Open any quant's notebook — the pricing model behind an options desk, the interest-rate model behind a swaps book, the volatility model behind a risk report — and one sentence appears over and over, in a hundred dialects:

Read it aloud: over the next instant, X moves by a predictable trend \mu\, dt plus a random kick \sigma\, dW_t. That is the entire grammar of continuous-time finance. Geometric Brownian motion — the Black–Scholes stock — speaks it. The Ornstein–Uhlenbeck process speaks it. Every classical interest-rate model — Vasicek, Cox–Ingersoll–Ross, Hull–White — speaks it. They differ only in which drift \mu and which volatility \sigma they plug in. A process built this way is called an Itô process, and this page is about learning to read that sentence fluently — and to multiply its differentials, which is where the calculus gets strange.

The definition

With the Itô integral and its properties in hand we can name the central object of mathematical finance precisely. An Itô process is anything built from a drift integral plus a stochastic integral:

with adapted integrands a and b. The first integral is an ordinary (finite-variation) Riemann integral; the second is the Itô integral, a martingale. The whole thing is written compactly in differential form:

Here a_t is the drift — the predictable trend, the smooth part — and b_t is the diffusion (or volatility) — the size of the random kick. Finance usually writes the drift as \mu and the diffusion as \sigma, and in the most important models the coefficients are functions of the current state:

An equation of this shape — the coefficients fed back the process's own value — is a stochastic differential equation (SDE), and its solution is an Itô process. Almost every model you will meet, from a stock price to an interest rate to a volatility surface, is one of these; the job of stochastic calculus is to compute with them.

The anatomy of one tiny step

What does dX = \mu\, dt + \sigma\, dW actually promise over a small interval of length dt? Condition on everything known at time t and look at the increment \Delta X = X_{t+dt} - X_t:

• the drift term contributes a mean move of \mu\, dt — deterministic, proportional to the time elapsed;
• the noise term contributes zero mean (the Itô integral is a martingale) but a standard deviation of \sigma\sqrt{dt}, because dW_t \sim \mathcal{N}(0, dt);
• so, to leading order, \Delta X \approx \mathcal{N}\!\left(\mu\, dt,\; \sigma^2\, dt\right) — a tiny Gaussian nudge whose centre drifts and whose width is set by \sigma.

Put numbers on it. Model a price level with constant coefficients \mu = 8 per year and \sigma = 20 per \sqrt{\text{year}}, and step forward one trading day, dt = 1/252 \approx 0.004:

Over a single day the random kick is roughly forty times the size of the trend. Zoom out instead: over a year the mean is 8 and the s.d. is 20 — comparable; over ten years the drift has accumulated 80 while the s.d. has only grown to 20\sqrt{10} \approx 63 — now the trend wins. The drift grows like t, the noise only like \sqrt{t}: noise rules the short run, drift rules the long run. This one scaling fact explains why a day trader stares at volatility while a pension fund stares at drift — and it is the seed of everything on this page.

Worked example: arithmetic Brownian motion, solved in full

The simplest SDE has constant coefficients — and unlike almost every other SDE, we can solve it by just reading the definition. Take

Step 1 — write out the integral form. The differential is shorthand for

Step 2 — do both integrals. The first is an ordinary integral of a constant, \int_0^t \mu\, ds = \mu t. The second is the Itô integral of a constant, which just scales the Brownian increment: \int_0^t \sigma\, dW_s = \sigma (W_t - W_0) = \sigma W_t. So the solution is explicit:

This is arithmetic Brownian motion: a straight line of slope \mu with Brownian noise of size \sigma laid on top.

Step 3 — read off the distribution. Since W_t \sim \mathcal{N}(0, t) and the rest is deterministic, X_t is Gaussian with mean X_0 + \mu t and variance \sigma^2 t:

Step 4 — numbers. Start at X_0 = 100 with \mu = 4 and \sigma = 6, and ask about time t = 9: the mean is 100 + 4 \times 9 = 136 and the standard deviation is 6\sqrt{9} = 18, so X_9 \sim \mathcal{N}(136, 324) — about a two-thirds chance of landing in [118, 154]. Notice the division of labour once more: the drift \mu sets where the distribution is centred, the diffusion \sigma sets how wide it has spread — and the spread grows with \sqrt{t}, not t.

Drift plus diffusion, live

The figure below shows the arithmetic Brownian motion you just solved: X_t = X_0 + a\,t + b\,W_t, a straight drift line a\,t (dashed) with a diffusion wobble b\,W_t laid on top. The underlying Brownian path is frozen, so every move you see comes from the coefficients alone.

