Martingale: The Math of Why Doubling Down Fails

Martingale does not work, and the reason is one line of algebra rather than a matter of bankroll size or discipline. Every bet you place has an expected value of minus the house edge multiplied by its stake. Expectation is additive. Therefore the expected value of any sequence of those bets — in any order, at any sizes, chosen by any rule that looks only at past results — is minus the house edge multiplied by your total turnover. Doubling after losses changes the shape of the outcome distribution. It cannot change the sum.

What follows is the exact arithmetic, because the specific numbers are more persuasive than the argument.

The system, and why it feels like it works

Bet 1 unit. If you lose, bet 2. Lose again, bet 4, then 8, 16, and so on. The first win recovers everything staked in the sequence plus one unit of profit, and you reset to 1.

On an even-money bet the maths of the recovery is exact: after n − 1 losses you have staked 2ⁿ⁻¹ − 1 units, the next bet of 2ⁿ⁻¹ pays 2ⁿ⁻¹ in profit, and the net is +1. It works every single time — right up until it doesn’t.

It feels like a system because it delivers what a system is supposed to deliver: near-constant small wins. Set the base at 1 unit on a 49.5% win-chance game and you complete a sequence in an average of 2.02 rounds, each ending +1. Hours of play look like a slow, steady climb. The distribution has been reshaped so that almost all the probability mass is on small gains and a tiny sliver holds a loss large enough to pay for all of them.

The exact arithmetic of the losing branch

Take the canonical configuration: a bet with a 49.5% win chance paying 2.00×, which is a 2.00× crash cash-out or the 49.50% dice slider position. Loss probability q = 0.505, house edge 1%.

Read the two right-hand columns together. Every extra step of protection doubles what you must be willing to stake and only halves the probability that you need it. The trade is exactly break-even in probability terms and strictly worse in practical ones, because the bet cap and your actual bankroll are fixed while the requirement is not.

At a $1 base, the tenth bet in a sequence is $512 placed to win $1.

Why the expected value is unchanged

The formal argument is short enough to state completely.

Let Xᵢ be your profit on round i, and Sᵢ the stake you place on it. Whatever your rule for choosing Sᵢ — martingale, flat, Fibonacci, coin flip, mood — it depends only on rounds 1 through i−1. Conditional on any such history, the round itself is an independent bet with E[Xᵢ | history] = −e × Sᵢ. Taking expectations:

E[total profit] = E[Σ Xᵢ] = Σ E[Xᵢ] = −e × E[Σ Sᵢ] = −e × E[turnover]

There is no step where the staking rule can intervene. Linearity of expectation holds whether or not the terms are independent, so no sequencing, martingale, stop-loss or “only bet after three reds” rule alters the sum. The only two quantities in the final expression are the edge, which the operator sets, and your turnover, which you control.

That is why the house edge and RTP framing is the whole story: your expected loss is edge × turnover, full stop.

Risk of ruin, computed properly

Take the ten-step version with a $1,023 bankroll — enough to cover exactly ten consecutive losses and nothing more.

Each sequence ends in one of two ways: a win somewhere in the first ten bets (probability 99.8921%, profit +$1) or ten straight losses (probability 0.1079%, loss −$1,023).

• Expected sequences until ruin: 1 / 0.0010787 = 927.0
• Expected rounds per sequence: (1 − q¹⁰) / (1 − q) = 2.018
• Expected rounds until ruin: 927.0 × 2.018 = 1,871

The standard formula for the expected wait until a run of ten losses, (1 − q¹⁰) / (q¹⁰ × (1 − q)), gives 1,870.7 — the same number by a different route, which is a good sign the arithmetic is right.

Now the accounting. Over those 927 sequences you win $1 on 926 of them and lose $1,023 on the last: −$96.99. Total turnover across the run is 927 × $10.46 = $9,699, and 1% of that is $96.99. The two figures agree to the cent, which they must — the ruin event is not a separate disaster, it is simply where the edge you were paying all along gets collected.

What that looks like in sessions

A 200-round evening carries a 10% chance of losing the entire bankroll. Play three such evenings a week for a year and the probability of surviving is effectively zero. “Rare disaster” and “certain eventual disaster” are the same event viewed over different horizons.

Bet caps make it worse, not safer

Suppose the maximum bet is 100× your base. That caps you at seven doublings (1, 2, 4, 8, 16, 32, 64), so the streak that breaks you is seven long, not ten — probability 0.8376%, one sequence in 119. You now expect to be wiped out roughly every 239 rounds, and you lose $127 rather than $1,023.

Lowering the base stake does not escape this, because the cap scales with it. Table limits exist partly because they make the operator’s exposure finite; they also happen to guarantee the martingale player meets the wall.

The variants, and why they end the same way

• Grand martingale. Double and add one unit, so the stakes run 1, 3, 7, 15, 31, 63, 127, 255, 511, 1,023 — $2,036 committed by step ten instead of $1,023. Bigger profit per sequence, proportionally bigger loss when the streak arrives, same expected value.
• Reverse (anti-)martingale. Double after wins instead of losses. Failure mode is much friendlier — you are risking accumulated profit rather than bankroll — but the expected return is still 99% of turnover, because every individual bet is still a 99% bet.
• D’Alembert. Increase the stake by one unit after a loss, decrease by one after a win. Far gentler growth, so the disaster is slower and shallower, and the expected value is unchanged for the same reason as all the others.
• Fibonacci. Step up the Fibonacci sequence on losses, back two on wins. It grows more slowly than doubling and therefore recovers more slowly, and it changes nothing about the sum of negative-expectation bets.

Every one of these is the same object: a rule for choosing Sᵢ from past results. The proof above covers all of them at once, including any variant not yet invented.

What staking systems can legitimately do

They are not useless — they are just not what people think they are. A staking rule genuinely controls:

• Win rate. Martingale wins the vast majority of sequences. That is real, and if you value frequent small confirmations, it delivers them.
• Session length. Flat betting a small base makes a fixed bankroll last far longer in rounds played, which matters if the entertainment is the point.
• Swing size. Reducing stake size shrinks every percentile of the outcome distribution toward its mean.
• Turnover. This is the important one, because turnover is the only variable in −e × turnover you can move.

None of them touches the edge. On a 1% game you can expect to lose about $1 for every $100 you put through it, and the systems only decide how quickly you put it through and how it feels while you do.

The corollary is unglamorous and completely reliable: on a negative-edge game, expected loss equals edge multiplied by turnover, so the only staking decision that improves it is a smaller one. Betting less, less often, on the lowest-edge game available is not a system — it is the only thing the arithmetic permits. And verifying that the edge is what the operator claims, which is what provably fair systems let you check, is worth more than every staking spreadsheet ever written.