A non-local estimator for locally stationary Hawkes processes

Thomas Deschatre 1 Pierre Gruet 1 Antoine Lotz 1,2 · 2025

Read A non-local estimator for locally stationary Hawkes processes by Thomas Deschatre 1 Pierre Gruet 1 Antoine Lotz 1,2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-02631.txt.

Introduction

Thomas Deschatre, Pierre Gruet, and Antoine Lotz build a 2025 statistics method for Hawkes processes whose rates change over time. Hawkes processes model events that trigger more events, but real rates rarely stay fixed. The team wanted estimates that stay sound when the rate drifts.

Their method adapts the standard maximum likelihood approach and does not rely on set time bins. They show the estimator stays consistent and well behaved as the watch period grows. They also build a test for whether the birth rate is steady, and apply the method to German intraday power market data.

Catalog record

CREATORThomas Deschatre 1 Pierre Gruet 1 Antoine Lotz 1,2
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-02631

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A non-local estimator for locally stationary Hawkes processes Thomas Deschatre 1 Pierre Gruet 1 Antoine Lotz 1,2 Abstract . We consider the problem of estimating the parameters of a non-stationary Hawkes process with time-dependent reproduction rate and baseline intensity. Our approach relies on the standard maximum likelihood estimator ( mle ), coinciding with the conventional approach for stationary point processes characterised by [ Ogata, 1978 ] . In the fully parametric setting, we find that the mle over a single observation of the process over [ 0 , T ] [0,T] remains consistent and asymptotically normal as T → ∞ T\to\infty . Our results extend partially to the semi-nonparametric setting where no specific shape is assumed for the reproduction rate g : [ 0 , 1 ] ↦ ℝ + g\colon[0,1]\mapsto\mathbb{R}_{+} . We construct a time invariance test with null hypothesis that g g is constant against the alternative that it is not, and find that it remains consistent over the whole space of continuous functions of [ 0 , 1 ] [0,1] . As an application, we employ our procedure in the context of the German intraday power market, where we provide evidence of fluctuations in the endogeneity rate of the order flow. Mathematics Subject Classification (2020) : 62F03, 62F12, 60G55 Keywords : Hawkes processes, locally stationary process, maximum likelihood estimation. 1. Introduction 1.1. Context and motivation A Hawkes process (\boldsymbol{N}_{t})=(N_{k,t}) , k=1\ldots K , is a path-dependent point process, wherein past events in a component (N_{k,t}) may induce later occurrences at any other (N_{l,t}) via mutual excitation. The structure of the process is encoded within a kernel function \phi=(\phi_{kl})\colon\mathbb{R}\mapsto\mathbb{R}_{+}^{K}\times\mathbb{R}_{+}^{K} , with the integral \int_{0}^{\infty}\phi_{kl}(s)\mathop{}\!\mathrm{d}s quantifying the strength of the influence of (N_{l,t}) upon (N_{k,t}) . Formally, the intensity (\lambda_{k,t}) of the process writes \lambda_{k,t}=\mu_{k}+\sum_{l=1}^{K}\int_{0}^{t}\phi_{kl}(t-s)\mathop{}\!\mathrm{d}N_{l,s}, where \mu_{k}\in\mathbb{R}_{+} and the integral runs over [0,t) . In its seminal construction at K=1 due to [ Hawkes and Oakes, 1974 ] , the process is endowed a cluster representation. Clusters arrive following a Poisson process with rate \mu , and expand according to a Galton-Watson dynamic with mean offspring number \varrho=\int_{0}^{\infty}\phi(s)\mathop{}\!\mathrm{d}s . The parameter \varrho is thus referred to as a reproduction rate, or endogeneity ratio. The process (\boldsymbol{N}_{t}) is then known to reach stationarity in the sub-critical case \varrho<1 . The cluster representation and the stationarity condition both extend to the multivariate setting (see e.g [ Reynaud-Bouret et al., 2013 ] and [ Brémaud and Massoulié, 1996 ] ), the latter then depending on the spectral radius of \int_{0}^{\infty}\phi(s)\mathop{}\!\mathrm{d}s . Hawkes processes have found productive use cases in seismology [ Ogata, 1988 ] , neuroscience [ Reynaud-Bouret et al., 2013 ] , statistical finance [ Bacry et al., 2011 ] , among other domains. In the wake of these many applications, some recent attention has been dedicated to time-dependent extensions of the model. Time-dependent cluster arrivals were introduced in [ Chen and Hall, 2013 ] so as to model non-stationary phenomena in financial markets, and time-dependent reproduction kernels \phi considered by [ Roueff et al., 2016 ] , with comparable applications. The process (\boldsymbol{N}^{T}_{t}) then lives on the compact timeframe [0,T] , its kernel depending on the normalized duration \nicefrac{{t}}{{T}} and its intensity generally expressing as (1) \lambda^{T}_{k,t}=\Phi_{k}\Bigl[\mu_{k}\bigl(\frac{t}{T}\bigr)+\sum_{l=1}^{K}\int_{0}^{t}\phi_{l,s}\bigl(t-s,\frac{t}{T}\bigr)\mathop{}\!