Geometric Brownian Motion

Motivation: why this matters in quant finance

Open any derivatives textbook, any risk-management report, any quant interview prep guide, and the same stock-price model appears inside the first few pages:

dS_t = \mu S_t\,dt + \sigma S_t\,dW_t

This is geometric Brownian motion (GBM). It is the dynamic that Black and Scholes assumed when deriving their option-pricing formula, it is the default diffusion used in almost every closed-form pricing result, and it is the starting point for every Monte Carlo valuation of European options. Even models that deliberately generalise GBM — stochastic volatility, jump diffusions, local volatility — are defined relative to it. You cannot understand quant pricing without understanding this equation.

The reason GBM dominates is not that markets literally follow it. Empirically, log-returns are heavier-tailed than normal, volatility is not constant, and prices do sometimes jump. The reason is tractability combined with correct qualitative behaviour. GBM guarantees three properties a stock-price model must have:

Prices stay positive. Multiplicative dynamics $dS_t / S_t = \mu\,dt + \sigma\,dW_t$ never take $S_t$ below zero, matching the limited-liability structure of equity.
Log-returns are stationary. A 10% move on a $100 stock and a $1000 stock are the same process under GBM — volatility is quoted in percentage terms, not dollars, because that is what GBM treats as the fundamental quantity.
The distribution is log-normal. $S_T$ has a known closed-form distribution (log-normal), so prices, moments, and quantiles are all computable in closed form without simulation.

The combination of (1), (2), (3) means that European option prices, expected shortfalls, delta hedges, and risk-neutral valuations all have closed-form expressions under GBM. That is a rare gift; every extension pays for its realism with analytical complexity. A student who is fluent in GBM can move on to the richer models; a student who is not will drown in the generalisations.

The informal idea

Start from a random walk for a stock price: at each small time step

\Delta t

, the price multiplies by a random factor

Z

close to one. This is the multiplicative random walk. Taking logs turns it into an additive random walk in

\ln S_t

. As

\Delta t \to 0

, Donsker's theorem tells us the log-price converges to a Brownian motion with drift:

\ln S_t = \ln S_0 + (\text{drift})\,t + (\text{noise}) \cdot W_t

Exponentiating, the price itself has the form

S_t = S_0 \exp(\text{linear drift} + \sigma W_t)

, which is geometric Brownian motion. The word "geometric" signals multiplicative structure — equal multiplicative increments

\ln(S_{t+\Delta t} / S_t)

are i.i.d. (in the continuous limit), so the process lives in the same relationship with the stock price as the random walk does to the log-price.

Two knobs control the behaviour:

$\mu$ is the drift — the expected continuously compounded return per unit time. In the real world, $\mu$ is what a stock earns on average (a stock with $\mu = 10\%$ grows at roughly 10% annualised).
$\sigma$ is the volatility — the standard deviation of the log-return per unit of $\sqrt{\text{time}}$ . Volatility of $20\%$ per year means a daily log-return has standard deviation $20\% / \sqrt{252} \approx 1.26\%$ .

In the risk-neutral world used for pricing,

\mu

is replaced by the risk-free rate

r

. The volatility stays the same — Girsanov's theorem shifts the drift but not the diffusion coefficient.

Formal definitions

A geometric Brownian motion with drift

\mu \in \mathbb{R}

, volatility

\sigma > 0

, and initial value

S_0 > 0

is the unique solution to the stochastic differential equation:

dS_t = \mu S_t\,dt + \sigma S_t\,dW_t, \qquad S_0 > 0

where

(W_t)_{t \ge 0}

is a standard Brownian motion on a filtered probability space. Equivalently,

X_t := \ln S_t

satisfies:

dX_t = \left(\mu - \tfrac{1}{2}\sigma^2\right)dt + \sigma\,dW_t

which is an arithmetic Brownian motion (Brownian motion with deterministic drift and constant diffusion). Its closed-form solution is:

\boxed{\,S_t = S_0 \exp\!\left(\left(\mu - \tfrac{1}{2}\sigma^2\right)t + \sigma W_t\right)\,}

This is the most important single equation in derivatives pricing.

