The usable output of this paper is a fence around the rough Hawkes-Heston parameter space. Wang, Wu and Zhu ask what the model has to look like if it comes out of a marked Hawkes order flow. Under their subcritical scaling, the Volterra drift b has to be negative and the leverage correlation has to sit strictly inside (-1,0). Price jumps can only go down. The jump measure has to be finite-activity with a finite second moment, and the forward variance curve has to be realizable from the microscopic immigration rate, that is, reproducible from the model's own baseline arrival profile. Critical and supercritical scalings are separate regimes, and the drift restriction does not apply there. The leverage constraint, and the arithmetic that rides along with it, is where the paper earns its keep.
What does the construction actually build?
Rough Heston writes the spot variance as a Volterra equation with the fractional kernel k_α(t) = t^(α-1)/Γ(α), α in (1/2,1), driven by a Brownian motion correlated at ρ with the return driver. El Euch, Fukasawa and Rosenbaum got that out of order flow. A nearly unstable Hawkes process with a heavy-tailed kernel, pushed near criticality, has a Mittag-Leffler renewal resolvent, and the resulting mild equation rewrites in Riemann-Liouville form. Bondi, Pulido and Scotti then bolted on common jumps. In their rough Hawkes-Heston model the same compensated jump measure appears in both the log-return and the variance equation, with intensity proportional to spot variance. One affine model can then chase SPX and VIX smiles at once. Bondi et al. wrote those jumps straight into the price and variance equations. This paper supplies the order-flow mechanism behind the variance and common-jump part of them.
Two sign-specific Poisson random measures are thinned by a single shared intensity λ^T. Every accepted event is either regular, or a rare shock carrying a positive mark z drawn from F_J. Regular events move the price by a tick, with sign weights ω+ = 2/(1+q) and ω- = 2q/(1+q) chosen so that the uncentered signed flow is a local martingale, and negative events carry a heavier variance loading β > 1. A rare event drops the log price by Λz and injects extra excitation c_J r_T z a_T ϕ into future activity. The memory kernel is fixed at ϕ(t) = α(1+t)^(-1-α). The scaling is 1 - a_T ~ μ0 T^(-α), p_T = (1-a_T)² p̄_T and r_T ~ r*/(1-a_T). Under it, the rescaled price, variance and jump measure converge along the full sequence to the unique canonical rough Hawkes-Heston solution. The limits come with a coefficient map, the formulas taking the microscopic parameters to the reduced-form b, c̄_B, ξ̄, ν_J and ρ.
There is no market data anywhere in it.
The evidence is a weak-convergence theorem plus Monte Carlo. At each integer T from 2 to 50 they simulate 100,000 exact immigration-branching paths, with each Hawkes cluster generated from its parent event. There is one option experiment, at T = 1000.
Rare shocks, loud descendants
The accounting that makes the jump component survive is the best part of the construction. An order-one variance level corresponds to microscopic intensity of order T^(2α-1), so roughly T^(2α) events over the horizon [0,T]. Shock probability scales as p_T ≍ (1-a_T)² ≍ T^(-2α). Multiply and the number of marked shocks on the macroscopic clock stays of order one. Each one is individually negligible at the event level. But r_T ≍ T^α amplifies its excitation before the near-critical cascade multiplies it again, so its volatility footprint does not vanish. Vanishing event probability, finite-activity common jumps in the limit. Getting there needs time simplicity of the limiting jump measure, the property that no two macroscopic jumps land at the same instant. It is proved via a no-clustering argument. Without it, J1 convergence of the jump-price functional fails, meaning convergence in the Skorokhod topology that allows jump times to shift slightly.
Subcriticality is what sets the fence
Subcriticality of the marked branching, 2 c_J r* p̄ m1 < 1, is equivalent to ξ∫z ν_J(dz) < μ0. After the fractional rewriting that is exactly b = ξ̄∫z ν_J(dz) - μ0/Γ(1-α) < 0. Subcriticality is what keeps the cascade from exploding, and it is what forces b < 0. Critical and supercritical scalings are separate regimes, and the paper's own well-posedness result covers every real b, so nothing here forbids calibrating outside the fence. It only forbids claiming an order-flow origin while you do.
