Martingale: Fool’s Gold for Advantage Players and Alpha Hunters

Written By Paradis

Why doubling after losses cannot turn a negative-EV game into a positive one

Martingale systems have survived for centuries because they are very good at producing the result people want to see.

Bet one unit on red. If it loses, bet two. Lose again and bet four, followed by eight, sixteen and however many further powers of two the bankroll and table limit will tolerate. The first win recovers every earlier loss and leaves a profit of one unit.

Most sequences end quickly. The player records another successful cycle, the system appears reliable and the person selling it gains another example for the website. The occasional failure is much larger, although it arrives infrequently enough that the system can build trust before taking the money back.

Martingale usually wins. Then it loses once and takes the previous wins, the bankroll and part of next month with it.

I think this is worth addressing directly because Martingale often appears near the beginning of somebody’s interest in advantage play, sports betting or alpha hunting. It looks systematic, responds to changing results and supplies precise instructions at the point where emotion would otherwise take over. Those are qualities worth having in a serious staking framework.

The problem lies in the information driving the stake. Martingale increases exposure because the previous wager lost. In a game where successive outcomes are independent, that loss tells us nothing favourable about the next result.

The system changes the shape of the outcome distribution. The expected value remains exactly where the casino left it.

The Betting System and the Mathematical Martingale

The word martingale is used for two related ideas that should be kept separate.

The gambling system is a staking progression. The player raises the wager after losses, traditionally by doubling, until a win recovers the accumulated deficit and produces a small profit.

In probability theory, a stochastic process (M_t) is a martingale when its expected future value, conditional on everything currently known, equals its present value:

E[M(t + 1) | information available at time t] = M(t)

A fair gambling process can satisfy this condition because the next step has zero conditional expectation. A casino game with a house edge has negative conditional drift, making the player’s wealth process closer to a supermartingale: expected future wealth is no greater than current wealth.

The staking progression does not repair that drift. It rearranges the outcomes so that the player experiences frequent small profits and rare large losses.

This distribution is psychologically attractive because win percentage is simple and feels like evidence. A system that wins 99% of its cycles sounds far safer than one that wins 55%, provided nobody asks how much is lost during the remaining 1%.

The missing part of the distribution is where the bodies are buried.

A Simple Single-Zero Roulette Example

A single-zero roulette wheel contains 37 numbers:

18 red
18 black
1 green zero

An even-money bet on red wins with probability:

p = 18 / 37

and loses with probability:

q = 19 / 37

The expected profit from betting one unit is:

E[X]
    =
(18 / 37)(+1)
+
(19 / 37)(-1)

    =
-1 / 37

The house edge is therefore approximately 2.70% of every unit wagered.

Suppose the player uses a Martingale sequence containing at most (n) bets:

1, 2, 4, 8, ..., 2^(n - 1)

If any wager wins before the sequence ends, the cycle produces a profit of one unit. If all (n) wagers lose, the total loss is:

1 + 2 + 4 + ... + 2^(n - 1)
    =
2^n - 1

The probability of losing the complete sequence is:

q^n

The expected profit from one cycle is therefore:

E[cycle]
    =
(1 - q^n)(+1)
+
q^n[-(2^n - 1)]

    =
1 - (2q)^n

For single-zero roulette:

2q = 38 / 37

which gives:

E[cycle]
    =
1 - (38 / 37)^n

This quantity is negative for every positive value of (n).

Extending the sequence increases the percentage of winning cycles. It also increases the capital exposed during the losing sequences, leaving the expected result progressively worse.

Maximum bets Capital required Winning-cycle probability Full-sequence loss probability Expected profit per cycle
4 15 units 93.046% 6.954% -0.1126 units
6 63 units 98.166% 1.834% -0.1735 units
8 255 units 99.516% 0.484% -0.2378 units
10 1,023 units 99.872% 0.128% -0.3056 units

The ten-step version wins approximately 99.87% of its cycles. The remaining 0.13% contains a loss of 1,023 units.

A complete ten-loss sequence occurs about once every 780 cycles on average. Across 1,000 independent cycles, the probability of encountering at least one is approximately 72%.

The system can look almost impossible to lose with over a short demonstration. Repeated play eventually gives the tail event enough chances to arrive.

Where the Expected Loss Goes

The same calculation can be viewed through the total amount wagered.

The first one-unit bet occurs during every cycle. The two-unit bet appears only after one loss, with probability (q). The four-unit bet appears after two losses, with probability (q^2), and the pattern continues.

