Local Martingales and Semimartingales

Itô's formula, stochastic differential equations, and arbitrage-free pricing are all built on top of the martingale, but the strict definition of a martingale — a genuinely fair game at every point in time, forever — turns out to be too narrow for what actually shows up when you integrate against Brownian motion. The fix isn't to abandon the martingale idea; it's to weaken it just enough to cover every process that behaves like one locally, even if something subtle breaks at the boundary.

The analogy: fair on every leg of the trip, but not necessarily the whole journey

Imagine a road trip broken into legs by rest stops, where on every single leg you're driving at a genuinely random, unpredictable pace with no net tendency to speed up or slow down — each leg, by itself, is a fair coin flip on average distance covered. A local martingale is a process built exactly that way: fair on every leg (defined via a sequence of stopping times, the "rest stops," that eventually cover the whole timeline), but the legs can behave badly enough near their edges that stitching them together doesn't guarantee the whole trip averages out. Most of the time it does. Occasionally it doesn't, and that gap is precisely where some of the trickiest failures in stochastic calculus live.

Writing it down

A process $M_t$ is a local martingale if there exists a sequence of stopping times $\tau_n \to \infty$ such that the stopped process $M_{t \wedge \tau_n}$ is a genuine martingale for every $n$. In words: you can't necessarily prove $M_t$ is fair all the way out to infinity, but you can find an escalating sequence of cutoff times, each pushing further out, such that "the process, halted at that cutoff" is provably fair. Every true martingale is automatically a local martingale (just take $\tau_n = n$), but the reverse can fail — a local martingale can have $E[M_t] \ne E[M_0]$ for some finite $t$, something a genuine martingale can never do.

A semimartingale is the still broader, practically important category: any process that can be written as

$$X_t = X_0 + M_t + A_t,$$

where $M_t$ is a local martingale (the "unpredictable, no-drift" part) and $A_t$ is a process of finite variation (a "smoothly moving, drift or trend" part). In words: a semimartingale decomposes cleanly into noise plus trend. This is exactly the class of processes for which the Itô integral is defined and Itô's formula holds — asset prices and interest rates are assumed to be semimartingales for precisely this reason.

Geometric Brownian motion, sampled above, is the standard example of a semimartingale: it decomposes into a driftless local martingale piece (the Brownian noise, scaled by the price) plus a deterministic drift piece growing at rate $\mu$. Set the drift to a nonzero value and watch the trend component visibly separate from the noise around it — that separation is the $M_t + A_t$ decomposition, made visible.

Worked example 1: a genuine local martingale that isn't a martingale

The classic textbook example: let $W_t$ be 3-dimensional Brownian motion started away from the origin, and define $M_t = 1/|W_t|$ (the reciprocal of its distance from the origin). This process is a local martingale — using stopping times $\tau_n$ that halt the process before $|W_t|$ gets too close to zero, the stopped process is a genuine martingale each time. But it is not a true martingale: $E[M_t]$ actually decreases strictly over time (it's a classical fact that $1/|W_t|$ is a strict local martingale in three or more dimensions), so $E[M_t] < E[M_0]$ for $t>0$ — a violation that would be impossible for a real martingale, where $E[M_t] = E[M_0]$ always. The gap comes entirely from the small but nonzero chance of $|W_t|$ passing arbitrarily close to zero before any finite cutoff, which the localizing stopping times can delay but never fully rule out.

Worked example 2: decomposing a strategy's P&L

A trend-following strategy's cumulative P&L $X_t$ is modeled as $X_t = A_t + M_t$ where $A_t = 0.02t$ (a steady drift of 2 percent P&L accumulation per unit time from the trend edge) and $M_t$ is a driftless martingale capturing day-to-day noise around that trend, with, say, daily standard deviation 1. After 100 days, the trend component contributes exactly $0.02 \times 100 = 2.0$ (in P&L units) with certainty, while the noise component $M_{100}$ has mean 0 but a standard deviation of $\sqrt{100} \times 1 = 10$ (since a driftless martingale with unit-variance independent increments accumulates variance linearly). So the decomposition tells you precisely how much of the day-100 P&L to attribute to genuine edge (2.0, guaranteed) versus noise (mean 0, but with a standard deviation five times larger) — exactly the split a Sharpe-ratio-style analysis is trying to estimate.

What this means in practice

• Itô's formula and stochastic differential equations are stated for semimartingales, not for arbitrary random processes — this is the formal reason SDE-based models (Black-Scholes, Heston, short-rate models) are built the way they are.
• Risk-neutral pricing relies on turning an asset's price process into a true martingale (under a suitable measure) — checking that a candidate discounted price process is genuinely a martingale, not merely a local martingale, is what separates a valid no-arbitrage model from a subtly broken one (see Novikov's Condition).
• P&L attribution — splitting realized returns into "edge" (the $A_t$ trend term) versus "noise" (the $M_t$ martingale term) — is exactly the semimartingale decomposition applied to a trading strategy.