LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
Arithmetic Brownian motion adds a drift and a shock. Geometric Brownian motion multiplies both by the current price, so the process compounds and cannot cross below zero. The log of the geometric process is an arithmetic process whose drift is the original drift minus half the variance. Both are models for building expectations, not machines that pick the path you will get.
Why it matters
Derivative pricing needs a random engine and a choice of dynamics. The lecture compares the two standard choices by simulation before it solves them. On the arithmetic side, the expected path is a straight line, like simple interest that is not reinvested. On the geometric side, the expected path curves. A few extreme paths pull the arithmetic average up. The geometric mean, the compounded path, lies below it. The gap is volatility drag. Outliers dominate the average you see near the end of a large simulation.
Each simulated step is one draw from a random variable. A stochastic process is a family of those variables indexed by time. For the arithmetic model, the level at a future date is normal. For the geometric model, it is lognormal, which keeps the price positive. That positivity is useful for equities. It is a bad fit when the underlying can print a negative price, which the lecture illustrates with crude oil when storage costs overwhelmed the market around the pandemic period. In that setting they would rather use the arithmetic process.
The idea in plain words
The arithmetic change is drift times a small time step, plus volatility times a Brownian shock. Drift is a constant you choose. The shock is not. Integrate to a horizon T and the level equals today's price, plus drift times T, plus volatility times a normal increment with mean zero and variance T. The future level is therefore normal. Its mean is today's price plus drift times T, and its variance is volatility squared times T. More simulated paths, inside a fixed model, line up with that normal curve. A real firm is not a fixed generator, so twenty years of prices need not settle to one mean.
The geometric version multiplies both the drift and the shock by the current price, so a plain integral cannot isolate the level. Take the log and apply Ito's lemma. Squared time steps drop out, and a squared Brownian increment counts as a time step. The log then moves with a drift equal to the original drift minus half the variance. Exponentiate and the price at T is today's price times an exponential of that reduced drift times T, plus volatility times the Brownian increment. The distribution is lognormal. The mean of the log is the log of today's price plus the reduced drift times T, and the variance of the log is volatility squared times T.
How to use it
Pick the arithmetic process when the level can be negative or when you truly want additive shocks. Pick the geometric process when the variable is a positive price that should scale with itself. When you read a geometric simulation, do not confuse the average of the paths with the compounded growth. The average includes the right tail. The growth an investor compounds is the lower curve, the one with half the variance removed from the drift. If you estimate drift and volatility from history, remember the lecture's limit: those inputs are choices or estimates, and the model only converges to its own distribution while the distribution stays put.
Increments differ in a way that matters for simulation. An arithmetic step does not look at the previous price. A geometric step scales with it, so later increments can be more extreme simply because the level is higher. That is the source of the outliers, not a separate jump process.
What usually goes wrong
Using a geometric model for a price that can go negative. Using an arithmetic model for a stock and then being surprised by negative simulated prices. Reporting the average of geometric paths as if it were the growth rate. And treating classroom convergence as a property of live markets. The lecture's endpoint is the pair of solutions and the drag term, not a claim that either process is the real data generator.
Questions people ask
Why does half the variance show up in the geometric drift?
It is the second-order Ito term from the logarithm. Once the squared shock is replaced by a time step, that term subtracts half the variance from the drift of the log. Exponentiating puts the curvature back. The arithmetic process never takes that logarithm, so its mean has no such subtraction.
Can geometric Brownian motion produce the extreme paths in a picture?
Yes. Because steps scale with the current level, a run of gains makes later gains larger. Those paths pull the arithmetic average above the geometric mean. They are still draws from the same lognormal solution. They are not a different model.
What are these solutions for?
They give the distribution of the level at a chosen time, so you can form an expectation and, later, a derivative price. The lecture is explicit that the purpose is not to predict which path arrives. Randomness stays random. The mean is the model's best guess under the parameters you supplied.