Beyond Rough Volatility: Decoupling Memory and Scaling via a Generalized Langevin Equation

Battle Count: 50

The paper is highly relevant to quantitative trading through its improved stochastic volatility modeling framework. The decoupling of memory and scaling provides a richer model class for volatility dynamics, directly applicable to option pricing, VIX modeling, and risk management. The leverage term structure findings (rejecting instantaneous correlation) have direct implications for hedging and trading strategies. However, the paper is primarily theoretical and does not provide a complete pricing framework (deferred to companion paper). The physical-measure focus means direct trading applications require the risk-neutral construction. The joint SPX-VIX calibration problem addressed is central to volatility trading desks. The empirical tests on order book data connect to market microstructure and high-frequency trading.

The framework is mathematically sophisticated, requiring expertise in non-equilibrium statistical mechanics, stochastic Volterra equations, Malliavin calculus, and Bayesian inference. The GLE with two-exponent kernels, non-quadratic potentials, and two-dimensional kernel matrices introduces significant complexity. The Markovian lift requires Prony approximation of kernels. The non-affine structure (non-quadratic potential) eliminates closed-form solutions, requiring Monte Carlo or numerical PDE solvers (e.g., Diagonal Frog scheme). The P-to-Q construction for a non-traded variance driver under market incompleteness requires indifference pricing. Bayesian kernel estimation with NUTS/HMC on high-frequency data adds computational burden. The paper itself does not provide code or a complete implementation pipeline.

The paper uses publicly available datasets (FI-2010 benchmark, WRDS TAQ data, Yahoo Finance, TrueFX, Kraken). The theoretical framework is fully specified with proofs in appendices. However, the paper is primarily theoretical with limited empirical implementation. The companion paper [Itkin, 2026a] is referenced but not yet published, containing the risk-neutral construction and joint calibration. No GitHub repository is mentioned. The Bayesian estimation uses standard NUTS/HMC samplers. Key parameters and priors are specified in detail. The paper acknowledges that validation on industry-grade data remains a future direction.