Martingales and Optional Stopping: The Fair-Game Machine

A handful of interview questions — gambler's ruin, "can I quit while ahead?", first-passage times — all dissolve under one tool. This article is that tool.

The definition that matters

A process $M_t$ is a martingale if its expected future value, given everything known now, is its current value:

$$E[M_{t+1} \mid \mathcal{F}_t] = M_t$$

A fair game: no information available today predicts drift. Your bankroll in fair coin-flipping is a martingale; in roulette it's a supermartingale (drifts down). Two non-obvious examples worth carrying:

• If $S_n$ is a fair ±1 walk, then $S_n^2 - n$ is a martingale (the walk's square grows exactly like $n$ on average — variance accumulation made literal).
• For a biased walk with up-probability $p \neq \tfrac12$, the process $\left(\frac{1-p}{p}\right)^{S_n}$ is a martingale — the exponential trick that unlocks biased-ruin problems.

Optional stopping: no system beats a fair game

The optional stopping theorem (OST): for a martingale and a "reasonable" stopping rule $\tau$ (bounded, or bounded increments with finite expected time — conditions that hold in every interview setting),

$$E[M_\tau] = M_0.$$

No strategy for deciding when to quit — however clever, however adapted to the history — changes the expected outcome of a fair game. This is the mathematically precise version of "there is no betting system."

The machine in action

Gambler's ruin, one line. Start with \20, opponent \80, fair \$1 bets until someone is broke. Bankroll is a martingale and the stopping time is legal, so$E[\text{final}] = 20. The final bankroll is \100 (probability $p$) or \$0:$100p = 20 \Rightarrow p = 1/5$. No difference equations needed.

Expected duration, two lines. Using the second martingale: $E[S_\tau^2 - \tau] = 0$ where $S$ measures your fortune around... concretely, with absorbing barriers at $0$ and $a$ starting from $x$, $E[\tau] = x(a - x)$. For 20/80: $20 \times 80 = 1600$ expected bets. The quadratic — leaving from the middle takes longest — is the $\sqrt{t}$ law again from the other side.

"Quit while I'm ahead." "I'll flip fair coins and stop the moment I'm +1. I always end positive — free money?" OST says $E[\text{profit}] = 0$... yet you do reach +1 with probability 1. The resolution: the strategy needs an unbounded bankroll and unbounded time — the rare, arbitrarily long excursions deep into the red exactly cancel the certain small win. With any finite bankroll $B$, you sometimes bust, and $P(\text{bust}) \times B$ eats precisely your expected gain. The theorem's "reasonable stopping rule" conditions aren't legal fine print; they're where the economics lives — and explaining that is what separates a real answer from a memorized one.

Why traders should care beyond puzzles

• Prices as martingales. Efficient-market logic says properly discounted prices should be (approximately) martingales under the right measure — if tomorrow's price were predictably above today's, buying today would already have moved it. The risk-neutral measure of options theory is exactly the measure making discounted prices martingales; "pricing = expectation" is martingale theory in production.
• Stopped strategies aren't magic. Stop-losses and take-profits reshape the distribution of P&L (clip a tail here, fatten one there) but cannot manufacture expectation from a zero-edge process. Any backtest whose profitability comes purely from exit rules on a driftless underlying is a bug hunt.
• Fair-value bookkeeping. In market-making games, "my expected P&L from here is my current mark" is the martingale property; deviations from it are either edge or error.

The interview version

The biased follow-up completes the pattern: with win probability $p < \tfrac12$, apply OST to the exponential martingale $r^{S_n}$, $r = \frac{1-p}{p}$, to get $P(\text{reach } a \text{ before } 0 \text{ from } x) = \frac{1 - r^{x}}{1 - r^{a}}$. Sanity check it: at $p = 0.49$ starting with 20 versus 80, ruin is already overwhelming — small edges compound viciously over many bets. Casinos are not lucky; they own a supermartingale and the OST guarantees the rest.