The expected total amount wagered during an (n)-step cycle is:

E[total wagered]
    =
1 + 2q + 4q² + ... + 2^(n - 1)q^(n - 1)

    =
Σ(2q)^k

Every unit placed on red loses (1/37) units in expectation. The expected cycle profit can therefore be written as:

E[cycle profit]
    =
-(1 / 37) × E[total wagered]

The progression creates more turnover during the sequences where the player is losing. Since the casino earns an expected percentage of turnover, this arrangement causes no distress at the roulette table.

The high win rate comes from the way a cycle is defined. Every sequence ending in a win is counted as one successful cycle, regardless of whether the player wagered one unit or hundreds of units to obtain the same one-unit profit. The rare losing cycle absorbs a large number of those wins at once.

A system can look excellent when its success metric counts sessions. Expected value counts money.

Previous Losses Do Not Improve the Next Spin

A roulette wheel has no memory of earlier outcomes. After six consecutive black results, the probability of red on the next spin remains:

18 / 37

The run may look unusual after it has happened, but the next spin receives no instruction to repair the sequence.

The wheel has no obligation to return your money because black appeared six times. It does not know you are upset.

The gambler’s fallacy comes from expecting short sequences to resemble the long-run distribution more closely than probability requires. If red appears with long-run frequency (18/37), people often feel that a recent shortage of red should make it more likely to appear soon.

The law of large numbers says that the observed proportion converges towards the underlying probability over a sufficiently large number of spins. It does not require the process to compensate for an earlier imbalance during the next spin, the next hour or before the player’s bankroll reaches zero.

A sequence containing ten black results followed by one red result remains heavily black. The red result has not restored balance. It appeared with the same probability it always had.

Martingale converts this misunderstanding into a position-sizing rule.

The General Expected-Value Proof

The argument applies to much more than doubling.

Let (B_t) be the amount wagered on spin (t). The player can choose (B_t) using every result observed before that spin, provided the stake is fixed before the wheel produces the next outcome.

Let (X_t) equal (+1) when the even-money bet wins and (-1) when it loses. For single-zero roulette:

E[Xt | all previous information]
    =
-1 / 37

Because (B_t) has already been selected using past information:

E[BtXt | previous information]
    =
Bt E[Xt | previous information]

    =
-Bt / 37

Taking expectations and summing across any finite sequence gives:

E[total profit]
    =
-(1 / 37) E[total amount wagered]

The same reasoning covers Martingale, reverse Martingale, Fibonacci systems, cancellation systems, stop-win rules and any other staking plan whose next wager depends only on earlier roulette results.

The stake can change dramatically. The expected loss attached to each dollar remains unchanged.

A double-zero wheel adds a second green number and raises the house edge on an even-money wager to:

2 / 38
    =
5.263%

The progression behaves the same way while charging a larger expected fee on every dollar wagered.

The Unlimited-Bankroll Defence

Martingale defenders often argue that the system must eventually win if the player has unlimited capital and the casino has no table limit.

That statement is mathematically interesting because an infinite sequence of losses has probability zero when each wager carries a fixed positive chance of winning. Given unlimited doubling, a win eventually arrives with probability one and leaves a profit of one unit.

The required stake has no finite upper bound:

Consecutive losses Next wager Capital already required
5 32 units 31 units
10 1,024 units 1,023 units
20 1,048,576 units 1,048,575 units
30 1,073,741,824 units 1,073,741,823 units

With a $10 starting unit, the wager after 30 consecutive losses exceeds $10 billion.

The theoretical version requires unlimited capital, unlimited credit, no table limit and no cost attached to using the money. None of these conditions survives contact with a casino.

If the system requires unlimited capital and no table limit, it may be worth checking whether either came with the course.

Every version that can actually be played ends after a finite number of bets. Every finite version has negative expected value in a negative-EV game.

The unlimited construction also illustrates why stopping-time results require conditions. Standard optional-stopping conclusions rely on forms of boundedness or integrability that prevent a player from hiding an arbitrarily large exposure inside a rare future event. Infinite Martingale betting escapes those conditions by allowing stakes and temporary losses to grow without limit.

The mathematics has not been defeated. The proposed strategy has become financially impossible.

