Bankroll Management: Staking, Variance and Ruin

Bankroll management is the set of rules that decide how much you bet, and its purpose is survival rather than profit. A bankroll is a fixed sum of money, held separately from everything else, that you have already decided you can afford to lose in full. Every stake is a percentage of that number and nothing else.

It cannot make a losing approach profitable. What it does is stop variance from removing you from the game before the arithmetic has had time to work — which matters whether the arithmetic is in your favor or not.

The bankroll as a fixed, separate, loseable amount

Three properties, all load-bearing.

Fixed. You decide the amount once. Topping it up from your checking account after a bad week is not bankroll management; it is a slow-motion loss you have chosen not to count.

Separate. In its own account, or at minimum tracked in its own ledger. Money that is simultaneously the rent and the bankroll causes bad decisions, because the pressure to get back to even is the pressure to make the rent.

Loseable. The whole thing. Not “unlikely to be lost” — actually gone, with no consequence other than not betting for a while. If that sentence is uncomfortable, the number is too large.

The psychological function is as important as the mathematical one. A fixed bankroll converts an open-ended activity with no natural stopping point into a bounded one with a defined maximum cost.

Unit sizing: 1u = 1–2% of the roll

A unit is your standard bet size, expressed as a percentage of the bankroll. The common convention is 1% to 2%, and it is the convention for good reasons that the variance math below makes concrete.

Two rules make the unit useful rather than decorative:

• Bet one unit on everything by default. “Confidence” is not a measurable quantity for most bettors, and multi-unit plays are where a good month turns into a bad one.
• Recalculate the unit rarely. Monthly, or after a large move in the bankroll. Not after each bet.

If you genuinely want to vary stakes, vary them by measured edge, not by feeling — and if you cannot state your edge as a number, you are varying by feeling. The concept you would need is set out in expected value betting.

Flat staking vs percentage staking

Flat staking means every bet is the same dollar amount, recalculated occasionally. Percentage staking means every bet is the same percentage of the current bankroll, so stakes shrink as you lose and grow as you win.

Percentage staking is theoretically safer — betting 2% of what remains means you can never quite reach zero. In practice it demands recalculating before every bet, and that constant contact with the bankroll number is exactly where discipline breaks down.

Flat staking is the right default. It is simple, it is honest, and it makes your records comparable across a season.

The Kelly criterion, and why not to use all of it

Kelly gives the stake that maximizes the long-run growth rate of a bankroll:

f = (bp − q) / b

Where f is the fraction of the bankroll to bet, b is the profit per unit staked in decimal terms, p is your probability of winning and q is 1 − p.

Worked example. You believe you have a 5% edge at even money — your side wins 52.5% of the time at +100. Then b = 1, p = 0.525, q = 0.475:

f = (1 × 0.525 − 0.475) ÷ 1 = 0.05

Five percent of the bankroll. At even money, Kelly is simply your edge.

At other prices it is not. Here is the same formula at −110, where b = 100 ÷ 110 = 0.9091:

Notice how demanding this is. A 54% win rate at −110 is an excellent long-run result, and Kelly still only sanctions 3.4% of the bankroll.

Why full Kelly is too aggressive

Kelly is optimal on one condition: that p is exactly right. Yours is an estimate, and estimates of your own edge skew high — the errors you notice are the ones that lost.

The consequence is severe, because Kelly’s growth curve is asymmetric. Using the 5%-edge-at-even-money example, here is the long-run growth rate per bet at multiples of the Kelly stake:

Read the 2× row carefully. Betting twice the Kelly stake reduces the long-run growth rate to zero, even though the edge is entirely real. Bet more than that and a genuine edge produces a bankroll that trends to nothing.

Now combine that with estimation error. If you think your edge is 5% and it is actually 2.5%, then betting “full Kelly” on your estimate is betting double Kelly on reality — and you have volunteered for the zero-growth row while believing you were optimizing.

What variance actually looks like

Two computed examples, so the word “variance” stops being an abstraction. Both assume flat $100 stakes at −110, where a win pays $90.91.

A break-even bettor, 100 bets

At exactly 52.38% — the break-even rate at −110 — expected profit over 100 bets is $0.00. The standard deviation is $953.

That means a completely edgeless bettor routinely finishes 100 bets up $900 or down $900. Nine units in either direction, from zero skill. Any conclusion you draw about your ability from a 100-bet sample is being drawn from noise.

