Martingales (Discrete Time)

A martingale is the precise mathematical statement of a fair game. Given a filtration (\mathcal{F}_n), an adapted, integrable process (M_n) is a martingale when the best forecast of tomorrow's value, using everything known today, is exactly today's value:

\mathbb{E}[M_{n+1} \mid \mathcal{F}_n] = M_n.

Read it as a conditional expectation: no information available at time n lets you predict an up-move or a down-move on average. A submartingale relaxes this to \mathbb{E}[M_{n+1} \mid \mathcal{F}_n] \ge M_n (the game drifts in your favour), and a supermartingale to \le M_n (it drifts against you).

The canonical example

Take the symmetric random walk: independent fair steps \xi_i = \pm 1 (each sign with probability \tfrac12, so \mathbb{E}[\xi_i] = \tfrac12(+1) + \tfrac12(-1) = 0), and partial sums S_n = \xi_1 + \xi_2 + \cdots + \xi_n on the natural filtration \mathcal{F}_n = \sigma(\xi_1, \dots, \xi_n). We verify the martingale property \mathbb{E}[S_{n+1} \mid \mathcal{F}_n] = S_n line by line, naming the rule that licenses each step.

Split tomorrow into today plus the new step. By definition S_{n+1} = S_n + \xi_{n+1}, so

\mathbb{E}[S_{n+1} \mid \mathcal{F}_n] = \mathbb{E}[\,S_n + \xi_{n+1} \mid \mathcal{F}_n\,].

Linearity of conditional expectation separates the two pieces:

\mathbb{E}[\,S_n + \xi_{n+1} \mid \mathcal{F}_n\,] = \mathbb{E}[S_n \mid \mathcal{F}_n] + \mathbb{E}[\xi_{n+1} \mid \mathcal{F}_n].

First term — S_n is already known. Since S_n is \mathcal{F}_n-measurable (it is built from \xi_1, \dots, \xi_n), conditioning on \mathcal{F}_n leaves it untouched — "taking out what is known":

\mathbb{E}[S_n \mid \mathcal{F}_n] = S_n.

Second term — the new step is independent of the past. Because \xi_{n+1} is independent of \mathcal{F}_n = \sigma(\xi_1, \dots, \xi_n), the conditioning information is useless and the conditional expectation collapses to the ordinary one:

\mathbb{E}[\xi_{n+1} \mid \mathcal{F}_n] = \mathbb{E}[\xi_{n+1}] = 0.

Recombine. Adding the two results,

\mathbb{E}[S_{n+1} \mid \mathcal{F}_n] = S_n + 0 = S_n,

which is exactly the martingale property. More generally the same three moves — split off the new increment, keep the known part, and use mean-zero independence on the rest — show that any running sum of independent mean-zero increments is a martingale.

Refresh the figure to draw a fresh path. It wanders, but it has no preferred direction: it is just as likely to drift above the dashed mean level M_0 = 0 as below it.

Constant mean, and why finance cares

The tower property makes the mean stand perfectly still. The cleanest statement is that \mathbb{E}[M_n] = \mathbb{E}[M_0] for every n, and we prove it by induction.

One-step result. Start from the defining equation \mathbb{E}[M_{n+1} \mid \mathcal{F}_n] = M_n and take an ordinary (unconditional) expectation of both sides. On the left, the tower property \mathbb{E}[\mathbb{E}[\,\cdot \mid \mathcal{F}_n]] = \mathbb{E}[\,\cdot\,] applies:

\mathbb{E}[M_{n+1}] = \mathbb{E}\!\big[\,\mathbb{E}[M_{n+1} \mid \mathcal{F}_n]\,\big].

Now substitute the martingale property inside the outer expectation, \mathbb{E}[M_{n+1} \mid \mathcal{F}_n] = M_n:

\mathbb{E}\!\big[\,\mathbb{E}[M_{n+1} \mid \mathcal{F}_n]\,\big] = \mathbb{E}[M_n].

Chaining the last two displays gives the single-step conclusion

\mathbb{E}[M_{n+1}] = \mathbb{E}[M_n] \qquad \text{for every } n \ge 0.

Induction. This says consecutive means are equal. Take the base case \mathbb{E}[M_0] = \mathbb{E}[M_0] (trivially true), and assume as the inductive hypothesis that \mathbb{E}[M_n] = \mathbb{E}[M_0]. The one-step result just proved then advances the hypothesis by one index:

\mathbb{E}[M_{n+1}] = \mathbb{E}[M_n] = \mathbb{E}[M_0].

By induction the equality holds for all n, which we may display as the telescoping chain

\mathbb{E}[M_n] = \mathbb{E}[M_{n-1}] = \cdots = \mathbb{E}[M_1] = \mathbb{E}[M_0].

A fair game is fair forever: your expected fortune at any future time equals your starting stake. This is exactly why martingales sit at the heart of pricing. Under the risk-neutral measure, a properly discounted, arbitrage-free asset price is a martingale — its expected discounted value tomorrow is its price today — and that single condition is what makes option prices well defined.

