Comparison of the Hawkes and SEIR models for

the spread of Covid-19

Conor Kresin∗, Frederic Paik Schoenberg∗, and George Mohler†

Abstract
Two models that capture the spread of infectious diseases, the Hawkes
point process model and the SEIR compartmental model, are compared
with regard to their use in modeling the Covid-19 pandemic. The physi-
cal plausibility of the SEIR model is weighed against the parsimony and
flexibility of the Hawkes model. The mathematical connection between
Hawkes and SEIR models is described.

∗ Department of Statistics, University of California Los Angeles

† Department of Computer and Information Science, Indiana University - Purdue University
Indianapolis
MSC 2020 subject classifications. Primary-60G55; secondary-92D30
Key words: compartmental models, Coronavirus, Hawkes models, model evaluation, point
process models, self-exciting point process models

1
1 Introduction
The SARS-CoV2 (Covid-19) pandemic spread from China to at least 188 coun-
tries or regions in the first six months of 2020 [15]. The characteristics of the
Covid-19 virus have been estimated and forecasted by numerous researchers
with highly variable results. Estimates of properties such as reproduction rate
(or time-varying reproduction number), numbers of individuals infected, hospi-
talization rates, fatality rates, and efficacy of containment measures have varied
widely [4, 22]. Accurate real-time estimates of the spread of Covid-19 are dif-
ficult to achieve without population-wide testing [18, 6]. Nevertheless, it is
important for researchers to accurately estimate and forecast the dynamics of
Covid-19 so that optimal public policy measures and other responses can be
adopted.

Several different frameworks have been proposed for modeling the spread of
Covid-19, including compartmental models such as the SEIR (Susceptible →
Exposed → Infectious → Removed) differential equation model, and branching
point process models such as the Hawkes point process model [70, 37, 6, 10].
This paper compares these two approaches for forecasting Covid-19. Relative
to Hawkes models, SEIR models and their variants have been used far more
widely to describe the Covid-19 pandemic [7, 24, 31, 42] as well as other infec-
tious diseases such as Ebola [43] and SARS [19]. However, recent studies have
suggested that Hawkes models may be more accurate [87]. For general discus-
sion of mathematical and statistical models of epidemiological phenomena, see
[60, 28].

This paper is structured as follows. Following a review of Hawkes and SEIR

models in Section 1, we compare their advantages and disadvantages, especially
with respect to forecasting Covid-19 cases or deaths in Section 2. In Section
3, we detail the mathematical connection between Hawkes processes and SEIR
models, and in Section 4 we provide concluding remarks.

1.1 The Hawkes Model

The Hawkes model or self-exciting point process model is commonly used to
model clustered point patterns in applications such as seismology, finance, crime,
and infectious diseases [16, 69, 57, 9]. A spatial-temporal Hawkes process is
specified by the model
Z
λ(s, t) = µ(s) + K g(s − s0 , t − t0 )dN (s0 , t0 ) (1)
t0 <t
X
= µ(s) + K g(s − s0 , t − t0 ),
(s0 ,t0 ):t0 <t

for s ∈ X ⊆ R2 and t ∈ [0, T ), where λ(s, t|Ht ) is the conditional rate at

which points (events) are expected to accumulate around spatial-temporal lo-
cation (s, t), given information on all previous events. The conditional intensity
uniquely characterizes the finite-dimensional distribution of any simple point
process (see Prop. 7.2.IV of [16]), and thus equation (1) fully specifies the model.

2
The function g is typically assumed to be a density, i.e. to be nonnegative and to
integrate to 1 over all time and space, and is called the triggering density. Com-
mon choices for g are the exponential or Pareto densities in time, and the Gaus-
sian or Pareto densities in space [69]. The constant K is called the productivity.
Provided g is a density function, K is the expected number of points triggered
directly by each point, and is thus closely connected to the reproduction number
in compartmental models such as SEIR. Each background point, associated with
µ(s), is expected to generate K + K 2 + K 3 + ... = 1/(1 − K) − 1 = K/(1 − K)
triggered points. As a result, in a Hawkes process, the expected fraction of
background points is 1 − K.

