LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
Realized volatility is a backward-looking measure of how dispersed returns were. Implied volatility is the volatility input that makes a pricing formula match an option's market price. They are different numbers that often move near each other. The lecture's Nvidia series from 2020 to 2025 shows an implied path, a proprietary realized path, and a plain quadratic-variation path tracking a shared level, sometimes near 40 percent and sometimes near 60 percent, without a fixed mean. A scatter of implied volatility against its own next 30 days suggests high readings tend to be followed by lower averages. A backtest that freezes every other pricing input can turn that tendency into a rising equity curve. The lecture lists the risks that curve ignores.
Why it matters
Variance needs a mean, and the mean of returns does not stay put, so an outlier in one regime is ordinary in another. The firm itself changes, leadership changes, and policy regimes change. Realized volatility is also late: you need the returns before you can measure them. Implied volatility is observable now, because the market price is observable and the formula can be inverted. Unless someone says otherwise, the lecture means a Black-Scholes inversion. Other formulas, such as a Heston implied volatility, exist and are a different number.
The idea in plain words
Start with a stock simulated at a constant 20 percent volatility. A 30-day rolling standard deviation of returns, annualized with the square root of 252, printed a path between about 0.10 and 0.20. The truth was 0.20. The estimate never saw the truth. It was still more informative than a guess that might have been 1 percent or 120 percent. Replace the constant with a Heston-style mean-reverting variance. The true variance path and the 30-day estimate move together, with the estimate lagging. A 15-day estimate is a second, quicker, noisier track. Neither is "correct" at every date. Which one is closer can change. Time-series language for the three jobs is filtering (where is volatility now), smoothing (where was it), and forecasting (where will it be). ARCH, GARCH, a mean-reverting variance, and an exponentially weighted moving average are the lecture's menu, not a fitted winner. Implied volatility answers a different question. Put the market price and the other formula inputs in, solve for volatility, and you have the number traders are embedding in the premium. Higher uncertainty tends to mean a higher premium, the way insurance costs more when the outcome is less clear. On the Nvidia history the lecture queried, implied and realized measures did not sit at absurdly different levels, such as 120 percent against 20 percent. They shared a moving level. Three of those series, compressed with principal components, put about 97 percent of their variation in the first two components. A larger set of thirty volatility features would compress too. The loadings have to stay stable or the compressed feature is a different feature next month. The trading picture plots today's implied volatility against the mean of that series across the following month.
How to use it
Keep a realized estimate and an implied series on the same chart before you invent a spread between them. If they do not even share a level, the spread is a unit error. If they do, a regression of current implied volatility on the subsequent average is a description of the past sample, not a position. The lecture's own next step is a list of ways the sandbox curve dies: time decay, a volatility move that keeps going, gap risk, margin, and correlation. An idiosyncratic spike that has faded before is a different bet from a macro spike that has not. The research order the lecture describes is the honest one. Hypothesis, simple world, real data, costs, and only then a decision about whether any average survives.
What usually goes wrong
Calling the 30-day tracker the true 20 percent. It wandered between about 10 and 20 while the truth was constant, and it lagged once the truth moved. Selling every print above 64 percent because the historical scatter slopes down. The 64.26 figure is a fit on one implied series, and the lecture says a shock that does not normalize will lose money on that rule. Treating the up-sloping sandbox curve as a live gain. Other inputs were frozen so that curve could isolate the volatility effect. Live prices do not freeze theta, the spot, or the spread.
Questions people ask
What did the constant-volatility check show?
Truth was 20 percent. A 30-day annualized realized measure stayed roughly between 10 and 20 percent. Informative, lagged, and not equal to the truth.
What crossed at about 64 percent?
The lecture's fit of Nvidia implied volatility against the following 30-day average. Above roughly 64.26 percent, that average tended to be lower. The sandbox sell threshold used in the equity curve was higher, around 80 percent.
Did principal components replace the series?
They compressed three volatility measures so that two components held about 97 percent of the variation. The lecture treats that as a feature-engineering option, useful only while the loadings stay put.