Let (\mathcal{F}_n) be a filtration. An adapted, integrable process (M_n) is a martingale when

\mathbb{E}[M_{n+1} \mid \mathcal{F}_n] = M_n \qquad \text{for every } n \ge 0,

(a submartingale if \ge M_n, a supermartingale if \le M_n). For a martingale the unconditional mean is constant in time,

\mathbb{E}[M_n] = \mathbb{E}[M_0] \qquad \text{for all } n,

obtained by taking expectations of the defining equation (tower property) to get \mathbb{E}[M_{n+1}] = \mathbb{E}[M_n], then inducting on n.

The same random walk hides a second martingale. With S_n = \xi_1 + \cdots + \xi_n and \xi_i = \pm 1 fair and independent, note first that \xi_{n+1}^2 = 1 always (a sign squared is 1), so \mathbb{E}[\xi_{n+1}^2] = 1. Expand the square of tomorrow's position:

S_{n+1}^2 = (S_n + \xi_{n+1})^2 = S_n^2 + 2 S_n \xi_{n+1} + \xi_{n+1}^2.

Condition on \mathcal{F}_n and take the three terms in turn. S_n^2 is \mathcal{F}_n-measurable, so it stays; in the cross term S_n comes out as known and \mathbb{E}[\xi_{n+1} \mid \mathcal{F}_n] = \mathbb{E}[\xi_{n+1}] = 0 kills it; and \mathbb{E}[\xi_{n+1}^2 \mid \mathcal{F}_n] = \mathbb{E}[\xi_{n+1}^2] = 1:

\mathbb{E}[S_{n+1}^2 \mid \mathcal{F}_n] = S_n^2 + 2 S_n \cdot 0 + 1 = S_n^2 + 1.

Now subtract (n+1) from both sides so the extra +1 is absorbed by the bookkeeping:

\mathbb{E}[\,S_{n+1}^2 - (n+1) \mid \mathcal{F}_n\,] = S_n^2 + 1 - (n+1) = S_n^2 - n.

So M_n = S_n^2 - n is a martingale. Its constant mean \mathbb{E}[S_n^2 - n] = \mathbb{E}[S_0^2 - 0] = 0 immediately gives \mathbb{E}[S_n^2] = n — the variance of the walk grows linearly in time, the discrete seed of the \sqrt{t} scaling of Brownian motion.

Can a clever betting rule turn a fair game into a winning one? Let (M_n) be a martingale and let (H_n) be a predictable strategy — H_n, your stake on step n, is fixed using only \mathcal{F}_{n-1}, the information available before the step. Your accumulated winnings are the martingale transform

(H \cdot M)_n = \sum_{k=1}^{n} H_k\,(M_k - M_{k-1}).

Look at the conditional expectation of the next gain. The increment is H_{n+1}(M_{n+1} - M_n); since H_{n+1} is \mathcal{F}_n-measurable it comes out as known, and the martingale property makes the bracket's conditional mean vanish:

\mathbb{E}\big[H_{n+1}(M_{n+1} - M_n) \mid \mathcal{F}_n\big] = H_{n+1}\big(\mathbb{E}[M_{n+1} \mid \mathcal{F}_n] - M_n\big) = H_{n+1}\,(M_n - M_n) = 0.

Hence \mathbb{E}[(H \cdot M)_{n+1} \mid \mathcal{F}_n] = (H \cdot M)_n: the transformed process is itself a martingale, so \mathbb{E}[(H \cdot M)_n] = 0 for every strategy. No predictable rule can give you a positive expected profit — the precise sense in which "you can't beat a fair game". This is the combinatorial heart of arbitrage-free pricing: if discounted prices are a martingale under the risk-neutral measure, then every self-financing trading strategy has zero expected discounted gain, so no strategy manufactures a riskless profit.

It is tempting to picture a martingale as a process that stays constant, or one that drifts back toward its mean. Both pictures are wrong. The defining property constrains only the expected future value given the past:

\mathbb{E}[M_{n+1} \mid \mathcal{F}_n] = M_n,

not any single realisation. An individual path may fluctuate violently — the symmetric random walk above is a martingale yet is unbounded, wandering arbitrarily far from 0 with probability one. What is pinned down is the lack of predictable drift: no information available today lets you forecast an up-move or a down-move on average. A martingale is a fair game, not a still one.

Two further traps. First, don't confuse the three siblings: a martingale has zero expected drift, a submartingale (\mathbb{E}[M_{n+1} \mid \mathcal{F}_n] \ge M_n) tends to rise, and a supermartingale (\le M_n) tends to fall — the inequality direction, not the intuition of the words, is what counts.

Second, the martingale property is never absolute: it is always relative to a specified filtration (\mathcal{F}_n) and a probability measure \mathbb{P}. Change the information structure or the measure and the same process can stop being a martingale — or start being one. That is not a technicality but the entire engine of risk-neutral pricing: a discounted stock price is generally not a martingale under the real-world measure (it drifts upward as a reward for risk), yet becomes one under the equivalent risk-neutral measure \mathbb{Q}, and that switch is exactly what makes arbitrage-free option prices well defined.