Self- and Mutually-Exciting Processes across the Microstructure Stack

Bacry, Mastromatteo, and Muzy provide a comprehensive survey of Hawkes processes applied across the full microstructure stack, unifying under a single mathematical framework several applications that are often treated as methodologically disjoint. The survey covers: tick-level volatility estimation via the identification of quadratic variation with the integral of the Hawkes intensity; measurement of market endogeneity via the scalar branching ratio; cross-asset and cross-venue systemic contagion analysis via the spectral radius of the kernel matrix; optimal execution under Hawkes-driven market impact in the Almgren–Chriss framework; and full order-book modeling via bid- and ask-side excitation intensities. The central thesis is that a single mathematical object — the matrix kernel Φ(τ) of a multivariate Hawkes process — encodes the cross-event dynamics at all of these levels, and that calibrating Φ from high-frequency data provides a unified basis for volatility modeling, contagion measurement, and momentum signal interpretation. The survey's taxonomic value for a quantitative desk is substantial. By demonstrating that the branching ratio (Hardiman–Bouchaud), the rough-vol kernel (El Euch–Rosenbaum), and the order-book excitation (various authors) are all special cases of the multivariate Hawkes framework at different levels of aggregation, the paper creates a common vocabulary and a common set of estimation tools that can be applied across the desk's different analytical functions without requiring the construction of separate theoretical foundations for each application. The spectral radius ρ(∫Φ) — the largest eigenvalue of the matrix of integrated kernels — emerges as a single aggregate endogeneity indicator that collapses the multivariate excitation structure to a scalar, providing a natural generalization of the scalar branching ratio to the cross-asset setting. For this desk's methodology, the Bacry–Mastromatteo–Muzy survey occupies the role of an architectural reference: it is the document that establishes the common process-level language shared across microstructure, volatility, and cross-asset analysis, and it is the background against which all specific Hawkes-based methodological choices should be understood and justified.

The multivariate Hawkes process on d event types has conditional intensity vector λ_t = (λ^1_t, ..., λ^d_t) with λ^i_t = μ^i + Σ_{j=1}^d ∫_0^t Φ_{ij}(t − s) dN^j_s where μ^i > 0 is the exogenous baseline for type i, N^j_t counts events of type j up to time t, and Φ_{ij}(τ) ≥ 0 is the excitation kernel from type j to type i. The matrix of integrated kernels n̂ = ∫_0^∞ Φ(τ) dτ has components n̂_{ij} = ∫_0^∞ Φ_{ij}(τ) dτ, and the stability condition for the process to be stationary is that ρ(n̂) < 1 where ρ denotes the spectral radius. The covariance matrix of the increment process is Σ = (I − n̂)^{−1} diag(μ) (I − n̂)^{−T}, generalizing the univariate result. The Bartlett spectral matrix is S(ω) = (I − Φ̂(ω))^{−1} diag(μ) (I − Φ̂(ω)*)^{−1} where Φ̂(ω) = ∫_0^∞ e^{iωτ} Φ(τ) dτ is the Fourier transform of the kernel matrix; near-criticality appears as a divergence of S(0) = (I − n̂)^{−1} diag(μ) (I − n̂)^{−T} as ρ(n̂) → 1. The quadratic variation of a Hawkes-driven log-price process equals ∫_0^T λ^{price}_t dt in the diffusive limit, linking the intensity directly to instantaneous variance.

The taxonomic value of the survey is best understood by comparison with the alternative — maintaining separate modeling frameworks for volatility, contagion, and momentum — which would require separate calibration pipelines, separate vocabularies for discussing results, and separate theoretical foundations that may contradict each other when applied to the same underlying data. The Hawkes framework avoids this fragmentation by identifying a single object (Φ) that must be estimated once and can then be queried for all downstream applications, and the survey's contribution is to make this unification explicit and to demonstrate its consistency across the range of microstructure applications considered. The distinction between time-domain and spectral estimation of Φ is practically significant. Time-domain MLE requires specifying a parametric form for Φ_{ij}(τ) — typically exponential or power-law — and fitting its parameters via numerical optimization of the likelihood, which is computationally tractable for low-dimensional processes (d ≤ 5) but becomes expensive for larger d due to the quadratic scaling of the parameter count and the convolution structure of the likelihood gradient. Spectral estimation, which fits Φ̂(ω) directly from the empirical Bartlett spectrum, avoids the parametric assumption but requires dense frequency grids and careful bias correction, and its estimates of Φ in the time domain via inverse Fourier transform are subject to boundary effects and Gibbs phenomena. For the desk's purposes, the MLE approach with power-law kernel parameterization is the default, with spectral estimation used as a cross-check on the tail exponent. The curse of dimensionality in multivariate calibration is the primary practical limitation: for d assets or event types, the kernel matrix Φ has d² components, each of which is a function of lag τ, and estimating the full matrix reliably requires event counts scaling as d² times the count needed for the univariate case. This makes full multivariate Hawkes calibration feasible only for small d at the individual-name level, while cross-asset analysis is typically conducted at the index or sector level.

The shared state description that the Hawkes framework provides — a single kernel matrix Φ calibrated from data — enables unified reasoning across microstructure, volatility, and cross-asset memo sections without requiring translation between different model languages at each interface. In practice, this means that the desk's daily calibration output from the Hawkes estimation pipeline — the matrix n̂ and its spectral radius ρ(n̂) — can serve simultaneously as: (a) an endogeneity indicator for momentum signal confidence; (b) a roughness prior via the power-law tail exponent identification; and (c) a contagion indicator for cross-asset risk aggregation. Having these three functions share a common calibration object is a methodological efficiency that compounds over time as the calibration history accumulates and cross-functional regime analysis becomes possible. The spectral radius ρ(n̂) as a single endogeneity indicator aggregating all cross-asset channels is particularly useful for the desk's risk management function: a rising ρ(n̂) indicates that the cross-asset excitation structure is approaching criticality in aggregate, even if no individual univariate branching ratio n_{ii} has crossed a threshold, and the desk's position sizing protocols should respond to ρ(n̂) as a primary regime-change early warning signal.

The practical application of multivariate Hawkes calibration requires venue-level, self-trade-filtered tick data for each of the d event types included in the process, and the data requirements grow quadratically with d, creating a hard boundary on the dimensionality of multivariate analysis that can be conducted with the data infrastructure available to a typical desk. Non-stationarity of Φ across trading sessions — intraday seasonality, pre- and post-announcement dynamics, regime shifts in market microstructure — means that a single Φ calibrated over a multi-day or multi-week window conflates genuinely different microstructure regimes, and the calibration should ideally be conducted separately for intraday sessions and then aggregated with appropriate weighting. The parametric kernel assumption, required for tractable MLE, restricts the class of excitation dynamics to those expressible as sums of exponential or power-law functions, which may be misspecified for some event-type pairs. Nonparametric kernel estimation exists but requires substantially more data and does not naturally connect to the theoretical rough-vol limit. The tension between the parametric assumption needed for tractability and the nonparametric truth of the kernel is a standing limitation of the Hawkes framework applied to financial data, and it means that the kernel matrix Φ should be interpreted as a projection of the true excitation structure onto the assumed parametric family rather than as a direct estimate of the underlying process.