Martingales

A martingale is the mathematical distillation of a fair game: a process whose expected future value, given everything you know now, is exactly its current value. That one condition is astonishingly powerful. It formalises "you can't beat a fair game with clever bet timing", it is the language in which arbitrage-free pricing is stated, and it turns hard questions about random walks, where they end up, how long they last, into one-line calculations. Martingales are where the information-theoretic view of Probability Spaces pays off, because the whole definition is about conditioning on what you know.

Definition: conditional expectation is the engine

Fix a filtration $\{\mathcal{F}_n\}$, an increasing family of σ-algebras with $\mathcal{F}_n$ representing all information available at time $n$ (see Probability Spaces). A process $\{M_n\}$ adapted to the filtration (each $M_n$ is $\mathcal{F}_n$-measurable, i.e. knowable at time $n$), with $\mathbb{E}|M_n| < \infty$, is a martingale if $\mathbb{E}[M_{n+1} \mid \mathcal{F}_n] = M_n \quad \text{for all } n.$ The best prediction of tomorrow, given all of today's information, is today's value. Two cousins relax the equality: a submartingale has $\mathbb{E}[M_{n+1}\mid\mathcal{F}_n] \ge M_n$ (a favourable game, drifting up) and a supermartingale has $\le$ (unfavourable, drifting down, the "super" investor's wealth under a house edge sinks). Taking unconditional expectations in the martingale property and iterating gives the immediate corollary $\mathbb{E}[M_n] = \mathbb{E}[M_0]$ for all $n$: a martingale has constant mean. Note a martingale need not be Markov and a Markov Chains chain need not be a martingale, they are different structures that happen to overlap (e.g. a symmetric random walk is both).

Canonical examples

• Symmetric random walk. $S_n = \sum_{i=1}^n X_i$ with $X_i$ i.i.d. mean-zero. Then $\mathbb{E}[S_{n+1}\mid\mathcal{F}_n] = S_n + \mathbb{E}[X_{n+1}] = S_n$. The prototypical martingale.
• Product form. If the $X_i$ have mean 1 and are independent, $M_n = \prod_{i\le n} X_i$ is a martingale, the multiplicative version, and the shape of likelihood ratios and discounted prices.
• Wealth in a fair game. Bet any $\mathcal{F}_n$-measurable stake $B_n$ (your bet can depend on history but not the future). Winnings $W_n = \sum B_k X_k$ form a martingale transform, and it is still a martingale. This is the theorem that no strategy of varying bet sizes turns a fair game into a favourable one; the best you can do in expectation is break even, no matter how clever the system.
• Doob martingale. For any integrable $Y$, $M_n = \mathbb{E}[Y \mid \mathcal{F}_n]$ is a martingale, your evolving best estimate of a final quantity as information arrives. This is the Bayes' Theorem belief-updating process, and it converges (martingale convergence theorem) to $\mathbb{E}[Y\mid \mathcal{F}_\infty]$.

The optional stopping theorem

The crown jewel. A stopping time $\tau$ is a random time decidable from information available as it happens: $\{\tau \le n\} \in \mathcal{F}_n$ (you know when you've stopped without seeing the future, "stop when I hit $100" is valid, "stop the day before the peak" is not). The Optional Stopping Theorem (OST) says that under mild regularity (e.g. $\tau$ bounded, or bounded increments with $\mathbb{E}[\tau] < \infty$), $\mathbb{E}[M_\tau] = \mathbb{E}[M_0].$ Stopping a fair game at a cleverly chosen (but non-anticipating) time still leaves you with your starting expectation. This is the rigorous statement of "you cannot beat a fair game," and it is a computational powerhouse: it collapses questions about the value of a random process at a random time into its value at time zero.

