Rough volatility · teaching module · Layer 4 — calibration (Q4)

Can you back it
out of the smile?

The other route to the roughness runs through option prices. Fit the rough-Heston model to a market implied-vol smile and read off its parameters — level ξ₀, skew ρ, vol-of-vol ν, and the roughness H. The model fits beautifully. But watch what H does when you slide it.

Start here — a two-minute guided tour

hide ▲
01 · the fit

Rough-Heston fits the smile

The dots are the market smile; the teal curve is the calibrated model. It sits right on the data — and it pins the level (ξ₀), the skew (ρ) and the vol-of-vol (ν) cleanly.

02 · slide H

Now drag the roughness

The fit doesn't budge. As you change H, the vol-of-vol ν silently recalibrates to compensate — the faint dashed line shows what H alone would do, and ν cancels it. Same smile, any H.

03 · the wall

H is not identified

A single smile fixes ξ₀, ρ and ν, but leaves H on a flat, ν-degenerate direction — the calibration is ill-posed (cond ≈ 6×105). You can fit the market perfectly and still not know the roughness. Same wall as the price-history route.

Fit rough-Heston to the smile H = 0.10 · ν = 0.35 · IV-RMSE = 0.13pp
market smile best fit (ν recalibrated) same H, ν held fixed
H · roughness
0.10
the parameter you're sliding
ν* · best-fit vol-of-vol
0.35
moves to compensate H
IV-RMSE
0.13
fit error (pp) — stays tiny
cond(JᵀJ)
6.2×10⁵
how ill-posed — huge
The fit stays perfect as H moves — ν absorbs it. A single smile fixes ξ₀, ρ, ν but not H.
Set the roughness
0.10H
— the true value the market was built from
0.05 · rough0.50 · smooth
0.05
0.10
0.20
0.35
0.50
Every H you pick refits the smile just as well. Try the chips — the dashed line jumps, the fit doesn't. That flatness is the non-identifiability.

Advanced — does more data pin it down?

THE FINDING show ▼
Rough-Heston CF calibration geometry (D37–D39): cond(JᵀJ), the H~ν degeneracy, live Deribit BTC · illustrative smile, reported numbers are the project’s measured values · runs in your browser