Skip to content
The Chat Archive
gemini-2.5-pro-preview-05-06· Google LMArena Human Preference 140K

describe me the rough volatility methodology used in quantitative finance, give me all…

  1. User

    describe me the rough volatility methodology used in quantitative finance, give me all mathematical details required

  2. Assistant

    Okay, let's dive into the rough volatility methodology. It's a fascinating and relatively recent development in quantitative finance that has gained significant traction due to its ability to better capture certain stylized facts of financial markets, particularly the behavior of implied volatility smiles and skews.

    1. Context: Why Move Beyond Traditional Models?

    Traditional volatility models include:

    • Black-Scholes (Constant Volatility): dS/S = μdt + σdW. Assumes σ is constant. Fails to explain the volatility smile/skew.
    • Local Volatility (Dupire): σ_LV(S,t). Volatility is a deterministic function of stock price and time. Can fit any smile but has poor dynamics (e.g., sticky smile).
    • Stochastic Volatility (e.g., Heston, SABR):
      • Heston: dS/S = μdt + √v_t dW1_t dv_t = κ(θ - v_t)dt + ξ√v_t dW2_t dW1_t dW2_t = ρdt These models introduce a separate stochastic process for variance (v_t), allowing for dynamic smiles. However, they often struggle to simultaneously match the term structure of ATM volatility and the skew for short-dated options without ad-hoc parameterizations or more factors.

    Empirical Observations Leading to Rough Volatility:

    • Term Structure of ATM Implied Volatility: Often exhibits a power-law decay σ_ATM(T) ≈ a + b T^(H-0.5) for small T, where H is empirically found to be very small (around 0.1).
    • Explosive Behavior of Skew for Short Maturities: The skew |∂σ_imp/∂log(K)| near the money K=S_0 blows up very fast as T → 0.
    • Realized Volatility Time Series: Log-volatility appears to have long memory but is "rougher" (more jagged) than a standard Brownian motion.

    2. What is Rough Volatility? The Hurst Parameter (H)

    The core idea of rough volatility is that the (log-)volatility process is driven by a Fractional Brownian Motion (fBm) with a Hurst parameter H < 0.5.

    • Standard Brownian Motion (BM): Corresponds to H = 0.5. Increments are independent. Paths are continuous but nowhere differentiable, with Hölder continuity of order α for any α < 0.5.
    • Fractional Brownian Motion (fBm): B_H(t)
      • Gaussian process, starts at 0, E[B_H(t)] = 0.
      • Covariance: E[B_H(t)B_H(s)] = 0.5 ( |t|^(2H) + |s|^(2H) - |t-s|^(2H) ).
      • Increments are stationary but not independent for H ≠ 0.5.
        • H > 0.5: Positive correlation in increments (persistence). Smoother paths than BM.
        • H < 0.5: Negative correlation in increments (anti-persistence). Rougher paths than BM. Paths are Hölder continuous of order α for any α < H.
      • Self-similarity: B_H(at) is equal in law to a^H B_H(t).

    Empirically, for log-volatility time series, H is found to be around 0.05 - 0.15, hence "rough."

    3. Mathematical Details of Fractional Brownian Motion

    fBm is not a semi-martingale for H ≠ 0.5, so standard Itô calculus cannot be directly applied to dB_H(t). Instead, stochastic integrals with respect to fBm are often defined in the Wick-Itô sense or using fractional calculus.

    A common representation for fBm (Mandelbrot-Van Ness): B_H(t) = (1/Γ(H+0.5)) * [ ∫(-∞,0) ((t-s)^(H-0.5) - (-s)^(H-0.5)) dW(s) + ∫(0,t) (t-s)^(H-0.5) dW(s) ] where W(s) is a standard Brownian motion and Γ is the Gamma function. This representation highlights the long memory aspect, as B_H(t) depends on the entire past of W(s).

    For modeling, it's often more convenient to work with Volterra processes: A Volterra process X_t is of the form: X_t = ∫_0^t K(t-s) dW_s where K(·) is a deterministic kernel. If K(t-s) ∝ (t-s)^(H-0.5), then X_t behaves like a Riemann-Liouville fractional integral of a Brownian motion, which is closely related to fBm.

