Brownian Motion and Martingales: What Random Paths Can Tell Us

Brownian motion is useful in finance because it connects a tractable distributional model with a precise information structure. Its increments are simple; its paths are rough. That combination explains both the appeal of diffusion models and the need for stochastic calculus.

This reference consolidates the path properties, martingale identities and stopping-time arguments needed to interpret continuous-time financial models. It is a methods reference, with no directional market forecast.

What standard Brownian motion specifies

A standard Brownian motion $W_t$, defined on an appropriate filtered probability space, starts at zero, has continuous paths, and has independent Gaussian increments. For $0\leq s<t$,

$$W_t-W_s\sim N(0,t-s).$$

The second parameter is variance. Consequently, the standard deviation of an increment is $\sqrt{t-s}$, not $t-s$. Independence refers to nonoverlapping increments; it does not mean that the levels $W_s$ and $W_t$ are independent.

Their covariance is

$$\operatorname{Cov}(W_s,W_t)=\min(s,t).$$

For $s<t$, the later level includes the earlier level plus an independent innovation. This is why a simulated Brownian path looks persistent even though its successive increments contain no serial dependence.

For a step of length $\Delta t$, simulation uses $\Delta W=\sqrt{\Delta t}Z$, where $Z$ is standard normal. Drawing $\Delta t Z$ instead produces too little variance as the grid becomes finer. The error changes the model rather than merely reducing numerical accuracy.

Time scaling and volatility units

For a positive constant $c$, the process $W_{ct}/\sqrt c$ has the same law as standard Brownian motion. This scaling property explains square-root-of-time volatility conversion within the model.

It does not justify unconditional square-root scaling for every financial time series. Autocorrelation, volatility clustering, jumps and changing market conditions can all invalidate an independent, stationary-increment approximation. A model-based scaling rule should therefore be labelled with its assumptions.

Likewise, using annualised volatility in a daily simulation requires time in years. With 250 equally spaced trading intervals per year, $\Delta t=1/250$. Calendar-time discounting and trading-day volatility can coexist, but their conventions must be reconciled explicitly.

Continuous paths can have infinite variation

A smooth path has increments approximately proportional to the time step. Brownian increments have a typical scale proportional to its square root. As the grid becomes fine, adding absolute Brownian increments gives an unbounded total variation almost surely.

Adding squared increments behaves differently. For a partition $0=t_0<\cdots<t_n=T$, define

$$Q_n=\sum_{i=1}^n(W_{t_i}-W_{t_{i-1}})^2.$$

Its expectation is $T$. Independence and Gaussian fourth moments give

$$\operatorname{Var}(Q_n)=2\sum_{i=1}^n(t_i-t_{i-1})^2 \leq 2T\max_i(t_i-t_{i-1}).$$

In words: The uncertainty in the sum of squared changes vanishes as the largest time step shrinks.

Thus the variance tends to zero as the largest step tends to zero. This establishes mean-square, and hence probability, convergence to $T$. Along suitable nested partitions, such as the usual dyadic construction, an almost-sure version also holds.

The limiting quadratic variation is written $[W]_T=T$. The mnemonic $(dW_t)^2=dt$ records this limiting behaviour; it is not ordinary algebra with two infinitely small numbers.

For a continuous diffusion $dX_t=a_tdt+b_tdW_t$, quadratic variation is $\int_0^T b_t^2dt$. The continuous finite-variation drift contributes none. Processes with jumps require their squared jumps to be included, so the continuous-path qualification matters.

Martingales concern conditional means

An adapted, integrable process $M_t$ is a martingale if

$$\mathbb E[M_t\mid\mathcal F_s]=M_s,\qquad s\leq t.$$

Its current value is its conditional expectation at a later time. This does not imply constant realised paths or constant variance. Brownian motion is a martingale because its next increment has conditional mean zero.

Under a pricing measure, appropriately discounted tradable gains often have a martingale or local-martingale property. The relevant measure, numeraire, dividend convention and integrability conditions must be specified. “Prices are martingales” is too imprecise to be a general statement about actual market prices.

A submartingale has conditional future expectation at least as high as its current value; a supermartingale has the reverse inequality. These properties alone do not identify arbitrage. A positive expected return under the physical measure may compensate for risk, while a nonnegative local martingale can be a strict supermartingale.

Useful martingale identities

The squared process is not a martingale. Its conditional expectation is

$$\mathbb E[W_t^2\mid\mathcal F_s]=W_s^2+(t-s).$$

Subtracting elapsed time removes this drift:

$$M_t=W_t^2-t.$$

Similarly, $W_t^3$ is not a martingale merely because its unconditional mean is zero. A corrected version is

$$M_t=W_t^3-3tW_t.$$

The distinction between zero unconditional mean and the conditional martingale property is essential. A process can average to zero from the initial date while still having predictable changes given later information.

For a constant $\theta$, the exponential process

$$Z_t=\exp\left(\theta W_t-\tfrac12\theta^2t\right)$$

is a martingale. Its correction term compensates for the convexity of the exponential function. Without the correction, $\mathbb E[e^{\theta W_t}]=e^{\theta^2t/2}$, which grows with time. These exponentials support changes of measure, but random or unbounded integrands need additional conditions before the local-martingale expression can be treated as a true martingale.

Stopping rules must be observable

A stopping time $\tau$ has the property that one can determine whether it has occurred by time $t$ using information available then. First crossing of an observable barrier and exercise decisions based on current state variables are typical examples.

The time of the maximum price over a future interval is different. A price that appears to be the maximum today may be exceeded tomorrow. Selecting the eventual maximum uses future information and cannot be used as a feasible trading rule.