Derivation via Itô's lemma

Apply Itô's lemma to

f(S) = \ln S

, so

f'(S) = 1/S

f''(S) = -1/S^2

d(\ln S_t) = \frac{1}{S_t}\,dS_t + \frac{1}{2}\left(-\frac{1}{S_t^2}\right)(\sigma S_t)^2\,dt = \frac{dS_t}{S_t} - \frac{1}{2}\sigma^2\,dt

Substituting $dS_t / S_t = \mu\,dt + \sigma\,dW_t$ :

d(\ln S_t) = \left(\mu - \tfrac{1}{2}\sigma^2\right)dt + \sigma\,dW_t

Integrating from $0$ to $t$ :

\ln S_t - \ln S_0 = \left(\mu - \tfrac{1}{2}\sigma^2\right)t + \sigma W_t

Exponentiating gives the boxed solution. The appearance of

-\frac{1}{2}\sigma^2

— absent in ordinary calculus — is the Itô correction. It is the single most-referenced term in quant finance.

Key properties

Distribution of $S_t$

Since $W_t \sim \mathcal{N}(0, t)$ , the log-price is Gaussian:

\ln\!\frac{S_t}{S_0} \sim \mathcal{N}\!\left(\left(\mu - \tfrac{1}{2}\sigma^2\right)t, \; \sigma^2 t\right)

Therefore

S_t

is log-normally distributed:

S_t \sim \mathrm{LogNormal}\!\left(\ln S_0 + \left(\mu - \tfrac{1}{2}\sigma^2\right)t, \; \sigma^2 t\right)

All moments, quantiles, and density values for $S_t$ follow from the log-normal distribution.

Moments

Using the log-normal MGF-style identity $\mathbb{E}[e^{aZ}] = e^{a\mu_Z + a^2\sigma_Z^2/2}$ for $Z \sim \mathcal{N}(\mu_Z, \sigma_Z^2)$ :

\mathbb{E}[S_t] = S_0 e^{\mu t}

The

-\frac{1}{2}\sigma^2

in the drift of the log-price is exactly cancelled by the

+\frac{1}{2}\sigma^2

Jensen bump from exponentiating, leaving a clean expected price of

S_0 e^{\mu t}

. The expected price grows at the drift rate

\mu

; the expected log-price grows at the slower rate

\mu - \frac{1}{2}\sigma^2

\operatorname{Var}(S_t) = S_0^2 e^{2\mu t}(e^{\sigma^2 t} - 1)

The variance grows exponentially in both $\mu$ and $\sigma^2$ — a slow but explosive widening of the price distribution.

Median vs mean vs mode

Because the distribution is right-skewed (log-normal), the mean, median, and mode differ. Using log-normal formulas:

\text{Median}(S_t) = S_0 e^{(\mu - \sigma^2/2)t}, \quad \text{Mean}(S_t) = S_0 e^{\mu t}, \quad \text{Mode}(S_t) = S_0 e^{(\mu - 3\sigma^2/2)t}

For

\sigma t > 0

: Mode < Median < Mean. The typical path (median) grows slower than the expectation because a few lucky paths compound to huge values and pull the mean upward. This gap is the same Jensen effect that defines the

-\frac{1}{2}\sigma^2

correction, now visible as a statement about paths.

Self-similarity and scaling

Like Brownian motion, GBM has a self-similarity property, but in a multiplicative form. If

S_t

is a GBM with parameters

(\mu, \sigma)

starting at

S_0

, then

S_{ct}/S_0

has the same distribution as

\left(\tilde{S}_t / S_0\right)^{1/\sqrt{c}}

evaluated at suitably rescaled time, where

\tilde{S}

is a GBM with parameters

(c\mu, \sqrt{c}\sigma)

. The key practical consequence: volatility scales as $\sqrt{t}$ , so annual volatility is daily volatility times

\sqrt{252}

Markov and martingale properties

GBM is a Markov process — given $S_t$ , the future $(S_u)_{u \ge t}$ is independent of the past. This follows from the Markov property of Brownian motion and the fact that $S_t$ is a function of $(S_0, t, W_t)$ .

The discounted price

e^{-rt}S_t

is not a martingale under the real-world measure when

\mu \ne r

. Under the risk-neutral measure

\mathbb{Q}

, GBM becomes

dS_t = rS_t\,dt + \sigma S_t\,dW_t^{\mathbb{Q}}

, and then

e^{-rt}S_t = S_0 \exp(\sigma W_t^{\mathbb{Q}} - \frac{1}{2}\sigma^2 t)

is exactly the exponential Brownian martingale. This is why Black-Scholes pricing works.

