Rough volatility · teaching module · Layer 1c — identifiability

Can you measure the
roughness back?

You can set a true roughness to build a market. But with real data, nobody hands you the answer — you have to estimate it back from what you observe. The catch: different estimators give different answers, and they disagree most exactly where markets live. Set a true H below, and watch three standard estimators try to recover it.

Start here — a two-minute guided tour

hide ▲
01 · set the truth

Start at H = 0.5

The green marker is the true roughness. At 0.5 (ordinary Brownian motion) all three estimators land near it — they agree. Roughness is identifiable here.

02 · go rough

Drag H down toward 0.1

Watch the three coloured markers spread apart and drift off the true value — two bias up, one biases down. The rougher the truth, the worse they disagree.

03 · the trap

Watch them straddle 0.5

In the rough regime the estimators can sit on both sides of Brownian — some say “rough”, some “smooth”. When they can't even agree on that, the roughness is not identifiable. That ambiguity is the finding.

Three estimators vs the truth true H = 0.50 · n = 2000 obs
0.5 · Brownian
← rougher smoother →
truth
GJR
Cont–Das
MF-DFA
0.00.250.50.751.0
TRUE H
0.50
what you set
GJR
biases up at low H
Cont–Das
biases up at low H
MF-DFA
biases down at low H
The estimators agree and land near the truth — roughness is identifiable here.
Set the true roughness
0.50true H
— ordinary Brownian motion
0.05 · rough0.95 · smooth
0.1
0.15
0.5
0.7
0.9
Sample length2000
more data → less noise, but the bias stays

Advanced — the identifiability map, and where real markets sit

THE FINDING show ▼
GJR structure-function estimator (faithful) + the audit's documented small-H biases · runs in your browser