There is no options-market forecast to trade from Itkin's paper. He acknowledges the gap: "Convergence to models with singular kernels, statistical estimation and calibration to market data are not treated here". His closing case for the work is that the results "supply the probabilistic and pricing foundation for those investigations, while their empirical adequacy and computational performance remain to be established". The paper establishes a sufficient growth condition for the general model and an exact explosion boundary for a one-mode family. Rebuilding it could be worthwhile, so long as the choices left to the implementer are made explicit.
In local stochastic volatility, a stochastic variance is multiplied by L(t,S). Calibration chooses L so that L²·E[F²(Y)|S=K] matches Dupire local variance. The expectation depends on the law produced by the model itself, making calibration a McKean, Vlasov fixed point. Rough-volatility LSV also calls for an SPDE or a rough PDE.
Itkin instead gives the variance driver a generalized Langevin equation (GLE), which carries past disturbances forward through a memory kernel. The kernel c t^(−β0)(1+κt)^(β0−β∞) separates short-lag roughness from long-lag memory. Approximate it with a finite exponential sum and the state becomes a finite Markov diffusion; an ordinary forward equation then supplies the Dupire term. Past returns reach variance through leverage memory, with a response that grows stronger as variance rises. Joint SPX/VIX pricing under one measure is the trading application. There is no market sample or sample period here, only theorems and Monte Carlo on simulated paths.
Our own run adapts the leverage mechanism to end-of-day SPY variance forecasting. It leaves the paper's utility-selected measure and SPX-VIX calibration problem untested. Bid-ask calibration cannot be reproduced from end-of-day data.
How the memory becomes a state
Three forces act on the GLE driver Y: a restoring force from the potential U, friction that remembers earlier states, and colored noise, a random force correlated across time. Exponentially weighted past log returns form the leverage states. Each has its own fixed decay rate, while its amplitude varies with current variance. The example sets that amplitude to c0(1 + η tanh((y − y⋆)/s⋆)). A positive η makes the return response stronger in stress and weaker in calm.
The intrinsic kernel c t^(−β0)(1 + κt)^(β0 − β∞) has separate exponents at short and long lags. Itkin replaces it with a finite sum of exponentials. Every exact result applies to that finite sum. A Hörmander bracket argument gives a smooth density for a reduced state when the physical drift uses the friction and noise states only through their difference, |ρ| < 1, the leverage rates are distinct, and the coefficients are C∞. The Dupire expectation E[F²(Y)|S = K] becomes a ratio of integrals over that density. Hence the paper says it "requires no stochastic partial differential equation". The pricing-measure version needs a C∞ premium as well.
Exponential utility selects the minimal entropy measure under a bounded Sharpe ratio. Given a regularity hypothesis on the premium, VIX² is the expected average of L²F² over the next thirty calendar days. The paper verifies that hypothesis only in special cases with and without leverage. Its regularized calibration with bid-ask constraints has a minimizer if a feasible point has finite objective. Market data play no part, and estimation is deferred to a companion paper still in preparation.
What can be implemented directly?
The text specifies noise loadings q_i = √(2Θa_iγ_i), replacing N separate noise weights with one scale Θ. It also specifies an admissible variance map, φ(y) = v_min + ½(y + √(y² + ϵ²)); the simulations set v_min = 0.04 and ϵ = 0.1. Near-optimal exponential sums require O(log(1/ε)(log(1/ε) + log(T/t0))) terms for accuracy ε, according to the quoted quadrature cost. No number of modes N is fitted to data. An implementer must choose it.
The explosion boundary in one mode
Log returns carry an Itô drift of −½ς². Through the leverage kernel, that drift reaches the variance driver and opposes the restoring force. Theorem 3 locates the boundary for a one-mode family whose potential has degree m and whose F² grows like y^p. With constant, nonzero leverage amplitude and a polynomial potential, the solution is global from every initial state if and only if p ≤ m − 1. For p above that line, suitable starting points give an explosion before T with probability at least 1 − e^(−R/(2c)).
A quartic potential therefore permits F² growth like y³ in that family. The general sufficient condition stops at p ≤ m/2, or y² for a quartic. Arithmetic returns remain global through growth like y⁴. Itkin uses log returns because proving that the index is a true martingale requires the same growth condition under either return convention.
This constrains the optimizer before fitting begins. A lognormal map φ(y) = e^y paired with a polynomial potential violates the growth condition; the parameter box must impose the cap on p.
Hand-chosen paths
The Monte Carlo uses one mode, λ = 0.2, Θ = 0.5, ρ1 = 0 and L ≡ 1. It applies a deterministic drift shock of ±2.5 over [0.4, 0.8]. Each ensemble contains 5,000 paths, with 5,000 Euler steps through T = 2.
After a drawdown, E[Y_T] reaches 0.269 with constant leverage; after a rally, it reaches −0.212. At η = 0.9, the respective values are 0.377 and −0.160. The centered contrast m− + m+ − 2m0 rises from 0.003 under constant leverage to 0.124 under state-dependent leverage. Standard errors are at most 0.001.
Leverage also raises E[Y_T] without a shock, solely through the Itô drift: to 0.027 with constant leverage and 0.046 with state-dependent leverage. Under arithmetic returns, the drawdown and rally responses move to 0.309 and −0.175. The contrast falls to 0.109, while the no-shock lift falls from 0.046 to 0.012.
For the double-well case, cubic coefficient c3 = −2 places a barrier at y ≈ 1.27. Crossing occurs in 0.6% of paths without leverage. Following a drawdown, that share reaches 6.6% under constant leverage and 18.6% under state-dependent leverage. Following a rally, it stays at or below 0.3% in every specification.
These are demonstrations of mechanism with parameters chosen by hand. They measure neither forecast accuracy nor option P&L. E[S_T] equals 0.370 in every run by construction. The geometric mean exp(E[X_T]) drops from 0.320 without leverage to 0.282 with state-dependent leverage.
Our SPY adaptation
We use SPY and its end-of-day options in place of SPX and VIX derivatives. Our end-of-day variance-forecasting run takes the paper's leverage mechanism, leaving its utility-selected pricing measure and joint SPX-VIX calibration question aside. Bid-ask-constrained calibration cannot be reproduced from end-of-day data.
The paper defines VIX through a continuous thirty-day log contract. The author acknowledges that this ignores CBOE truncation error, which a SPY strip inherits. SPY options are American-style contracts on a dividend-paying ETF, so the log-contract identity is only approximate.
The judgment calls are ours.
- N and the leverage rates need to be chosen. Distinct rates are required for the density result, yet the paper fits none to data; the simulation sets N = M = 1 and λ = 0.2 by hand.
- The admissible Sharpe-ratio example is constant, giving a premium Λ0 = 0. The author acknowledges that it yields no memory-risk premium. Theorem 7 supplies a nonzero feedback premium without leverage; Proposition 8 does so for a tuned constant-leverage family.
- Y needs filtering from SPY history. The driver has no direct Brownian forcing, and the paper supplies no filter.
- At the wings, the Dupire ratio can be sensitive. As the author notes, a smooth density supplies no lower bound on spot density there.
The benchmark comes from within the model: set η = 0, then ask whether state dependence pays for its extra parameters out of sample after option costs. We did not find an out-of-sample test or a cost assumption in the paper to use. Our SPY run can assess whether η > 0 improves variance forecasts. It cannot confirm or refute a theorem, and one automated pass is no verdict on the authors' work.