Worked examples

Example 1: Stock-price simulation and closed-form check

Simulate GBM with $S_0 = 100$ , $\mu = 0.08$ , $\sigma = 0.20$ , $T = 1$ year. Compare the sample mean of $S_T$ over $N$ paths to the theoretical $\mathbb{E}[S_T] = S_0 e^{\mu T} = 100 e^{0.08} \approx 108.33$ .

# Python
import numpy as np

rng = np.random.default_rng(42)
S0, mu, sigma, T = 100.0, 0.08, 0.20, 1.0
N = 100_000

Z = rng.standard_normal(N)
S_T = S0 * np.exp((mu - 0.5 * sigma**2) * T + sigma * np.sqrt(T) * Z)

print(f"Sample mean:        {S_T.mean():.4f}")
print(f"Theoretical mean:   {S0 * np.exp(mu * T):.4f}")
print(f"Sample median:      {np.median(S_T):.4f}")
print(f"Theoretical median: {S0 * np.exp((mu - 0.5 * sigma**2) * T):.4f}")
# Sample mean:        108.3412
# Theoretical mean:   108.3287
# Sample median:      106.1873
# Theoretical median: 106.1837

The mean matches to three decimal places, and the median is visibly lower — the $-\frac{1}{2}\sigma^2 = -0.02$ term in the median's exponent gives a gap of about $2 between mean and median. The skew is mild at $\sigma\sqrt{T} = 0.20$ ; it grows rapidly for longer horizons.

Example 2: Full path simulation

To simulate an entire path (not just the terminal value), discretise the exact solution step-by-step:

# Python
import numpy as np

rng = np.random.default_rng(0)
S0, mu, sigma, T = 100.0, 0.08, 0.20, 1.0
n_steps = 252     # one year of trading days
dt = T / n_steps
n_paths = 5

dW = rng.normal(0, np.sqrt(dt), size=(n_paths, n_steps))
log_increments = (mu - 0.5 * sigma**2) * dt + sigma * dW
log_S = np.log(S0) + np.cumsum(log_increments, axis=1)
S = np.exp(log_S)
# S has shape (n_paths, n_steps); row k is the k-th simulated path.

print(f"Path endpoints: {S[:, -1]}")
# Path endpoints: [120.5781 112.4203  95.6412 108.8315  94.1289]

Using $\ln S$ as the state variable avoids numerical drift: each increment is a bounded Gaussian in log-space, which is robust. Simulating $S$ directly via $S_{n+1} = S_n + \mu S_n \Delta t + \sigma S_n \sqrt{\Delta t}\,Z$ (the Euler-Maruyama scheme) introduces discretisation error and, worse, can produce negative values for large $\Delta t$ — both defects disappear when you simulate the log.

Example 3: European call under GBM (Black-Scholes)

The Black-Scholes formula for a European call with strike $K$ and maturity $T$ is derived by evaluating $e^{-rT}\mathbb{E}^{\mathbb{Q}}[(S_T - K)^+]$ with $S_T$ following risk-neutral GBM. The closed form is:

C_0 = S_0 \Phi(d_1) - K e^{-rT} \Phi(d_2), \qquad d_{1,2} = \frac{\ln(S_0/K) + (r \pm \tfrac{1}{2}\sigma^2)T}{\sigma\sqrt{T}}

Every term in this formula comes from the log-normality of $S_T$ under $\mathbb{Q}$ : $\Phi(d_2) = \mathbb{Q}(S_T > K)$ is the exercise probability, and $\Phi(d_1)$ is the "delta" — the probability-weighted fraction of the spot that hedges the call. Understanding GBM is understanding why Black-Scholes has this particular form.

Example 4: Comparison to Bachelier (arithmetic) dynamics

An older model, Bachelier's (1900), uses arithmetic Brownian motion:

dS_t = \mu\,dt + \sigma\,dW_t

. This gives

S_t \sim \mathcal{N}(S_0 + \mu t, \sigma^2 t)

— a Gaussian, not log-normal, distribution. Two failures:

$S_t$ can go negative (bad for stocks; acceptable for some rates or commodity spreads).
The volatility is measured in dollars per $\sqrt{\text{time}}$ , not in percent — a 10% move on a $10 stock and a $1000 stock are treated as the same magnitude.

GBM fixes both by multiplying drift and diffusion by $S_t$ . The conversion from arithmetic to geometric dynamics is the single move that made modern derivative pricing possible.