    4. The Rough Volatility Model (e.g., rBergomi, rHeston)

    The general framework is a stochastic volatility model where the log-volatility process is driven by fBm or a related Volterra process.

    A. Conceptual Model for Instantaneous (Log-)Variance: Let v_t be the instantaneous variance. We model its logarithm: log v_t = Y_t Y_t = Y_0 + η ∫_0^t (t-s)^(H-0.5) dZ_s^v + drift_terms (Volterra form) where Z_s^v is a standard Brownian motion, η is the "vol-of-vol" parameter. H is the Hurst parameter. The (t-s)^(H-0.5) is the power-law kernel. The asset price process: dS_t/S_t = r dt + √v_t dZ_s^s dZ_s^v dZ_s^s = ρ ds (Correlation between price and volatility shocks)

    B. The rBergomi Model (Rough Bergomi): This is a popular specification, extending Bergomi's variance swap model. It models the forward variance curve. Let ξ_t(u) = E_t[σ_u^2] be the forward variance at time t for future date u > t. The instantaneous variance is σ_t^2 = ξ_t(t). The rBergomi model often specifies the dynamics of σ_t (or log σ_t) directly. A common formulation for the spot variance process v_t = σ_t^2: dS_t/S_t = √v_t dW_t v_t = ξ_0(t) * exp( η W_t^H - (1/2)η^2 t^(2H) ) (Simplified version, often appears in literature) More precisely, the log of the spot volatility log σ_t is modelled as a fractional Ornstein-Uhlenbeck process or directly using the Volterra representation driven by a BM Z_t: log σ_t = log σ_0 + η ∫_0^t (t-s)^(H-0.5) dZ_s - (1/2)η^2 ∫_0^t (t-s)^(2H-1) ds (The second term is the Itô correction analogue for fractional integrals. For H<0.5, 2H-1 is negative, so this integral needs careful definition/regularization, often t^(2H) is used as an approximation if log σ_0 is a constant or if the integral starts from -∞).

    A more canonical representation (Gatheral, Jaisson, Rosenbaum): The instantaneous variance v_t is given by: v_t = v_0 exp( η W_t^H - c η^2 t^{2H} ) where W_t^H is a specific type of fBm (e.g. Riemann-Liouville fBm). Or, more generally, for the log-forward variance: log ξ_0(T) = log v_0(T) (initial forward variance curve) The dynamics of the log-spot variance: Y_t = log √v_t dY_t = η K_H(t) dW_t^v - λ Y_t dt (Fractional Ornstein-Uhlenbeck, where K_H is the fractional kernel, or simply η dW_t^H if no mean reversion on Y_t). If log √v_t = X_t, then: X_t = X_0 + η ∫_0^t K(t-s) dW_s^{vol} - \frac{1}{2} η^2 ∫_0^t K(t-s)^2 ds where K(u) = u^{H-1/2} (or some normalization constant times this). dS_t/S_t = (r-q)dt + e^{X_t} dW_t^{spot} dW_t^{vol} dW_t^{spot} = ρ dt

    Parameters to calibrate:

    • H: Hurst parameter (often estimated from historical data first).
    • η: Volatility of volatility.
    • ρ: Correlation.
    • v_0(T) or ξ_0(T): Initial forward variance curve (calibrated to ATM option prices or variance swaps).

    5. Estimation of the Hurst Parameter (H)

    H is often estimated from high-frequency historical data before calibrating other parameters to option prices.