\mathrm{d}N^{T}_{l,s}\Bigr],\hskip 2.84544ptt\in[0,T] where k=1\ldots K , the \Phi_{k}\colon\mathbb{R}\mapsto\mathbb{R}_{+} are activation functions, and \mu_{k}\colon[0,1]\mapsto\mathbb{R}_{+} The results of [ Roueff et al., 2016 ] include the formal basis for the estimation of the model. This is completed in [ Roueff and Sachs, 2019 ] , where the asymptotics of local mean density and local Bartlett estimators are characterized. An analogous line of research is followed in [ Clinet and Potiron, 2018 ] , where the convergence of local maximum likelihood estimators (local mle s) is obtained in the case of gamma kernels \phi(s,\nicefrac{{t}}{{T}})=g(\nicefrac{{t}}{{T}})s^{m}\exp(-\beta(\nicefrac{{t}}{{T}})s) and mixtures thereof, where g and \beta are positive-valued real functions and m\in\mathbb{N}^{\star} . Finally, [ Mammen, 2017 ] obtain the convergence of local polynomial estimators in the compact support case \phi(s,\nicefrac{{t}}{{T}})=g(\nicefrac{{t}}{{T}})\mathbb{1}_{A}(s) , where A is a known compact interval of \mathbb{R}_{+} . In a companion paper [ Deschatre et al., 2025 ] , we consider a subclass of locally stationary Hawkes processes ( 1 ) with linear activation \Phi_{k}(x)\propto x and separable kernel (2) \phi_{kl}\bigl(s,\frac{t}{T}\bigr)=g\bigl(\frac{t}{T}\bigr)\varphi_{kl}(s), where g\colon[0,1]\mapsto\mathbb{R}_{+} and \varphi_{kl}\colon\mathbb{R}_{+}\mapsto\mathbb{R}_{+} . The function g thus corresponds to a time-dependent reproduction rate in the representation of [ Hawkes and Oakes, 1974 ] . We find that the asymptotic behaviour of (\boldsymbol{N}^{T}_{t}) may then be assimilated to that of its stationary counterpart, in that it obeys a functional central limit Theorem generalizing the limit results found in the case \mu\equiv 1 and g\equiv 1 (see [ Bacry et al., 2013 ] ). In this article, we show that a similar correspondence holds regarding the estimation of the process. Working in the fully parametric setting, we prove that the conventional maximum likelihood estimator ( mle ) for stationary point processes, as introduced in [ Ogata, 1978 ] , may be indifferently applied to non stationary processes with intensity ( 1 ) and kernel ( 2 ), for any Lipschitz-continuous (\Phi_{k}) and any sufficiently integrable (\varphi_{kl}) . We also deduce a semi-nonparametric test for time-dependencies in the reproduction rate g . In contrast with the results of [ Roueff and Sachs, 2019 ] [ Mammen, 2017 ] , and [ Clinet and Potiron, 2018 ] , the resulting procedure is not specifically adapted to the locally stationary context. Hence we refer to the present approach as naïve one. We are motivated by [ Kwan, 2023 , Chapter 3] , where the consistency and weak convergence of the conventional mle is obtained in the univariate case K=1 for a time-dependent baseline \mu , a constant reproduction rate g\equiv 1 , a linear activation \Phi_{k}(x)\propto x and any sufficiently integrable \varphi . We find that their results may be extended to any intensity of the form ( 2 ) as detailed in the following subsection. 1.2. Contributions Suppose that one observes over [0,T] a single trajectory of a locally stationary Hawkes process (\boldsymbol{N}^{T}_{t}) with intensity \lambda^{T}_{k,t}(\eta,g)=\Phi_{k}\Bigl[\mu\big(\frac{t}{T},\eta\big)+\sum_{l=1}^{K}\int_{0}^{t}g\big(\frac{t}{T}\big)\varphi_{kl}(t-s,\eta)\mathop{}\!\mathrm{d}N^{T}_{s}\Bigr], where \eta lies in a compact subspace of \mathbb{R}^{p} with non-empty interior, and g\in C[0,1] . Our primary concern lies with the time-inhomogeneous kernel k^{T}(t,s,\eta,g)=g(\frac{t}{T})\varphi(t-s,\eta) , and the amplitude of its fluctuations as \nicefrac{{t}}{{T}} varies. In the fully parametric setting where one has access to a family g(\cdot,\varpi) indexed on \varpi\in\mathbb{R}^{d+1} containing the true reproduction rate, we obtain in Theorem 1 a central limit Theorem for the maximum likelihood estimator ( mle ) of \vartheta=(\eta,\varpi) , generalizing some prior results of Ogata [ Ogata, 1978 ] and [ Kwan, 2023 ] to ( 1 ). In the more general case where no specification is imposed upon g apart from continuity, we construct a test with null hypothesis that g is constant, against the general alternative that it is not. Our test leverages the results of the parametric sub-case, and relies on a maximum likelihood estimator \hat{g}^{\mathfrak{d}}_{T} of g defined as a combination of Bernstein polynomials of fixed degree \mathfrak{d}\in\mathbb{N}^{\star} \hat{g}^{\mathfrak{d}}_{T}\bigl(\frac{t}{T}\bigr)=\sum_{k=0}^{\mathfrak{d}}\hat{\varpi}_{T}\binom{\mathfrak{d}}{k}\bigl(1-\frac{t}{T}\bigr)^{\mathfrak{d}-k}\bigl(\frac{t}{T}\bigr)^{k}, where (\hat{\varpi}_{i}) maximizes the model’s likelihood. We characterize the asymptotic distribution of the associated likelihood ratio statistic \Lambda^{\mathfrak{d}}_{T} as defined in 1 . Under the null hypothesis, the problem reduces to a purely

Read the complete text