LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
Rough volatility is the claim that changes in log volatility scale like a fractional process with a Hurst exponent near 0.1, not like ordinary Brownian motion at 0.5. The Hurst exponent is a roughness parameter: lower means a more jagged path. If simulated volatility is too smooth, volatility surfaces and exotic prices extrapolated from those paths will be the wrong objects, even when the algebra looks polished.
Why it matters
Volatility is latent. You never see a population parameter for it. You estimate a local mean of returns, then a second moment around that mean, and you call the result a proxy. The proxy depends on choices: bar size, and how many bars sit inside the window. Short windows jump around. Long windows look smooth because variance is an average squared departure from an expectation. If everything is a surprise, nothing is a surprise, so a dispersion process that runs off to infinity is a bad description of a market that keeps resetting what “normal” means.
The rough-volatility literature, especially Jim Gatheral’s work and the rough Bergomi model, grew up around a practical pain: fitting the S&P 500 and the VIX at the same time. The lecture is a guide to that idea, not a replacement for the papers. Its own regressions use a smaller sample, on the order of a few hundred days of S&P 500 data from 5-minute bars, so the slopes are teaching numbers, not a replica of the published tables.
The idea in plain words
Let the object be the log of a volatility proxy. A structure function of order q and lag delta is the expected absolute increment, raised to q, across that lag. You estimate it by averaging those powered increments. One picture fixes q at 2 and the lag at eight 5-minute bars, which is 40 minutes, draws the histogram, and takes the mean.
The rough-volatility claim says this object scales like a power of the lag. Take logs and the power becomes a slope in an ordinary regression. At q equal to 1, the lecture’s S&P sample produced a slope of 0.1555, near 0.15 rather than the Brownian benchmark of 0.5. Stretch the horizon by four. Brownian scaling multiplies a typical increment by two. A slope near 0.15 multiplies it by about 1.21. Fractional paths with Hurst 0.1 look much rougher than paths with Hurst 0.5, and more like that realized-volatility proxy.
Repeating the regression at orders including 0.5, 1, 1.5, 2, and 3 gave slopes that lined up roughly linearly in q, which is what a simple fractional process predicts. The sample is only a few hundred days, shorter than the papers, so the decimals are a classroom, not a replica.
Raw fractional Brownian motion is still a poor volatility model. It does not revert, so the variance envelope can wander without a bound. The repair in this literature is a fractional Ornstein-Uhlenbeck process: mean reversion, plus a rough driver. The memory is long, so practical calibration eventually wants a Markovian lift that replaces infinite memory with a finite state. The lecture points at that step and stops.
How to use it
Build the proxy on purpose. Pick a bar size that is not dominated by microstructure noise; the lecture uses 5-minute bars. Estimate the structure function over a range of lags and a range of q. Regress in log space. Compare the slope with 0.5 before you decide that a smooth Brownian variance is enough.
What usually goes wrong
Treating one window length as “the” volatility. Calling fat, jagged paths a visual impression without a structure function. Using Hurst 0.5 because the textbook Brownian motion is familiar. And forgetting mean reversion, so the variance model is free to explode. All models are wrong; the useful ones at least resemble the scaling you measured.
Questions people ask
What number in the lecture pointed at roughness?
A log-log regression of the order-1 structure function on that S&P sample had slope 0.1555, far from the Brownian benchmark of 0.5.
Why not model volatility as fractional Brownian motion directly?
That process does not revert to a mean, so its variance envelope grows. A fractional Ornstein-Uhlenbeck process keeps roughness and pulls volatility back.
Does a small sample prove the academic papers?
No. The lecture’s history is shorter than the papers it is illustrating. The point is to see power-law scaling and what it does to simulated paths.