Play with it deliberately. Set b = 0: the noise vanishes and the process collapses onto its backbone — an Itô process with zero diffusion is just ordinary calculus. Now push b up and watch the same wiggles amplify: the roughness tracks b alone (the quadratic variation is b^2 t, read out above the curve) while a only tilts the trend without adding a single wrinkle. Finally, make the drift large and negative with a small b, then small with a large b — you are re-enacting the day-trader-versus-pension-fund scaling from the anatomy card: which term you see depends on how the two sliders compare, and the box-calculus conclusion d[X] = b^2\, dt is visible as pure texture.

The box algebra, line by line

To do calculus with dX_t = a_t\, dt + b_t\, dW_t we need to know how the differentials multiply. The rules are forced by a single scaling fact: over a short step of length dt, a Brownian increment has standard deviation \sqrt{dt}, so it is of size

We work out each product by comparing its order of magnitude to dt, keeping anything of order dt and discarding anything smaller.

Step 1 — (dt)^2 = 0. The product of two time-steps is of order (dt)^2, which is smaller than dt as dt \to 0 ((dt)^2 / dt = dt \to 0). Anything of higher order than dt is negligible, so we set it to zero:

Step 2 — dt\cdot dW = 0. Using dW \sim \sqrt{dt}, the cross product is of order dt \cdot \sqrt{dt} = (dt)^{3/2} — again higher order than dt ((dt)^{3/2}/dt = \sqrt{dt} \to 0), hence negligible:

Step 3 — (dW)^2 = dt. This is the one that does not vanish. The squared Brownian increment is of order (\sqrt{dt})^2 = dt — exactly the order we keep — and we know its precise value from the quadratic variation: summing squared increments converges to elapsed time, the differential shorthand of which is

These three lines are the whole multiplication table — tabulated in the theorem below. Note the asymmetry that makes stochastic calculus its own subject: (dt)^2 and dt\, dW die, but (dW)^2 survives as a genuine dt.

The quadratic variation of X

Now apply the table to an Itô process. The quadratic variation tracks the squared infinitesimal moves, so we compute (dX_t)^2 directly.

Step 4 — square the differential. With dX_t = a_t\, dt + b_t\, dW_t, expand the square (p + q)^2 = p^2 + 2pq + q^2:

Step 5 — apply the box rules term by term. The first term has (dt)^2 = 0 (Step 1); the cross term has (dt)(dW_t) = 0 (Step 2); only the last term survives, with (dW_t)^2 = dt (Step 3):

Step 6 — read off the quadratic variation. Accumulating the squared moves gives

The drift a has disappeared from the quadratic variation entirely — the smooth part contributes no roughness. Only the diffusion b drives the quadratic variation, which is exactly the [M]_t = \int_0^t H^2\, ds we met for the Itô integral, with H = b. The drift moves the process; the diffusion makes it rough.

A lineup of famous Itô processes

Everything so far used constant coefficients. The real power of the form dX = \mu(t,X)\, dt + \sigma(t,X)\, dW is that the coefficients can depend on the current state — and each choice encodes a different piece of financial intuition. The single most important choice makes both coefficients proportional to the level:

This is geometric Brownian motion (GBM), the stock model at the heart of Black–Scholes. Why proportional? Because prices move in percentage terms, not absolute ones. A share at $10 and a share at $1000 both wobble by, say, 2% on a normal day — but 2% of $1000 is fifty times more dollars. Dividing through by S_t makes the point:

— the return over the next instant is drift-plus-noise with constant coefficients. GBM is just arithmetic Brownian motion applied to returns instead of levels. (Solving it properly requires the chain rule of this calculus — Itô's lemma, the next concept — which is exactly what the box algebra above was built for.)

Here is the standard zoo. In every row, cover the right-hand columns and practise reading \mu(t,x) and \sigma(t,x) straight off the SDE — it is the first thing you do with any model you meet:

Read the design decisions like a novel. Ornstein–Uhlenbeck and Vasicek have a drift that turns negative above \bar{x} and positive below it — an elastic band pulling the process back to a long-run level, which is exactly how interest rates and volatility behave. CIR keeps Vasicek's elastic band but multiplies the noise by \sqrt{r}, so the random kicks fade away as the rate approaches zero — a built-in floor. Same grammar, different sentences: choosing a model is choosing \mu and \sigma.