    1. Construct Realized Volatility: For a given day d, using intraday log-returns r_i = log(S_i) - log(S_{i-1}) at a high frequency (e.g., 5-minute): Realized Variance RV_d = Σ r_i^2. Realized Volatility σ_d = √RV_d.
    2. Analyze Log-Volatility Increments: Consider the time series of log σ_d. The scaling property of fBm implies for X_t = log σ_t: E[|X_{t+δ} - X_t|^q] ∝ δ^(qH) Taking logs: log E[|X_{t+δ} - X_t|^q] ≈ const + qH log δ.
    3. Regression: Calculate m_q(δ) = (1/(N-δ)) Σ_{i=1}^{N-δ} |log σ_{i+δ} - log σ_i|^q for various lags δ and orders q (e.g., q=1, 2). Plot log m_q(δ) against log δ. The slope of this log-log plot gives an estimate of qH. Divide by q to get H. Empirical estimates are typically in the range 0.05 < H < 0.15.

    6. Calibration and Pricing

    • Calibration: Given H (from historical data or also as a calibration parameter), η, ρ, and the initial forward variance curve ξ_0(T) are calibrated to market option prices.
      • Minimize Σ_i w_i (C_market_i - C_model_i(H, η, ρ, ξ_0))^2.
    • Pricing:
      • Monte Carlo Simulation: This is the most common method.
        • Simulating fBm paths: Can be done using Cholesky decomposition of the covariance matrix (slow for many steps), or FFT-based methods (e.g., Davies-Harte algorithm, or for Volterra forms, using FFT for the convolution ∫K dW).
        • Hybrid schemes (e.g., by Bennedsen, Lunde, Pakkanen) provide efficient approximations for simulating ∫ (t-s)^(H-0.5) dW_s.
        • Euler-Maruyama or Milstein type schemes for the SDE components (asset price).
      • Asymptotic Expansions: For short maturities, powerful asymptotic formulas for option prices and implied volatility can be derived (e.g., by Gatheral, Forde, Jacquier, Alòs, Vives). These often show the T^(H-0.5) behavior for ATM vol and skew explosions.
      • PDEs: Due to the non-Markovian nature of fBm, standard PDE methods are difficult. However, by lifting the process to an infinite-dimensional Markovian state (related to the entire past of W), some PDE approaches exist but are complex.

    7. Advantages of Rough Volatility Models

    • Excellent Fit to Implied Volatility Surface: Can capture the term structure of ATM vol and the skew of short-dated options very well with fewer parameters than multi-factor stochastic vol models.
    • Consistency: More consistent between historical volatility dynamics and implied volatility. H from historical data works well for pricing.
    • Realistic Vol-of-Vol: The roughness naturally generates high vol-of-vol, which is observed empirically.
    • Parsimony: A small number of parameters (H, η, ρ) plus the initial forward curve can describe complex market phenomena.

    8. Challenges

    • Computational Cost: Simulating fBm can be more intensive than standard BM, especially for long paths or many simulations.
    • Non-Markovian Nature: Makes many standard techniques (like simple PDE solvers, dynamic programming for American options) harder to apply directly.
    • Mathematical Sophistication: Requires understanding of fractional calculus and advanced stochastic processes.
    • Hedging: Hedging strategies are more complex due to the path-dependent nature of volatility.

    Mathematical Nuts and Bolts Summary:

    1. Core Idea: log σ_t or log v_t is driven by a process related to fBm with H < 0.5. dX_t = ... + η K_H(t-s) dW_s where K_H(u) = u^(H-0.5).
    2. fBm Properties: E[B_H(t)B_H(s)] = 0.5 ( |t|^(2H) + |s|^(2H) - |t-s|^(2H) ). Non-semimartingale for H ≠ 0.5.
    3. H Estimation: Log-log plot of E[|log σ_{t+δ} - log σ_t|^q] vs δ. Slope is qH.
    4. Pricing/Calibration: Primarily Monte Carlo, leveraging specific fBm simulation algorithms (Davies-Harte, FFT for Volterra convolution, hybrid schemes). Parameters: H, η, ρ, initial forward variance curve ξ_0(T).
    5. Key Kernel: (t-s)^(H-0.5) for H ∈ (0, 0.5). This kernel has a singularity at s=t, which contributes to the roughness.

    Rough volatility represents a significant step forward in financial modeling, offering a more realistic and parsimonious description of volatility dynamics. While mathematically and computationally challenging, its empirical success has made it an active area of research and application.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.