Rough Volatility Hurst Calibrator

License Hub Type Fractional Benchmarks Interactive Studio

πŸ“Œ Executive Summary

At1uri/rough-volatility-hurst-calibrator is a deep neural calibrator designed to extract the Hurst exponent $H \in (0, 0.50]$, volatility of volatility $\nu$, and spot-volatility correlation $\rho$ from raw forward variance paths and implied volatility smiles under the Rough Bergomi (rBergomi) and Rough Heston fractional SDE paradigms.

While standard numerical calibration of rough fractional Volterra SDEs requires heavy Monte Carlo simulations taking tens of minutes per maturity slice, this neural network calibrates all three key structural parameters in under $1.5\text{ ms}$ with a mean absolute error on $H$ of $\text{MAE} < 0.02$.


πŸ”¬ Mathematical Framework: Rough Volatility & Volterra SDEs

Under the rough Bergomi framework (Bayer, Friz & Gatheral, 2016), the instantaneous variance $v_t$ is modeled as a non-Markovian log-normal Volterra process:

vt=ΞΎ0(t)exp⁑(Ξ½2H∫0t(tβˆ’s)Hβˆ’1/2dWsβˆ’12Ξ½2t2H)v_t = \xi_0(t) \exp\left( \nu \sqrt{2H} \int_0^t (t - s)^{H - 1/2} dW_s - \frac{1}{2} \nu^2 t^{2H} \right)

where:

  • $\xi_0(t) = \mathbb{E}[v_t]$ denotes the initial forward variance curve extracted from market variance swaps.
  • $H \in (0, 1/2)$ is the Hurst exponent, controlling the HΓΆlder regularity $\alpha = H - \epsilon$.
  • $W_s$ is a standard Brownian motion correlated with the price-driving Brownian motion $B_s$: $d\langle W, B \rangle_t = \rho dt$.

The ATM Skew Singularity

Because $H < 1/2$, the kernel $K(t - s) = (t - s)^{H - 1/2}$ is singular at $s = t$. This singularity resolves the classic financial puzzle where market option prices display steep at-the-money (ATM) volatility skews for short maturities:

ψ(T)=βˆ‚ΟƒBS(k,T)βˆ‚k∣k=0β‰ˆΟΞ½2(H+1/2)THβˆ’1/2\psi(T) = \left. \frac{\partial \sigma_{\text{BS}}(k, T)}{\partial k} \right|_{k=0} \approx \frac{\rho \nu}{2(H + 1/2)} T^{H - 1/2}

As $T \to 0$, $\psi(T)$ explodes at rate $T^{H - 1/2} \approx T^{-0.40}$ for $H = 0.10$, matching empirical index options identically.


⚑ Calibration Performance & Benchmarks

Evaluated on $N = 200$ test benchmarks from At1uri/rough-volatility-fractional-benchmarks:

Calibrated Parameter Ground Truth Range Neural MAE Classical Variogram MAE Calibration Speed
Hurst Exponent $H$ $[0.03, 0.50]$ 0.0184 0.0421 1.2 ms
Vol-of-Vol $\nu$ $[1.20, 2.80]$ 0.0650 0.2840 1.2 ms
Correlation $\rho$ $[-0.90, -0.65]$ 0.0210 0.0950 1.2 ms

πŸš€ Quickstart & Inference

from rough_engine import RoughVolatilityHurstCalibrator
import numpy as np

calibrator = RoughVolatilityHurstCalibrator()

# Simulate a synthetic rough variance path (H = 0.10)
sim = calibrator.simulate_rbergomi(H=0.10, nu=2.0, rho=-0.75, n_steps=50)
v_path = np.array(sim["variance_path"])

# Calibrate parameters in 1.2 ms
res = calibrator.calibrate(v_path)
print(f"Calibrated Hurst H: {res['calibrated_H']} (Target: 0.10)")
print(f"Calibrated Vol-of-Vol nu: {res['calibrated_nu']}")
print(f"Rough Volatility Regime:  {res['regime']}")

πŸ”— Integrated Ecosystem


πŸ“œ Citation

@misc{atluri2026roughcalib,
  author       = {Atluri, Ramprasad},
  title        = {Rough Volatility Hurst Calibrator},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/At1uri/rough-volatility-hurst-calibrator}}
}
Downloads last month
24
Safetensors
Model size
112k params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Space using At1uri/rough-volatility-hurst-calibrator 1

Collection including At1uri/rough-volatility-hurst-calibrator

Evaluation results

  • Hurst MAE on Rough Volatility Fractional Benchmarks
    self-reported
    0.018
  • Vol-of-Vol MSE on Rough Volatility Fractional Benchmarks
    self-reported
    0.042