Random Walks & Martingales

You’ve met the idea that randomness can be modeled, step by step. Now we meet the two creatures that show up everywhere a price, a bankroll, or a forecast wanders through time: the random walk and the martingale. The first is the simplest possible model of accumulated chance — a running total of coin flips. The second is the precise mathematical statement of a fair game: a process with no predictable lean, where your best guess of tomorrow is exactly today. They are the skeleton under “stock prices look unpredictable,” and getting them exactly right — including what they do not say — is what separates a real understanding of market efficiency from a folk slogan.

The simple random walk

Analogy. Picture a coin-flipping drunk on a sidewalk. Heads, he lurches one step right; tails, one step left. After many flips, where is he? Not nowhere in particular — but not drifting anywhere either. His position is the running sum of independent shoves, and that sum has a beautiful, predictable shape even though every step is pure chance.

Definition. A simple random walk is the running total $S_n = X_1 + X_2 + \cdots + X_n,$ where each increment $X_i$ is an independent, identically distributed step — say $+1$ or $-1$ with equal probability (or $\pm c$ for a step of size $c$). $S_0 = 0$; each new position is just the last one plus a fresh independent shock: $S_n = S_{n-1} + X_n$.

Two facts do all the work, and both come straight from arithmetic on the increments.

Fact 1 — the mean stays put. Each fair step has $E[X_i] = \tfrac{1}{2}(+1) + \tfrac{1}{2}(-1) = 0$. Expectation adds, so $E[S_n] = E[X_1] + \cdots + E[X_n] = 0.$ On average the walk goes nowhere. A symmetric walk has no built-in lean.

Fact 2 — the spread grows like √n. Independent variances add. For a $\pm c$ step, $\mathrm{Var}(X_i) = c^2$ (since the mean is 0 and the squared deviation is always $c^2$). So $\mathrm{Var}(S_n) = n\,c^2, \qquad \mathrm{SD}(S_n) = c\sqrt{n}.$ The variance grows linearly with the number of steps, but the standard deviation — the typical distance from the start — grows only like the square root of time. This is the famous √t spreading law, and it is why uncertainty piles up slower than you’d guess.

Worked example. Take 100 fair steps of $1 each (win or lose a dollar on each flip), so the step size is $c = 1$.

• Expected final position: $E[S_{100}] = 0$ — you expect to break even.
• Standard deviation: $\mathrm{SD}(S_{100}) = 1 \times \sqrt{100} = 10$, i.e. $10.
• Because $S_{100}$ is (by the central-limit theorem) roughly normal, about 95% of the time you land within $\pm 2$ standard deviations — that is, between −$20 and +$20.

Notice the asymmetry of scale: 100 steps, each worth $1, could in principle leave you anywhere from −$100 to +$100 — but the typical outcome is only about $10 from zero. Cancellation is the rule; extreme runs are rare.

Adding drift

The fair walk is symmetric, but nothing forces a coin to be fair. Tilt it — make each step carry a small average push — and you get a biased random walk with drift.

Definition. Let each increment have mean $E[X_i] = \mu \neq 0$ while keeping a finite variance $\sigma^2$. Then $E[S_n] = n\mu, \qquad \mathrm{SD}(S_n) = \sigma\sqrt{n}.$ The expected position now grows linearly in $n$ (the drift $n\mu$), while the spread still grows only like $\sqrt{n}$. That mismatch — linear signal versus square-root noise — is the whole story of telling skill from luck.

Worked example — a tiny edge. Suppose a trader (or a loaded coin) has a per-step edge of $\mu = +0.02$ with step-scale $\sigma = 1$. Over $n = 100$ steps:

• Drift (signal): $n\mu = 100 \times 0.02 = +2$.
• Noise (spread): $\sigma\sqrt{n} = 1 \times \sqrt{100} = 10$.

After 100 steps the expected gain is $+2$, but the noise is $\pm 10$ — five times larger. The edge is real, yet it’s buried; a single 100-step sample could easily show a loss and tell you nothing. The drift is invisible short-term.

Now stretch the horizon. The drift $n\mu$ scales with $n$; the noise $\sigma\sqrt{n}$ scales with $\sqrt{n}$. So the signal-to-noise ratio is $\frac{n\mu}{\sigma\sqrt{n}} = \frac{\mu}{\sigma}\sqrt{n},$ which grows without bound as $n$ increases. At $n = 100$ it’s $0.02 \times 10 = 0.2$ (drift is a fifth of the noise). At $n = 10{,}000$ it’s $0.02 \times 100 = 2$ (drift is double the noise). Eventually drift always wins — but “eventually” can be a very long time.

The martingale: a fair game

Strip the drift back out and you get the single most important object in this lesson: a martingale. The word names a precise property — a fair game — and the precision matters.

Definition. A process $X_0, X_1, X_2, \dots$ is a martingale if, given everything known up to today (call that information set $\mathcal{F}_t$), the conditional expectation of tomorrow’s value equals today’s: $E[X_{t+1} \mid \mathcal{F}_t] = X_t.$ In words: your best forecast of tomorrow, using all the information you have right now, is exactly today’s value. No predictable push up, no predictable push down. The martingale property is about the conditional expectation — the forecast given the present — not about the variance.

That last point is the whole trap. A martingale is not a constant, and it is not low-variance. Tomorrow’s value is still uncertain and can swing wildly; the claim is only that the average of that swing, conditioned on today, is centered on today. A driftless random walk is the textbook example: $E[S_{n+1} \mid \mathcal{F}_n] = S_n + E[X_{n+1}] = S_n + 0 = S_n$. It’s a martingale, yet (we just saw) its spread grows like $\sqrt{n}$ forever.

