through a double or split. Those outcomes make session results much more variable than the single-unit win-or-loss picture suggests.

For one six-deck example with dealer standing on soft 17, double after split, late surrender and resplit aces, Wizard of Odds reports a per-hand variance of 1.303 betting units under basic strategy. That is an example, not a universal blackjack constant. The deck count, dealer rule, available actions, payout and number of simultaneous hands change both expected value and volatility.

Standard deviation is the more readable companion to variance. If a model treats hands as comparable, the expected spread grows roughly with the square root of the number of hands, while the expected loss from a fixed house edge grows with the number of hands. Real table play can depart from this simplification because bet sizes, decisions, multiple spots and changing conditions alter the distribution.

Open the House Edge workspace at /en/house-edge to record the posted rules before comparing a bankroll path. A number attached to a different ruleset is an estimate for a different game.

Expected loss does not predict the next session

Expected loss is a rate over repeated exposure. It gives a useful centre point for a model, yet it does not supply a schedule of wins and losses. Two sessions with the same number of hands and the same average wager can finish far apart because their sequences of doubles, splits, naturals and dealer outcomes differ.

This distinction matters when reading a projection. A projected loss of 20 virtual credits over a long sample does not mean that a 20-credit decline will appear smoothly. A short session can end above its starting point, below it by much more than 20 credits, or close to the projection. The distribution is part of the result.

The Long-Run Simulations workspace at /en/simulations can show repeatable virtual paths from one stated configuration. Keep the configuration visible: starting credits, ruleset, wager policy, number of hands, seed and any strategy assumption. A result without those inputs cannot be compared honestly with another run.

Drawdown is a path measure

Drawdown measures a decline from the highest bankroll value reached so far. A session can finish near its starting level after a deep interim decline, so ending balance and maximum drawdown answer different questions.

Consider a virtual path that starts at 1,000 credits, rises to 1,160, falls to 920 and closes at 1,020. The ending balance is positive relative to the start, while the maximum drawdown from the prior peak is 240 credits. The path had to absorb that 240-credit drop before it recovered.

Drawdown depends on the order of outcomes. It is useful for checking whether a proposed wager size would leave enough room for routine adverse stretches in a chosen time horizon. It cannot establish that a particular loss is due, deserved or likely to reverse. A prior low point does not alter the probability model for the next independently dealt hand.

Risk of ruin needs a complete model

Risk of ruin is the probability that a bankroll reaches a stopping threshold under stated assumptions. In its classical form, the calculation depends on the starting wealth and the full distribution of outcomes, not only on an average return. General ruin formulas also require a precise stopping rule and payoff distribution.

For blackjack, a useful model needs more than a starting balance. It needs the ruleset, wager schedule, number of hands or session endpoint, player policy, treatment of doubles and splits, and the definition of ruin. A zero-credit threshold, a table-minimum threshold and a voluntary stop-loss are different events.

Avoid importing a risk-of-ruin percentage from an unrelated ruleset or bet spread. A result built for a six-deck, double-after-split game does not describe a different table merely because both games are called blackjack. Treat a calculated percentage as a conditional model output, with its assumptions displayed beside it.

Build a fixed educational bankroll plan

Start with a virtual amount that is separate from essential spending and has no cash value. Set a session limit before opening the Simulator at /en/simulator, then choose a wager size that makes the limit meaningful in the number of hands you plan to model. Smaller virtual units create more observable decisions before the limit is reached; they do not remove volatility or alter a negative expectation.

Record these inputs in the Bankroll workspace at /en/bankroll:

1. Starting virtual credits and a stop-loss threshold.
2. The exact table rules and variant.
3. A fixed or explicitly defined wager policy.
4. A finite session length or number of hands.

Run several seeded simulations, then inspect the ending distribution and maximum drawdown. If the lower paths cross the planned stop-loss frequently, reduce the virtual wager or shorten the session and run the model again. Do not use a larger wager to recover a prior loss. That choice increases exposure while leaving the random process intact.