Random Walk

Motivation: why this matters in quant finance

The shortest route from a binomial tree to Black-Scholes is a random walk. A stock that moves up or down each period, a hedging error accumulated over rebalancing dates, and a cumulative P&L stream are all partial sums of random shocks. Before continuous-time notation writes $dS_t$ or $dW_t$ , the discrete object is

S_n=X_1+\cdots+X_n.

This matters because the most important continuous process in classical pricing, Brownian motion, is the scaling limit of random walks. Lawler's route into stochastic calculus makes this explicit: sums with many small independent finite-variance increments fall into the central-limit regime, while rare-event sums lead instead to Poisson jumps. That contrast is the first modelling fork between diffusion models and jump models.

A random walk also teaches the finance lesson that expected drift and realised risk live on different scales. Expected movement accumulates like

n\mu

, while noise accumulates like

\sqrt{n}\sigma

. That is why annualised volatility uses square-root-of-time scaling, why estimating expected returns is harder than estimating volatility, and why the binomial tree model becomes useful only after the time and space increments are calibrated together.

This lesson is the discrete-time foundation for martingales, geometric Brownian motion, stochastic differential equations, and the scaling intuition behind Itô's lemma.

The informal idea

A random walk is cumulative surprise. At each date a new shock arrives, and the process records the running total. The simplest case is a fair coin game: heads adds $+1$ , tails adds $-1$ . After ten flips, the position is not the last flip; it is the sum of all ten shocks.

That running-sum structure makes random walks the discrete prototype for price changes and trading gains. If $X_i$ is the $i$ th log-return shock, then $S_n$ is the cumulative log-return. If $H_i$ is the number of shares held before the $i$ th price change, then $\sum_i H_i\Delta S_i$ is the discrete trading gain. Lawler uses this predictable-betting viewpoint as the discrete ancestor of stochastic integration.

The core intuition is the square-root law. Independent shocks do not add their typical sizes linearly. Their variances add. After $n$ fair unit shocks, the variance is $n$ and the typical displacement is order $\sqrt{n}$ , not order $n$ . That is the statistical reason a many-step random walk can converge to a continuous process after space is scaled by $\sqrt{n}$ .

Formal definitions

Let $X_1,X_2,\ldots$ be independent identically distributed random variables and define the natural filtration

\mathcal{F}_n=\sigma(X_1,\ldots,X_n).

The random walk with increments

X_i

is the partial-sum process

S_0=0,\qquad S_n=\sum_{i=1}^n X_i.

The simple symmetric random walk is the special case

\mathbb{P}(X_i=1)=\mathbb{P}(X_i=-1)=\frac{1}{2}.

More generally, a biased nearest-neighbour walk has

\mathbb{P}(X_i=1)=p,\qquad \mathbb{P}(X_i=-1)=q=1-p.

Then

\mathbb{E}[X_i]=p-q=2p-1,\qquad \text{Var}(X_i)=4pq.

For stock prices, the additive walk usually belongs to log-prices rather than prices. If $R_i$ are log-return increments, then

\log S_n=\log S_0+\sum_{i=1}^n R_i,

so the price itself evolves multiplicatively:

S_n=S_0\exp(R_1+\cdots+R_n).

Key properties

Moments separate drift from noise

If the increments have mean $\mu$ and variance $\sigma^2<\infty$ , then

\mathbb{E}[S_n]=n\mu,\qquad \text{Var}(S_n)=n\sigma^2.

The expected displacement grows linearly in time, but the standard deviation grows like

\sqrt{n}\sigma

. In finance, this is the signal-to-noise problem for expected returns: drift estimates require far more data than volatility estimates.

Independent increments make the walk Markovian

For $m<n$ ,

S_n-S_m=X_{m+1}+\cdots+X_n

is independent of

\mathcal{F}_m

. Conditional on the current position, the future evolution depends only on new increments. This is the discrete ancestor of the independent-increment property in Brownian motion and Poisson processes.

A centred random walk is a martingale

If $\mathbb{E}[X_i]=0$ , then

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

The simple symmetric random walk is therefore a martingale. A biased random walk is not a martingale, but subtracting its predictable drift gives one:

M_n=S_n-n\mu.