Given a dataset consisting of n points within a space-time observation region

B, the parameters in Hawkes processes are typically fit by maximum likelihood
estimation (MLE), where one obtains parameter estimates Θ̂ maximizing
n
X Z
L(Θ) = log(λ(si , ti )) − λ(s, t)dtds.
i=1 B

The resulting estimates have desirable properties. For instance, Ogata (1978)
showed that the MLE Θ̂, is, under standard conditions, asymptotically un-
biased, consistent, asymptotically normal, and asymptotically efficient, with
standard errors readily constructed using the diagonal elements of the inverse
of the Hessian of L evaluated at Θ̂ [59]. Further, if the fitted model is missing
some relevant covariates, under general conditions the MLE will nevertheless be
consistent, provided the effect of the missing covariates is small [74]. The trig-
gering function can also be estimated non-parametrically [49], and some authors
have also estimated the background rate µ(s) nonparametrically, e.g. [91, 65].
Bayesian methods can also be used to estimate parameters and quantify uncer-
tainty in Hawkes process models [68, 53].

A host of variations of the Hawkes model have been proposed [10, 70]. The
HawkesN model, as defined in [70], has a Hawkes conditional intensity scaled by
the proportion of events which can still occur after time t, in order to account
for the dynamic decrease in the number of susceptible individuals in a given
location [70]: X
λ(t) = (1 − Ic (t)/N )(µ + K g(t − t0 )). (2)
t0 <t

In the context of a Hawkes process modeling the spread of an infectious disease,

Ic (t) is the cumulative number of infections that have been recorded up to time
t and N is the total population size.

Hawkes models and their slight variants such as the epidemic-type aftershock
sequence (ETAS) model [57, 58], HawkesN [70, 6, 54], and the recursive model
[76] have been shown to be useful in modeling infectious diseases such as Ebola
[39, 64], chlamydia [75], SARS [82, 9], measles [21], meningococcal disease
[51], and Rocky Mountain Spotted Fever [76]. Hawkes models have also been
shown to be the best fitting models for forecasting seismicity in rigorous, purely
prospective earthquake forecasting studies such as the Collaboratory for the
Study of Earthquake Predictability (CSEP) [13, 14, 90, 8, 27, 77].

3
1.2 The SEIR Model
SEIR models and their variants have been widely used to model and forecast
the spread of many contagious diseases including Covid-19 [81, 86, 7, 29, 44,
32, 30, 66, 52, 31]. Such models employ a wide variety of modifications to the
classic SEIR model, including using Bayesian inference [52], machine learning
[31], mobility networks and ensemble approaches [32], and mixed-effects curve
fitting [81] to fit parameters, as well as slight compartmental variants like SuIER
which account for unreported cases [30]. Other models explicitly define scenarios
for government interventions or enforcement of public health policies in specific
populations [44, 7, 29].

SEIR models assume that individuals within each category, or compartment

(susceptible, exposed, infectious, and recovered), share pertinent characteris-
tics, and the size of the population of interest N is equal to the total number
of individuals in the compartments [40]. SEIR models are a slight extension
of SIR (Susceptible → Infectious → Removed) models, generalized to account
for the fact that there is an incubation time for some infectious diseases like
Covid-19, during which the exposed host may be asymptomatic and thus not
recorded as infected. SEIR models can be either deterministic, in which case
they are comprised of a system of differential equations, or stochastic, in which
case they are based on a Markov chain framework. Given large populations,
sufficient initial spread, and enough time, the deterministic framework should
resemble the stochastic framework in expectation, assuming properly specified
models [70].