Examples.

• A driftless random walk — fair coin flips — is a martingale.
• A fair casino bankroll: each bet has expected change zero, so your expected bankroll tomorrow is your bankroll today. A martingale.
• Any process with no predictable trend, where knowing the past gives you no edge on the next move’s average.

Two cousins capture biased games:

• A submartingale satisfies $E[X_{t+1} \mid \mathcal{F}_t] \ge X_t$ — it tends to drift up. A stock earning a risk premium is a submartingale: on average it pays you to hold it.
• A supermartingale satisfies $E[X_{t+1} \mid \mathcal{F}_t] \le X_t$ — it tends to drift down. A gambler’s bankroll at an unfair casino (house edge) is a supermartingale; so is a fee-draining fund.

A handy mnemonic: a supermartingale is bad luck for the holder (it leaks down), a submartingale is good luck (it builds up) — the names feel backwards because they’re named after the inequality on the expectation, not the direction of the money.

Why efficient markets look like martingales

Here’s the payoff — the reason this abstract machinery is everywhere in finance.

The argument. Suppose a price already reflects all available information. Then the expected next change, beyond the baseline compensation for time and risk, must be zero. Why? Because if everyone could see that the price would predictably rise tomorrow, they’d buy today — pushing the price up now until the predictable gain vanished. Any forecastable excess move is free money, and free money gets competed away instantly. What survives is a process whose best forecast of the next (suitably adjusted) move is no move at all: a martingale.

More carefully: it’s the discounted, risk-adjusted price that behaves like a martingale. Strip out the risk-free growth and the risk premium, and what remains has expected change zero. (Pricing theory formalizes this as the existence of a martingale measure — a reweighting of probabilities under which discounted prices are exactly martingales. You’ll meet it properly later; for now, just hold the picture: efficiency = no predictable excess move.)

Pitfall 1 — “random walk” does not mean “no drift.” Real equity indices are submartingales, not martingales: they carry an upward drift, the equity risk premium, your reward for bearing risk. Efficiency kills predictable excess returns — the part you could arbitrage — not the risk premium, which is compensation, not a free lunch. Saying “the market is a random walk so it has no trend” fuses two different claims and gets the second one wrong.

Pitfall 2 — a martingale is not safe. “No predictable trend” says nothing about risk. A martingale’s variance still grows like $\sqrt{t}$ (or faster), so it can wander enormously far from where it started. Fair does not mean gentle. Confusing “you can’t forecast the direction” with “it won’t move much” is exactly the error the pretest flagged.

Gambler’s ruin & optional stopping (intuition)

One last pair of ideas that follow from the martingale property and demolish a lot of wishful thinking.

Gambler’s ruin. Play a fair game (a martingale bankroll) with a finite purse against a house with effectively infinite money, and keep playing forever. You go broke with probability 1. The game is fair — your expected bankroll never drops — yet ruin is certain, because the $\sqrt{t}$ wandering will eventually touch zero, and zero is an absorbing wall (broke is broke; the house, with infinite reserves, has no such wall). You cannot out-wait variance. Fairness protects your average, not your survival.

Worked ruin probability. Suppose you start with stake $a$ and quit either at $0 (ruin) or at a target $b > a$, playing a fair $\pm1$ walk. The probability of ruin before hitting the target is $P(\text{ruin}) = \frac{b - a}{b}.$ Start with $20 aiming for $100 (so $a = 20$, $b = 100$): ruin probability $= (100 - 20)/100 = 0.8$. An 80% chance of going broke before you ever see $100 — in a perfectly fair game. (And against an infinite house, set $b \to \infty$ and the ruin probability $\to 1$.)

Optional stopping (intuition). You might hope to beat a martingale with a clever exit rule — quit the instant you’re ahead, double down when behind, and so on. The optional stopping theorem says: under reasonable conditions, no stopping strategy changes your expected payoff. For a fair martingale, your expected value when you stop equals your starting value, whatever rule you use. The “quit while you’re ahead” schemes (like the classic Martingale doubling system) don’t manufacture an edge — they just reshuffle a near-certain small win against a rare catastrophic loss, leaving the expectation exactly where it started. There is no free lunch hiding in the timing.

Putting it together

A simple random walk is a running sum of independent shocks: its mean stays at the start (for fair steps) while its spread widens like $\sqrt{n}$ — uncertainty piling up slower than the step count. Drift adds a per-step lean $\mu$, so the expected position grows linearly ($n\mu$) while noise still grows like $\sqrt{n}$ — which is why a small edge is invisible short-term but inevitable long-term. A martingale is the precise statement of a fair game: $E[X_{t+1}\mid\mathcal{F}_t] = X_t$, a forecast-equals-today property about the conditional expectation, not about variance — so it can still wander enormously. Submartingales drift up (a stock’s risk premium), supermartingales drift down (a house edge). Efficient markets look like martingales because any predictable excess move would be arbitraged away — but real indices are submartingales, carrying a risk premium, and a martingale is never “safe.” And gambler’s ruin plus optional stopping seal the deal: you can’t out-wait variance, and no clever exit rule beats a fair game.

Next up — Brownian motion — we let the steps shrink to zero and arrive in a step at a time, taking the random walk’s √t spreading to its smooth, continuous limit. That continuous-time martingale is the raw material for every price model you’ll build from here: feed it a drift and a volatility, wrap it in an exponential, and you’ve reinvented the engine behind simulated asset prices.