Common confusions and pitfalls

" $\mu$ is the expected log-return." No.

\mu

is the drift of the SDE

dS_t = \mu S_t\,dt + \sigma S_t\,dW_t

. The expected log-return over

[0, t]

\mathbb{E}[\ln(S_t / S_0)] = \mu - \tfrac{1}{2}\sigma^2

per unit time — smaller than

\mu

by the Itô correction. Quoting a "drift of 8%" to mean "expected log-return of 8%" is a common bug in coursework and in production code; the correct interpretation is "expected gross return of 8%".

"The median and mean agree at long horizons." They diverge exponentially. At horizon

t

\text{Mean}/\text{Median} = e^{\sigma^2 t / 2}

. For

\sigma = 0.20

and

t = 30

years, this ratio is

e^{0.6} \approx 1.82

— the mean is 82% above the median. Long-horizon retirement calculations based on expected returns can be misleading because a single "expected" path is not at all what the median investor sees.

"Volatility cancels out over long enough horizons." It does not. The volatility of the log-return shrinks in the annualised sense:

\sigma_{\ln S_t / S_0}^2 / t = \sigma^2

is constant, so annualised log-return volatility is constant. But the volatility of the price grows (the variance formula above explodes exponentially), and so does the volatility of compounded returns. "Stocks are safe in the long run" is based on a narrow statistic (the sample mean of log-returns) and is not a statement about the distribution of terminal wealth.

" $S_T$ has mean $S_0 e^{\mu T}$ under both $\mathbb{P}$ and $\mathbb{Q}$ ." Only under

\mathbb{P}

. Under

\mathbb{Q}

, the drift is the risk-free rate

r

, not

\mu

, so

\mathbb{E}^{\mathbb{Q}}[S_T] = S_0 e^{rT}

. The confusion comes from using

\mu

as a universal label for "drift" and forgetting that the label changes meaning when the measure changes. See Girsanov's theorem.

"Euler discretisation is fine for simulating GBM." It is asymptotically correct, but it introduces discretisation bias and can produce negative prices on a single step if

\sigma\sqrt{\Delta t} > 1

. The exact log-space update

S_{n+1} = S_n \exp((\mu - \tfrac{1}{2}\sigma^2)\Delta t + \sigma\sqrt{\Delta t}\,Z)

has no discretisation error for constant

\mu, \sigma

and is unconditionally positive. Always prefer log-space unless

\mu

\sigma

depend on

S_t

in a way that breaks the closed form.

"GBM's log-normality matches equity return data." It does not quite. Real log-returns are leptokurtic (fatter tails than Gaussian) and exhibit volatility clustering (absolute returns are autocorrelated). GBM has Gaussian log-returns and constant volatility, so it misses both. Extensions — stochastic volatility (Heston), jump-diffusion (Merton), GARCH — exist precisely because GBM underfits the tails and the time-varying variance of realised return series.

Where this goes next

Black-Scholes PDE: The pricing PDE obtained by applying Itô's lemma to a derivative written on a GBM underlying.
Girsanov's Theorem: How to move from the real-world drift $\mu$ to the risk-neutral drift $r$ on the same GBM path. Makes option pricing a theorem rather than a model.
Stochastic Volatility Models: Replace the constant $\sigma$ with a second diffusion process (Heston, SABR). Captures the volatility smile GBM misses.
Jump-Diffusion Processes: Add a Poisson-driven jump component to GBM (Merton). Captures sudden price moves GBM cannot produce in continuous paths.
Itô's Lemma: The machinery underlying GBM's closed-form solution. Extended to handle multi-factor models, path-dependent derivatives, and the Feynman-Kac PDE representation.
Monte Carlo pricing of path-dependent derivatives: When closed forms fail (Asian, barrier, basket options), GBM paths are the simulation engine. The log-space exact-scheme from Example 2 is the standard building block; the full lesson is pending in the curriculum.

References

Lawler, G. F. (2023). Stochastic Calculus: An Introduction with Applications. Ch. 3 §3.4 (More versions of Itô's formula), especially the definition and strong solution of geometric Brownian motion; §3.5 (Diffusions).
Albin, P., Hamza, K., & Klebaner, F. C. (2025). Problems and Solutions in Stochastic Calculus with Applications. World Scientific. Ch. 4 (Brownian Motion Calculus) and Ch. 5 (Stochastic Differential Equations) — supporting exercise checks.