Predictable betting produces a discrete stochastic integral

A sequence $J_1,J_2,\ldots$ is predictable if $J_n$ is $\mathcal{F}_{n-1}$ -measurable. It is chosen before the $n$ th shock arrives. The discrete gain process is

Z_n=\sum_{j=1}^n J_j\Delta S_j,\qquad \Delta S_j=X_j.

Lawler stresses three properties that survive into continuous stochastic integration: $Z_n$ is a martingale when the walk is centred, the integral is linear in $J$ , and its variance is controlled by the accumulated $J_j^2$ terms.

The Brownian scaling is forced by variance

If one unit of continuous time contains $N$ random-walk steps, then $\Delta t=1/N$ . To keep terminal variance of order one, the space step must satisfy

(\Delta x)^2N=1,

\Delta x=\frac{1}{\sqrt{N}}=\sqrt{\Delta t}.

This is the discrete origin of the stochastic-calculus rule of thumb that Brownian increments have size $\sqrt{dt}$ .

Worked examples

Example 1: a biased one-year P&L walk

Suppose a daily P&L model has increments $X_i\in\{1,-1\}$ with $p=0.52$ for a gain and $q=0.48$ for a loss. Then

\mu=\mathbb{E}[X_i]=0.04,\qquad \sigma^2=\text{Var}(X_i)=4pq=0.9984.

Over $250$ trading days,

\mathbb{E}[S_{250}]=250(0.04)=10,

while

\text{SD}(S_{250})=\sqrt{250(0.9984)}\approx 15.8.

The expected gain is positive, but it is smaller than one standard deviation. This is the arithmetic behind the practical difficulty of detecting small return premia from noisy price histories.

Example 2: why log-prices, not prices, carry the additive walk

Let a stock multiply each period by $u$ or $d$ , with $u>1$ and $0<d<1$ . Then

S_n=S_0\prod_{i=1}^n Z_i,\qquad Z_i\in\{u,d\}.

Taking logs gives

\log S_n=\log S_0+\sum_{i=1}^n \log Z_i.

The log-price is the additive random walk. This is why the multiplicative binomial tree model converges naturally to geometric Brownian motion, while an additive price walk can cross below zero.

Example 3: Brownian scaling from a coin walk

Let $X_i=\pm 1$ with equal probability and define

W^{(N)}_t=\frac{1}{\sqrt{N}}S_{\lfloor Nt\rfloor}.

At $t=1$ ,

\text{Var}(W^{(N)}_1)=\frac{1}{N}\text{Var}(S_N)=1.

At time $t$ ,

\text{Var}(W^{(N)}_t)\approx t.

The central limit theorem gives Gaussian one-time limits. Donsker's theorem strengthens this to convergence of the whole path, which is why the limiting continuous model has Brownian rather than arbitrary noise.

Common confusions and pitfalls

"A fair walk is safe because it has zero expectation." Zero expectation means fair, not low risk. The variance still grows linearly with the number of steps.

"Drift dominates noise over any useful horizon." Drift grows like $n\mu$ , but noise grows like $\sqrt{n}\sigma$ . If $\mu$ is small relative to $\sigma$ , the sample size needed to see the drift can be enormous.

"The square-root-of-time rule is a market convention." It is the random-walk variance formula in disguise. It applies when increments are independent with stable variance; it can fail under volatility clustering, jumps, or serial dependence.

"An additive random walk is a good stock-price model if the step is small." Even small additive steps can eventually make prices negative. For equities, the additive walk usually belongs to log-prices.

"The random walk becomes Brownian motion path by path." The convergence is distributional. A fixed discrete path does not literally become a Brownian path; the law of the rescaled process approaches the Brownian law.

Where this goes next

Brownian Motion: Defines the continuous Gaussian process obtained from the random-walk scaling limit.
Geometric Brownian Motion: Applies the additive walk to log-prices and keeps prices positive.
Martingales: Generalises the fair-game property of the centred walk.
Binomial Tree Model: Uses a multiplicative random walk for option pricing by backward induction.
Poisson Processes: Shows the other major scaling limit when rare jumps, not small finite-variance shocks, dominate.
Itô's Lemma: Uses the Brownian $\sqrt{dt}$ scaling that is already visible in the random-walk limit.

References

Lawler, G. F. (2023). Stochastic Calculus: An Introduction with Applications. Ch. 1 §1.6 (Integrals with respect to random walk), Ch. 2 §2.1 (Limits of sums of independent variables), §2.3 (Limits of random walks).