LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
"Time in the market beats timing the market" is the proverb this lecture ties to non-stationarity. With a fair six-sided die, you do not need to guess the face. The best guess is 3.5, and trading only when you can buy below it or sell above it compounds in the simulation. Replace the die with a ten-sided one and keep the old quote, and the same rule gives the gains back. The hard part is that a roll of five could have come from either die, so you do not know when to update.
Why it matters
Probability is a framework for uncertainty, not a preview. A die face, or whether a market is hit hard next month or in five years, is unknown, and the true probabilities are unknown too. Courses often hand you a Bloomberg pull and a parametric model and skip what happens when the source changes. A time-series course, the lecture says, usually assumes stationarity on day one so the semester can proceed. That is a classroom convenience.
You do not need to be right on each outcome. You need an average that is sound often enough for the book you are running, whether that book is making a market, trading, or investing. The die makes the ideal case. Each face from one to six has probability one-sixth, and nothing else has any mass. The mean-squared-error minimizer is 3.5. Quote around it, trade only the favorable side, and the lecture's equity curve rises while every individual face stays random.
The idea in plain words
Drop the known mass function and keep the same die, and fifty rolls produce a working mean around 3.66. Trading around 3.66 instead of 3.5 does not harvest the true edge, and the equity curve in that frame struggles. More rolls pull the average to 3.5 by the law of large numbers, and the curve turns up. The lecture's classroom moral is that a long enough sample can replace the mass function when the die is fixed. The same sentence is then withdrawn for markets.
Non-stationarity is not a series you can transform until a unit-root test passes. A return is the lecture's composite of a huge number of decisions, treated as uncertain because you cannot model the inputs. A stationarity test on one sample can look calm and then meet a regime change. What the lecture says you can look for is stability: a strategy whose behavior does not shatter when parameters are perturbed or when the regime shifts, or that can break and be rebuilt. You cannot turn the real world into a stationary random variable by a formula.
How to use it
The SPY illustration is the classroom exercise the lecture wants retired. Fit a normal to a histogram of returns and read the probability of a three-standard-deviation left-tail event. The normal sits well inside a kernel-density estimate of the same sample. The lecture names excess kurtosis, fat tails on both sides, as the shape, and names non-stationarity as the cause. The histogram has compressed time away. A low-volatility regime, a middle one, and a high-volatility one can all contribute bars to one picture, so the tails are not a single static law.
What usually goes wrong
Trading a fifty-roll mean as if it were 3.5 spends the edge before the law of large numbers has anything to say, and that law was only available because the die never changed. Applying it to SPY returns, or to a timing rule, imports a theorem whose hypothesis the lecture says is false.
Questions people ask
What is the best guess for a fair die?
3.5, the mean-squared-error minimizer. The trading rule buys below that value and sells above it. With the true die, the lecture's equity curve rises.
Why do fifty rolls fail in the picture?
The sample mean is about 3.66, not 3.5, so the quotes are centered on the wrong value. More rolls of the same die pull the mean back and the curve recovers.
Why is a roll of five ambiguous?
It is possible on both the six-sided and the ten-sided die. A ten would reveal the switch. A five does not, which is why the lecture says you cannot see the regime change on the data alone.