A strong winning bettor, 500 bets

Now take a genuine 54% bettor — a rate that would make someone a long-term winner in any market.

• Expected profit per bet: 0.54 × $90.91 − 0.46 × $100 = $3.09
• Over 500 bets: +$1,545, or about 15.5 units
• Standard deviation per bet: 190.91 × √(0.54 × 0.46) = $95.15
• Over 500 bets: 95.15 × √500 = $2,128, or about 21.3 units

A very good bettor has close to a one-in-four chance of losing money over 500 bets. Drawdowns of ten or fifteen units along the way are not a sign that something has broken — they are the normal texture of the distribution. This is why bettors who understand their own numbers track closing line value instead of their balance: it reads the signal months before the balance does.

Risk of ruin, and what overbetting does to it

Risk of ruin is the probability that your bankroll reaches zero before your edge can express itself. It depends on three things: the size of the edge, the size of the bet relative to the bankroll, and how deep the bankroll is in units.

For a 54% bettor at −110 flat-staking one unit and never resizing, the approximate probability of eventual ruin is:

These are computed under deliberately harsh assumptions — flat stakes forever, no reduction after losses, an infinite horizon — so treat them as the shape of the relationship rather than a forecast. The shape is the lesson: ruin risk falls steeply with bankroll depth, which is the entire argument for a 1–2% unit.

And for a bettor with no edge, the number is not 3% or 26%. Given enough bets, ruin is certain. Bankroll management buys time; it does not change the direction of travel.

Overbetting is the fastest route to the bad end of this table. Doubling your unit while keeping the same edge does not double your expected profit relative to risk — it multiplies your variance by four relative to the bankroll, and as the Kelly growth table showed, past a certain point it drives the long-run growth rate negative. A real edge, bet too large, still ends at zero.

Chasing losses

Every staking plan fails the same way: the bettor stops following it after losing.

The instinct is arithmetically coherent and practically fatal. If you are down five units, one five-unit bet gets it back. If that loses, one ten-unit bet gets it back. Each individual step is a valid solution to the problem of being down, and the sequence has a guaranteed terminal state.

Two things make it worse:

• Escalating stake size means your biggest bets are placed with your smallest bankroll. Risk of ruin climbs at exactly the moment you can least afford it.
• Recovery bets reach for variance. Getting a large amount back quickly means long odds, which usually means parlays — a product whose compounding margin makes it the most expensive place to attempt a recovery.

The defense is structural rather than psychological. Decide your unit while you are not betting, and treat any in-session change to it as prohibited. The rule has to be in place before you need it, because the moment you need it is the moment you will not want it.

Rules that survive contact with reality

• One unit, flat, on everything. Recalculate monthly at most.
• Never resize during a session. Not up after wins, not up after losses.
• Record every bet. Date, market, price, stake, closing price, result. The record is the only thing that will ever tell you the truth.
• Keep gambling money separate from life money. Different account, no exceptions.
• Set a deposit ceiling per month and treat it as the real limit. The bankroll is the number you can lose; the deposit ceiling is what stops that number from quietly growing.
• Shop your price before you size your bet. Getting −105 instead of −115 improves your break-even rate by 2.27 percentage points, which is a bigger effect than most staking tweaks.

When to move up or down

Move the unit when the bankroll has moved meaningfully — a common trigger is a 25% change in either direction — and move it by recalculating the same percentage, not by choosing a new number that feels right.

Move down without hesitation. A bankroll that has fallen 30% should be producing smaller bets; that is what percentage-based sizing means, and refusing to shrink is chasing by another name.

Move up slowly. A bankroll that has grown 30% can support a bigger unit, but the growth may have been variance, and 500 bets is not enough to know. There is no cost to raising stakes late and a large cost to raising them early.

The part nobody wants to hear

Bankroll management cannot create an edge. Nothing about how you size a bet changes whether the bet was worth making. A perfectly disciplined bettor with no edge loses at a rate set by the vig, just more slowly and with better records than an undisciplined one.

What it does is preserve the sample. It keeps you solvent long enough for a few hundred bets to become a few thousand, at which point your results start to mean something and you can find out — honestly, with numbers — whether you were ever any good at this. Most people find out that they were not. Knowing that cheaply, over a long time, is a better outcome than finding out expensively in a weekend.