A stopped process is $M_{t\wedge\tau}$, where $t\wedge\tau$ denotes the smaller of $t$ and $\tau$. It follows the original process until stopping and then remains fixed. This construction is useful for barrier options, credit events and controlled trading horizons.

Optional stopping does not mean that every stopping strategy preserves expectation without qualification. Bounded stopping times provide a standard sufficient setting; more general cases require conditions such as uniform integrability or other suitable bounds. An unbounded doubling strategy can violate those conditions even when each individual gamble is fair.

A two-barrier example

Consider Brownian motion starting at zero and stopped when it first reaches $a>0$ or $-b<0$. Let $p$ be the probability of reaching the upper barrier first. Under the standard bounded-domain stopping argument,

$$0=pa-(1-p)b,\qquad p=\frac{b}{a+b}.$$

Equal barrier distances give equal hitting probabilities. An upper barrier twice as far away as the lower one has one-third probability of being reached first in this driftless model. These are illustrative mathematical probabilities, not estimates for an actual security.

Applying the corrected squared martingale gives the expected stopping time

$$\mathbb E[\tau]=ab.$$

The calculation depends on Brownian time and variance units. Introducing drift, state-dependent volatility, jumps or transaction costs changes the economic interpretation and may change the result. It also does not mean that a trader can earn a positive expected gain by choosing asymmetric take-profit and stop-loss levels: payout sizes and probabilities offset under the driftless setup.

A Brownian bridge answers a different conditional question

Suppose the beginning and end of an interval are known. Intermediate Brownian values are no longer distributed like an unconditional new path. For $0<t<T$, conditioning on $W_T=w$ gives a normal distribution with mean $(t/T)w$ and variance $t(T-t)/T$.

The conditional mean joins the endpoints, while the variance is zero at both endpoints and largest in the middle. A bridge is therefore useful for simulating plausible intermediate paths conditional on a coarse simulation or observed endpoint. It is not a deterministic straight-line interpolation.

This distinction is useful in barrier monitoring. The probability of an unobserved crossing between two safe endpoints is not zero. Under a suitable Brownian model it can be calculated or sampled using bridge arguments. For processes with jumps or different boundary behaviour, the correction needs to change with the model.

Reflection connects extrema to endpoint distributions

For standard Brownian motion and a positive barrier $a$, the reflection principle gives

$$\mathbb P\left(\max_{0\leq s\leq T}W_s\geq a\right) =2\left[1-\Phi\left(\frac a{\sqrt T}\right)\right],$$

In words: The chance of touching the upper barrier is twice the chance of finishing above it in this Brownian model.

where $\Phi$ is the standard normal cumulative distribution function. Reaching the barrier at any time is more likely than finishing above it because some paths cross and then return below.

For the purely mathematical case $a=1$, $T=1$, the crossing probability is approximately 31.73%, while the probability of the terminal value exceeding one is approximately 15.87%. The factor of two belongs to this driftless Brownian setting and should not be applied indiscriminately to asset prices, drifted processes or discrete monitoring.

The calculation demonstrates why terminal-risk and path-risk questions are different. A loss limit triggered intraday and a loss threshold checked only at the close are different events even when they use the same numerical level.

Local martingales require care with expectations

Ito calculations often produce a process with no explicit drift term. Under suitable local conditions this establishes a local martingale. To conclude that its expectation is preserved over a full horizon, one must also establish the required integrability.

Localisation stops the process before it enters increasingly extreme regions. Each stopped process can satisfy useful martingale properties even when the unrestricted limiting process does not preserve expectation. A nonnegative local martingale is a supermartingale, but it can be a strict one.

This qualification matters when an argument jumps directly from “the drift vanishes” to “the discounted expected value equals today’s price.” In standard bounded or well-behaved lognormal textbook cases, the extra conditions can be satisfied. In more general stochastic-volatility or explosive specifications, they should be checked rather than assumed.

Translating path theory into measurements

Real prices are sampled discretely, quoted on a tick grid and affected by trading frictions. Observed high-frequency returns can contain bid-ask bounce and asynchronous observations. An empirical sum of squared returns therefore need not behave exactly like the ideal quadratic variation of a frictionless diffusion.

Barrier monitoring is another gap. A simulated path can cross a barrier between grid points even when both endpoints remain below it. Increasing the number of paths without refining monitoring leaves this discretisation error untouched. Brownian-bridge corrections or appropriate fine grids may be needed.

The most useful implementation checks are direct: simulated increments should have the intended variance; disjoint increments should show no systematic correlation; quadratic-variation estimates should approach the model’s integrated variance; and corrected martingales should have stable sample means within simulation uncertainty.

Those checks do not validate Brownian motion as a complete market model. They establish that the implementation is consistent with its stated model. Economic suitability remains a separate question about tails, jumps, dependence and the use being made of the calculation.

Preserve the distinction between paths and distributions

Two independently simulated Brownian paths will not become close to one another merely because both use a finer time grid. They are different random outcomes. A strong-convergence test couples approximations to the same underlying noise so that their pathwise difference measures numerical error rather than different randomness.

A weak-convergence test compares expectations or distributions and need not align individual paths in that way. For a terminal European payoff, weak accuracy may be the primary pricing concern; for barrier timing or a pathwise hedge simulation, a stronger path treatment can matter.

Selecting the wrong comparison can produce a misleading validation result. A visibly different path is not proof of a faulty simulator, and a close terminal histogram is not proof that intermediate stopping events are correctly represented. The validation target should match the financial quantity being estimated.