Deterministic SEIR models, such as that described in Figure 1, can provide a

reasonable approximation of the characteristics of a contagious disease such as
Covid-19. There are numerous variations, but the basic idea conveyed in Figure
1 common to compartmental models is that there is some rate at which people
shift from one portion of the population to another, e.g. from the susceptible
population to the exposed population, and these rates may be fixed or allowed
to vary over time subject to certain constraints. Deterministic models such as
that shown in Figure 1 can be extended to allow parameters governing the force
of infection, number of cases by symptom onset, and death rate, with move-
ment between compartments commonly specified as binomial random variables.
Such a model has been suggested for the transmission of Ebola, for instance
[43]. Number of cases or deaths are commonly specified as a negative binomial
random variable [45, 35].

Perhaps the most common method for estimating compartmental infectious dis-
ease model parameters is by using Bayesian estimation [62, 12]. Prior parame-
ters are often decided on using subject matter experts [52] or parameters fit to
prior outbreaks [25]. Bayesian SEIR models have been employed to model infec-
tious diseases such as Ebola [25], Visceral Leishmaniasis [62] and Covid-19 [52].
Prior distributions are typically specified by compartment. For instance, Frasso
et al. specified number of deaths as beta-distributed, duration of incubation
as normal, and observed cases as negative binomial [25]. Disease characteris-
tics such as reproduction number has been modelled within the context of SIR
models with a gamma prior [12]. Joint posteriors are then solved for using a

4
 Susceptible

dS β·I·S
dt =Λ−µ·S− N

Exposed

dE β·I·S
dt = N − (µ + α)E

Infectious

dI
dt = α · E − (γ + µ)I

Recovered

dR
dt =γ·I −µ·R

Figure 1: Diagram of the deterministic SEIR model. Definitions: N is a constant

number of individuals in a susceptible population, β · I is equal to the force of
infection, Λ equal to birth rate, µ equal to death rate, γ equal to mortality rate,
α−1 equal to the average incubation period. Such a model has reproduction
α·β
number R0 = (µ+α)(µ+γ) .

MCMC approach such as Metropolis-Hastings or Gibbs sampling [12, 43].

Stochastic versions of the SIR and SEIR models allow researchers to include the
effect of networks of individuals, but specification and parameter estimation can
be more challenging [3]. Various stochastic SEIR models have been developed to
model Covid-19 data. A stochastic SEIR model with parameters fit using grid
search, which may be viewed as a relatively agnostic machine learning approach,
was implemented in [31], and a stochastic SEIR model hybrid with agent-based
simulation was suggested in [42]. The compartmental approach of the SEIR
model is slightly modified to accommodate under-detection and differentiated
government intervention in the DELPHI model [7]. Their flexibility notwith-
standing, the difficulty in estimating time-varying parameters in real time for
stochastic SEIR models is well known, especially for large populations [55].

Parameters in SEIR models are often estimated using opinions of expert epi-
demiologists or using data from other locations or past epidemics [11]. This
is attractive in the sense that expert opinion is integrated, but there is ample
opportunity for bias as well as mis-specification, and the parameter estimates
have a covariance structure that can be difficult to estimate. Further, non-
identifiability is a known problem for compartmental models [26]. Although
there exists algebraic approaches for testing identifiability such as exhaustive
modeling [83], such methods are not implemented in any of the above refer-
enced Covid-19 SEIR models. Crucially, estimated SEIR parameters in the
early stages of a epidemic (before peak infection) have been shown to be struc-

5
turally nonidentifiable [73].

2 Comparison of Point Process and Compart-

mental Models
Hawkes and SEIR models both offer flexible (and somewhat complementary)
frameworks for modeling infectious diseases. Hawkes models allow for nonpara-
metric estimation of the triggering function g, as well as spatial covariates, and
an intrinsic network-effect. SEIR models offer a far more physically plausible
framework for describing Covid-19 relative to the Hawkes model. Specifically,
SEIR models allow for specification of stochastic movement between compart-
ments based on previous epidemics and expert opinions. The compartmental
model framework allows for natural implementation of known networks within
the population of interest [55]. Further, quantities of interest to epidemiologists
and policy makers such as infection rate within a population can be imputed
using SEIR models [43].