Table Limits Expose the Problem

Starting with a $10 unit, a ten-step sequence is:

$10
$20
$40
$80
$160
$320
$640
$1,280
$2,560
$5,120

The total capital required is:

$10,230

The profit if the final wager wins is:

$10

A smaller starting unit reduces the required capital while reducing the routine profit by the same proportion. The relationship between the usual reward and the tail loss remains unchanged.

A larger bankroll allows the player to survive more consecutive losses. Continued play then supplies more opportunities for the longer sequence to occur.

System sellers sometimes describe casino table limits as artificial interference with an otherwise successful plan. The limit did not create the exponential growth. It merely stopped the doubling at a number the player could finally see.

Why Martingale Backtests Look Convincing

A Martingale simulation can look excellent when it reports the metrics most favourable to the strategy.

Those commonly include:

  • percentage of winning sessions;
  • average result from successful cycles;
  • number of cycles completed before failure;
  • longest winning run;
  • smoothness of the bankroll before the large drawdown.

The missing outputs are more important:

  • expected profit;
  • maximum capital required;
  • size of the failed sequence;
  • probability of failure across repeated cycles;
  • total turnover;
  • maximum drawdown;
  • results after bankroll and table constraints.

A system that wins 99.8% of sessions sounds safe until the remaining 0.2% is allowed into the report.

This negative skew also produces convincing testimonials. Many users experience a long series of small wins and report that the system works before the tail event arrives. Those who hit the tail early are less likely to remain active in the community, sometimes because the gambling budget has entered administration.

The backtest needs to contain the full distribution. Showing only the frequent result is how the system sells itself.

Blackjack Does Not Rescue the Progression

Blackjack creates more confusion because the game can genuinely become positive expectation under certain conditions.

Cards are dealt without replacement, so the composition of the remaining shoe changes. When enough ten-value cards and aces remain, the player can gain an advantage. Card counting attempts to estimate that conditional edge from the cards already seen and increases the wager when the underlying opportunity improves.

Martingale increases the wager after a loss.

A losing hand supplies very little useful information by itself. The player can lose several hands while the count becomes worse, then arrive at the largest wager when the remaining shoe is least favourable.

A genuine counting strategy follows a chain resembling:

observed cards
    -> estimate of remaining composition
    -> estimate of conditional player edge
    -> stake based on edge and bankroll

Martingale uses:

previous loss
    -> larger stake

The first process responds to information about the next hand. The second responds to damage from the previous one.

Basic strategy can reduce the casino’s edge substantially, but it normally leaves the player with negative expectation. Adding a Martingale progression changes the distribution of the losses and increases turnover during bad sequences. It does not move the expectation above zero.

Any paid blackjack system should therefore answer one direct question: what observable information changes the conditional probability of the next hand?

If the answer is that the previous hand lost, the seller has supplied a staking plan where an edge was required.

Advantage Play Begins with the Edge

A legitimate advantage strategy starts by identifying a reason the next wager has positive expectation.

In blackjack, the information may come from the composition of the remaining shoe.

In sports betting, the market price may understate the true probability of an outcome.

In poker, weaker opponents, superior strategy and game selection can create the edge.

In a physical casino game, an opportunity might arise from a biased device, procedural information or a mechanical feature that changes the distribution. Any such claim requires evidence showing how the information becomes available before the wager.

Once the edge has been established, bankroll management determines how much capital should be exposed to it.

Martingale begins with the staking system and hopes the progression can compensate for the missing edge. This is the gambling version of levering a trading strategy before checking whether the signal predicts anything.

A more complicated loss-recovery sequence does not improve the underlying proposition. It only gives the eventual loss more paperwork.

What Kelly Says About Roulette

The Kelly criterion asks what fraction of bankroll maximises long-run logarithmic growth when the probability and payout are known.

For an even-money wager won with probability (p), the full-Kelly fraction is:

f* = 2p - 1

For red on a single-zero roulette wheel:

p = 18 / 37

which gives:

f*
    =
2(18 / 37) - 1

    =
-1 / 37

The negative result says that the proposed wager has negative growth value. When the player is restricted to non-negative bets, the growth-optimal stake is zero.

Kelly does not recommend increasing the roulette wager after several losses. It rejects the first wager.

Where a positive edge exists, Kelly or fractional Kelly adjusts exposure according to the size of the advantage and the bankroll. A smaller bankroll leads to a smaller dollar stake. A sequence of losses cannot justify a larger percentage wager unless the sequence has also changed the underlying opportunity.

That is close to the opposite of Martingale.

