Paper: arXiv 2406.02319

Authors: Guido Gazzani, Julien Guyon

Abstract

We consider the path-dependent volatility (PDV) model of Guyon and Lekeufack (2023), where the instantaneous volatility is a linear combination of a weighted sum of past returns and the square root of a weighted sum of past squared returns. We discuss the influence of an additional parameter that unlocks enough volatility on the upside to reproduce the implied volatility smiles of S&P 500 and VIX options. This PDV model, motivated by empirical studies, comes with computational challenges, especially in relation to VIX options pricing and calibration. We propose an accurate \emph{“pathwise”} neural network approximation of the VIX which leverages on the Markovianity of the 4-factor version of the model. The VIX is learned pathwise as a function of the Markovian factors and the model parameters. We use this approximation to tackle the joint calibration of S&P 500 and VIX options, quickly sample VIX paths, and price derivatives that jointly depend on S&P 500 and VIX. As an interesting aside, we also show that this \emph{“time-homogeneous”}, low-parametric, Markovian PDV model is able to fit the whole surface of S&P 500 implied volatilities remarkably well.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper involves advanced stochastic calculus, path-dependent volatility specifications, and novel neural network approximations for VIX pricing, indicating high mathematical complexity; it directly addresses joint calibration to SPX and VIX options with empirical backtesting and computational implementation, showing strong empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Accurate PDV Model Pricing & Calibration"] --> B["Data & Model Inputs"]
  B --> C["Methodology<br>Markovian 4-Factor PDV Model"]
  C --> D["Computational Process<br>Pathwise Neural Network VIX Approx"]
  D --> E["Joint Calibration<br>S&P 500 & VIX Options"]
  E --> F["Key Findings & Outcomes"]
  
  B -->|"S&P 500 & VIX Options Data"| D
  B -->|"Model Parameters"| C
  
  F -->|"Accurate VIX Pricing"| G1["VIX Derivatives Pricing"]
  F -->|"Fast Calibration"| G2["Efficient Joint Calibration"]
  F -->|"Surface Fit"| G3["S&P 500 Implied Volatility Smile"]
  F -->|"Markovian Structure"| G4["Time-Homogeneous Model"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a pathwise neural network approximation of the VIX within the 4‑factor PDV model and an automatic, robust joint calibration procedure for SPX and VIX options.

Source sentence
We propose an accurate pathwise neural network approximation of the VIX which leverages on the Markovianity of the 4-factor version of the model.

Assumes 5

  • Interest rate and dividend yield curves are deterministic. model
    Source sentence
    Throughout this article, (rt)t≥0 and (qt)t≥0 denote the interest rate and the dividend yield curves, respectively, and are assumed deterministic.
  • Asset price dynamics are under the unique risk‑neutral measure Q. model
    Source sentence
    The price (St)t≥0 of the asset is modeled under the unique risk‑neutral
  • VIX futures and options are priced via the Black formula using model VIX futures. model
    Source sentence
    the model VIX implied volatilities are then of course computed by inverting the Black formula using the corresponding model VIX future.
  • Market data (SPX and VIX options) are representative of the calibration date. data
    Source sentence
    The market data used in our numerical experiments was provided by BNP Paribas via the Chair Futures of Quantitative Finance.
Show 1 more
  • Bid‑ask spreads are used as weights in the calibration loss. execution
    Source sentence
    The weights (γVIX T,K )T,K are often taken to be the inverse of the market bid‑ask spreads, which gives more weight to those options where the bid‑ask spread is tighter.

Data

  • SPX and VIX options market data · BNP Paribas (Chair Futures of Quantitative Finance) options commercial
    S&P 500 and VIX contracts
    Source sentence
    The market data used in our numerical experiments was provided by BNP Paribas via the Chair Futures of Quantitative Finance.

Costs

Not applicable The paper reports no trading or execution result; the cost the extractor found is a model parameter.

The paper-text extraction found a cost parameter inside a model (Calibration loss weights use the inverse of market bid‑ask spreads, but trading results do not explicitly incorporate transaction costs.); the classifier found no reported trading result, so costs are not applicable.

Finds 4

  • The pathwise neural network approximation enables fast VIX simulation and joint calibration, reducing calibration time to about 8 minutes. pricing-calibration
    Source sentence
    The joint calibration takes around 8 minutes to complete.
  • Joint calibration of SPX and VIX monthly options achieves very good accuracy, reproducing SPX smiles and VIX futures/volatility surfaces. pricing-calibration
    Source sentence
    In Figures 10 and 11 we observe that we are able to calibrate jointly the first two monthly smiles of the SPX
  • Calibrated β‑parameters are remarkably stable over time, while kernel parameters show some variability but produce similar two‑exponential kernels. statistical
    Source sentence
    The β parameters (β0, β1, β2, β1,2) (Figure 12a) are remarkably stable over time. The kernel parameters
  • The neural network VIX approximation is accurate enough for trading purposes when compared to nested Monte Carlo estimates. performance
    Source sentence
    Figures 10 and 11 show that for the jointly calibrated parameters our neural network approximation of the VIX is accurate enough for trading purposes.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 29 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 8 papers in the archive (8 by independent authors): Forecasting implied volatility surface with generative … (uses its method or data)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.