Within the context of a SEIR model, the spread or transmission of an infectious

disease such as Covid-19 occurs via Markovian diffusion which, under certain
regularity conditions, ultimately converges to a stationary distribution. In the
context of a Hawkes model, background events trigger future events, and these
trigger subsequent events, ultimately resolving due to the decay of the chosen
triggering function if the productivity is less than one. In general, a point pro-
cess is considered to be stationary when for all bounded Borel subsets A1 , . . . , Ar
of the real line, the joint distribution of {N (A1 + t), . . . , N (Ar + t)} is indepen-
dent of t ∈ R [16]. A Hawkes process with K > 1 is not stationary [80]. The
HawkesN model is a stationary process for K > 1.

The link between Hawkes-like and stochastic SIR models is explored in detail in
[70], where it is shown that an exponentially decaying triggering function chosen
for a finite population Hawkes model (HawkesN) coincides in expectation with
the number of individuals infected in a stochastic SIR model as it approaches
stationarity. This connection between SIR and Hawkes models was explored in
particular in the context of Covid-19 [6], where it was shown that the HawkesN
and SIR models converge if the triggering function is exponential and the re-
production number in the SIR model is constant [6]. SIR and HawkesN models
are shown to provide similar fit to Twitter re-tweet diffusions in [70].

One may also compare features such as the doubling time for both Hawkes and
SIR/SEIR models. In the early exponential growth stage, the doubling time for
SIR is τ = log(2)/(γ(R0 − 1)) [1]. The relationship between estimates of R0
and doubling time for simulations of compartmental models is summarized in
[50]. The parameter K is intuitively similar to R0 , as it represents the expected
number of events triggered by a previous event. The doubling time for HawkesN
models as a function of K is shown next to the doubling time of a SIR model as
a function of R0 in Figure 2. It should be noted that doubling time for Hawkes
models quickly approaches zero for K > 1, justifying the finite population cor-
rection present in the HawkesN in the context of modeling infectious diseases

6
such as Covid-19 [70].

5.5 SIR
Average HawkesN
5
average doubling time
4.5

3.5

2.5

1.5
1.5 2 2.5 3 3.5 4
R0

Figure 2: Average doubling time for HawkesN model with β = 14 , I0 = 10,

population size N = 106 , and using mean intensity over 100 simulations per K
(notated R0 for SIR). Doubling time is defined as t such that N (t) = 20.

Hawkes models offer computationally inexpensive parametric and non-parametric

estimates for important characteristics of infectious diseases such as Covid-
19. Due to computational difficulty, and model-specification convenience, both
SEIR and Hawkes models often make assumptions such as fixed population size,
or homogeneity individuals within compartments. Despite this, the difficulty of
specifying large population size stochastic SEIR models in real time is not trivial.
In general, Hawkes models seem to be far simpler to implement than SEIR-type
models, and in a pandemic such as the spread of Covid-19, where resources can
be scarce and policies and health-allocations must be made in real-time, quick
and accurate short term forecasts are highly valuable [85].

Relative to Hawkes processes, SEIR models are more natural mathematical rep-
resentations of the spread of contagious diseases. However, in implementation
SEIR models often require estimation of more parameters and structural modifi-
cations. With complexity, there is more opportunity for bias and random errors
in parameter estimates, as well as large covariances between pairs of parameter
estimates, and in some cases problems of identifiability [20, 71]. More press-
ingly, each component of the model is susceptible to mis-specification, which
can result in highly variable estimates and large forecasting errors [34, 61].

Problems such as these can be particularly severe in the case of Covid-19, where
available data used to fit parameters can rely can have substantial errors, due to
undercounted infected populations and testing policies that vary over time and
space [41]. Both Hawkes and SEIR models assume a homogenous population and
do not explicitly account for testing errors, but Hawkes and HawkesN models
appear to perform better than their SEIR equivalents for modeling the spread
of infectious diseases (see Table 1 below).