The leverage constraint is sharper and more interesting. With ρ = -(β-1)√q / sqrt((1+q)(1+β²q)), the symmetric benchmark q = 1 caps the attainable correlation at (-1/√2, 0). The calibrated ρ = -0.731 of Bondi et al. sits outside that, so matching it forces β >= (1+0.731)/(1-0.731) = 6.435. The authors take β = 8, which pins q = 0.041605 and gives activity weights ω+ = 1.9201 against ω- = 0.07989. Read as microstructure, that says the equity leverage effect comes from frequent small upward ticks together with rare downward ticks carrying a much larger variance loading. A correlated Brownian input plays no part in producing it. This is falsifiable on quote data. It is the sharpest of the paper's testable claims.
Two combinations, nine parameters
The reduced-form input is the Bondi et al. vector: α = 0.527, ρ = -0.731, b = -1.812, c̄_B = 0.115, Λ = 0.276. Inverting the coefficient map on it, after normalizing the variance-jump loading to one and fixing F_J = Exp(1), produces a microscopic instance. It is (μ0, p̄, c_J, r, θ, m, β, q, F_J) = (5.2653, 0.006772, 1, 26.2566, 0.07131, 1, 8, 0.041605, Exp(1)). The authors state the problem with this outright. The map identifies θ and m only through the ratio θ/m, and c_J and r only through the product c_J r. Hence "the individual microscopic parameters therefore remain non-identifiable." They also flag that β = 8 is one admissible realization, and that the lower bound 6.435 is the only implication the correlation actually delivers. In their own framing the map characterizes an attainable region rather than estimating anything: contribution (C3) says they "characterize the structural parameter region selected by the microscopic model", and Proposition 5.2 proves the converse attainability statement conditional on baseline realizability. So the charge lands on anyone who reads the fitted tuple as microstructure, not on the theorem. The ω+ = 1.92 versus ω- = 0.08 picture illustrates the fence.
Prices constrain the microstructure to a manifold and cannot pick a point on it.
Convergence to the paper's own reference law
The Wasserstein-1 distance between the microscopic terminal variance and the rough limit falls from 2.37e-2 at T = 2 to 3.09e-3 at T = 50. The integrated-variance distance falls from 1.20e-2 to 7.63e-4, at 100,000 paths per point. The large-T plateau stays above the 5-95% Monte Carlo resolution band. The authors say so, and attribute it to a finite-T pre-asymptotic discrepancy together with the numerical approximation of the affine reference law. They do not separate the two. They also decline to fit a rate, writing that the dense grid "displays the distributional approximation over the pre-asymptotic range without fitting a rate below numerical resolution." Fair, and it costs something: with no rate, a practitioner has no way to judge how far a real horizon sits from the limit. This experiment also runs at α = 0.65, while the option experiment runs at the calibrated α = 0.527.
The smile comparison is a consistency check between two of the authors' own objects. Microscopic paths at T = 1000 are priced by simulation, the limit by the affine transform with 32 kernel factors. Both use the same borrowed parameter vector, at maturities 23/730 and 66/730. The microscopic prices reproduce the deep-put negative skew, the at-the-money minimum and the short right-wing upturn of the affine limit, with 95% Monte Carlo bars widest for the deepest puts at the shorter maturity. Nothing in it is a fit to quotes. Bondi et al. did the calibration; this paper checks that its microscopic model, run at a large but finite scale, agrees with the limit it converges to.
One thing the paper does not oversell. It scopes the claim to the variance and common-jump mechanism, and the conclusion says that "Our construction combined this endogenous mechanism with reduced-form exponential price normalization and direct marked-event price loading". They assume the -Λz price impact of a shock rather than deriving it, and they say so. The convergence results are also proved for the single canonical kernel ϕ(t) = α(1+t)^(-1-α); we did not find a statement extending them to other regularly varying memory profiles.
We could not run this on our data. The mechanism lives on signed event-level arrivals, with regular and rare marks separated and a branching structure estimated from them. We have one-minute OHLCV bars, with no trade prints and no quotes. A daily or minute-bar proxy would only test a realized-volatility model.
Tick-level evidence from index futures or SPX options would move me. The construction at ρ = -0.731 demands a frequency-versus-size asymmetry in quote revisions: an upward-tick activity weight about twenty-four times the downward one (ω+ = 1.9201 against ω- = 0.07989), with the downward side carrying most of the variance loading. Absent that, the paper is a clean theorem about which corners of the rough Hawkes-Heston box have an order-flow story behind them, and a warning that the box coordinates cannot be read back into microstructure.