Worked example, gambler's ruin via OST

A symmetric random walk $S_n$ starts at $i$ and stops at $\tau$, the first time it hits $0$ or $N$. Since $S_n$ is a martingale and $\tau$ is a legitimate stopping time with $\mathbb{E}[\tau] < \infty$, OST gives $\mathbb{E}[S_\tau] = S_0 = i$. But $S_\tau \in \{0, N\}$, so with $p = \mathbb{P}(\text{hit } N)$, $\mathbb{E}[S_\tau] = 0\cdot(1-p) + N\cdot p = Np = i \;\Longrightarrow\; p = \frac{i}{N}.$ The ruin probability is $1 - i/N$, derived in two lines, no recursion. For the expected duration, use a second martingale: $M_n = S_n^2 - n$ is a martingale (since $\mathbb{E}[S_{n+1}^2\mid\mathcal{F}_n] = S_n^2 + 1$ for unit-variance steps). OST gives $\mathbb{E}[S_\tau^2 - \tau] = i^2$, and $\mathbb{E}[S_\tau^2] = N^2 p = Ni$, so $\mathbb{E}[\tau] = \mathbb{E}[S_\tau^2] - i^2 = Ni - i^2 = i(N - i).$ Constructing the right martingale ($S_n$ for the probability, $S_n^2 - n$ for the time) turns a hard first-passage problem into arithmetic. This is the archetype of the martingale method.

The pricing connection

Modern derivatives pricing is a martingale statement. The Fundamental Theorem of Asset Pricing says a market is arbitrage-free if and only if there exists an equivalent risk-neutral measure $\mathbb{Q}$ under which every discounted asset price is a martingale: $\frac{S_t}{B_t} = \mathbb{E}^{\mathbb{Q}}\!\left[\frac{S_T}{B_T} \,\Big|\, \mathcal{F}_t\right],$ where $B_t$ is the money-market numeraire. The price of any claim is then its discounted expected payoff under $\mathbb{Q}$. The economic content: no arbitrage ⇔ prices are fair games after adjusting for the risk-free drift. Completeness (unique $\mathbb{Q}$) corresponds to every claim being replicable, and the martingale representation theorem produces the hedging strategy. This is why the whole apparatus above, filtrations, adaptedness, optional stopping, is the native language of quant pricing.

Failure modes and subtleties

• OST needs its regularity conditions. Drop them and it fails spectacularly: the martingale betting system (double your bet after each loss) on a fair coin has $\mathbb{E}[M_\tau] = M_0$ violated in the naive sense because $\tau$ (first win) is unbounded and requires unbounded capital. With finite bankroll you almost surely eventually hit a catastrophic loss, the classic "why doubling doesn't work" lesson. Always check boundedness of $\tau$ or the increments.
• A martingale is not a predictor of profit. It says the game is fair, not winnable; positive expectation requires a sub-martingale (a genuine edge), and even then sizing matters, see The Kelly Criterion.
• Real prices are not martingales under the physical measure. They carry a risk premium (drift), i.e. they are submartingales for a risk-averse investor; the martingale property holds only under the pricing measure $\mathbb{Q}$, not real-world $\mathbb{P}$. Confusing the two measures is a deep and common error.
• Constant mean ≠ constant variance. A martingale's variance typically grows ($\operatorname{Var}(S_n) = n\sigma^2$); "fair in expectation" says nothing about spread or drawdown risk.
• Stopping-time measurability. A rule that peeks at the future ("sell at the top") is not a stopping time, and OST does not apply, precisely why you cannot use it to manufacture profit.

In interviews

Gambler's ruin via martingales is a signature question, deriving $\mathbb{P}(\text{ruin}) = 1 - i/N$ and $\mathbb{E}[\tau] = i(N-i)$ with the two martingales $S_n$ and $S_n^2 - n$ marks you as fluent. Expect "does the doubling (martingale) betting system beat a fair game?", no, and the crisp reason is the OST regularity failure plus finite bankroll. Be ready to define a martingale precisely (conditional expectation, filtration, adaptedness) and to distinguish it from a Markov chain. For pricing desks, articulate the risk-neutral pricing statement: discounted prices are $\mathbb{Q}$-martingales, and no-arbitrage is the existence of such a measure. The through-line to emphasise is that martingales convert statements about random times and endpoints into statements about time zero.