Random Walks and Stopping Rules

Martingale systems are often explained using random-walk language. The bankroll moves up and down, the player chooses a profit target and play stops when one boundary is reached.

Stopping rules can change the probability of reaching one boundary before another, the expected duration of play and the shape of the terminal outcomes. These are legitimate questions in stochastic-process analysis.

A stopping rule does not create positive drift in a process that has negative drift.

Suppose the player has a negative expected return on every wager, conditional on all information available before the wager is made. Under ordinary finite-capital and bounded-stake conditions, changing the time at which the player stops cannot convert those conditional losses into positive expectation.

A fair random walk produces a related result. Under the usual optional-stopping conditions, choosing a clever stopping time does not create expected profit from a process with zero conditional drift.

Interesting random-walk problems begin when something in the process changes the transition probabilities, payouts or available information. Mean reversion, state-dependent dynamics, changing boundaries and structural constraints can create genuine questions worth investigating.

The fact that the previous step went down supplies no such structure.

A Quick Test for Any Gambling System

Anyone starting in advantage play or alpha hunting can save a great deal of time by asking a proposed system a few direct questions.

Question What a serious answer should contain
Where does the edge come from? Information, pricing error or structure that changes conditional expectation
What is the expected value per dollar wagered? A calculation covering every outcome, including rare losses
Why does the stake change? A reason connected to future expectation or bankroll risk
How much capital is required? Full-sequence exposure, drawdowns, liquidity and table limits
How was it validated? Complete distributions, controls and testing beyond one favourable period
What happens with repeated use? The probability and size of eventual failure across many cycles

A system reporting only win percentage has selected the metric it was designed to flatter.

A system claiming that the next result is due has confused long-run frequency with a debt owed by the random process.

A system requiring unlimited capital should begin by locating the unlimited capital.

When Progressions Can Be Useful

Changing stakes over time can be entirely rational inside a positive-EV strategy.

Kelly betting, fractional Kelly and game-selection thresholds all adjust exposure. A sports bettor may risk more when the estimated edge is larger. A card counter may raise the wager when the remaining shoe becomes favourable. A poker player may move between limits as the bankroll changes.

The relevant variables include the edge, bankroll, variance, correlation and uncertainty around the model.

The previous loss matters because it reduced the bankroll. That usually argues for less dollar exposure. A loss supports a larger future wager only when it also supplied information that improved the expected value of the next opportunity.

This distinction separates risk management from chasing.

Final Thought

Martingale survives because it produces frequent wins, simple instructions and the comforting belief that every loss remains temporary.

The arithmetic remains unchanged. Every roulette wager carries the house edge. Doubling after losses increases the total amount wagered during the sequences where the player is already behind. The standard profit remains one unit while the possible loss grows exponentially.

Blackjack, poker, sports betting and financial markets can all contain real edges. Those edges come from information, pricing errors, opponent mistakes or structure in the process generating the outcomes. Position sizing determines how much of that advantage reaches the bankroll and how much risk is taken along the way.

A staking system without an edge is an organised way of losing.

For anyone starting in advantage play or searching for alpha, the order matters. Find the favourable process first. Work out how much to risk afterwards.

References

  • Bill Chen and Jerrod Ankenman, The Mathematics of Poker. Relevant to expected value, variance, risk of ruin, the Kelly criterion and rational game selection.
  • Dimitri P. Bertsekas and John N. Tsitsiklis, Introduction to Probability. Relevant to conditional expectation, independence, sequential probabilistic models and stochastic processes.
  • Larry Wasserman, All of Statistics. Relevant to random walks, Markov chains, conditional expectation and stochastic-process modelling.
  • David Williams, Probability with Martingales. Relevant to mathematical martingales, stopping times and optional stopping.
  • Lester E. Dubins and Leonard J. Savage, How to Gamble If You Must: Inequalities for Stochastic Processes. Foundational treatment of gambling strategies and stochastic-process inequalities.
  • William Feller, An Introduction to Probability Theory and Its Applications, Volume I. Relevant to random walks, gambler’s ruin and repeated independent trials.
  • John L. Kelly Jr., A New Interpretation of Information Rate. Bell System Technical Journal, 1956.
  • Edward O. Thorp, Beat the Dealer. Relevant to blackjack advantage methods based on changing card composition.
Paradis
Next

Aggressive Poker Bankroll Management: How Little a Proven Winner Can Use Under Kelly