7
 2.1 Comparison of Covid-19 Results in the Current Lit-
erature
SEIR models appear to be far more widely used by State and Federal agen-
cies for forecasting Covid-19 cases and deaths, with a notable exception being
the State of New Jersey which is primarily using a multivariate Hawkes model
[36]. Table 1 summarizes results comparing the accuracy of Hawkes models and
their variants with SEIR models and their variants for forecasting infectious dis-
eases. Point process models have been found to forecast incidence of mumps in
Pennsylvania better than compartmental SVEILR models [38]. Further, point
process models have been found to improve fit and forecasting performance rel-
ative to SEIR models when applied to incidence of pertussis in [87]. Yuan et al.
find substantially improved accuracy of Hawkes models over SEIR models for
forecasting Covid-19 in the European Union, California, New York, and for the
United States as a whole [89].

Data Better Fit Worse Fit Reduction RSME Authors

Pertusis in NV Recursive Hawkes SEIR 19% [87]
Recursive Hawkes SVEILR 38%
Mumps in PA [38]
Hawkes SVEILR 26%
Covid-19 in CA
SEIR Hawkes
Covid-19 in IN (∗) [6]
Covid-19 in NY Hawkes SEIR
Covid-19 in CA 63%
Covid-19 in NY 21%
Hawkes SEIR [89]
Covid-19 in US 31%
Covid-19 in EU 27%
Covid-19 in US Hawkes Variants SEIR (∗∗) [10]
Ebola in W. Africa Hawkes SEIR 38% [64]

Table 1: Prior results comparing the forecasting accuracy of point process and compartmen-
tal models for infectious diseases. Errors reported are the root mean squared error (RMSE)
and (∗∗) mean absolute error of daily forecasts. Model selection using (∗) Akaike Information
Criterion (AIC) and (∗∗) Normalized Discounted Cumulative Gain.

Hawkes models are directly compared to SIR and SEIR models to explain the
spread of Covid-19 in California, Indiana, and New York in [6]. The Akaike
information criterion (AIC) is used to evaluate the candidate models, and by
this metric, HawkesN performs more poorly relative to its compartmental coun-
terparts for Covid-19 death data, and with mixed results for Covid-19 case data.
However, fitted parameters are found to vary materially across locations, and
relative fit of parameters across models is concluded to not be strongly indicated.
Rather than concluding on the merits of either type of model, the authors note
the difficulty of using limited data at the beginning of an epidemic such as that
of Covid-19 [6].

8
In the context of the Covid-19 pandemic, compartmental models such as SIR
and SEIR have been noted to generally have low accuracy for long-term fore-
casts, and machine learning models have been proposed as a superior alternative
[2]. Compartmental models also may be poorly calibrated for forecasting more
than five days out: forecast numbers of Covid-19 cases in Italy six days in the
future based on the SEIR model were 14% too low on average [2]. Various com-
partmental models for forecasting Covid-19 yield slightly different projections of
future cases or future deaths [72]. However, estimates of variability vary widely,
with prediction interval widths often varying by a factor of 3 [7, 24].

Some variation is to be expected in both mean predicted deaths and size of

prediction interval between the models as each are designed differently, and
with varying assumptions. Estimates of the initial reproduction number Rt for
COVID-19 vary around 3.28 (1.4, 6.5) [63]. Of course, values of Rt are observed
to vary substantially depending on social distance policies. In China, estimates
of Rt decreased from 2 to 1 when public health measures were put in place [88].
Estimates of Rt in Singapore correspondingly decreased over time by between
78.2% and 99.3% [37]. Similar results were observed in Europe as a result of
public health measures [23].

SEIR forecasts of future confirmed cases or deaths depend critically on estimates

of the total numbers of asymptomatic or mildly symptomatic cases, which are
highly uncertain [78, 5] and extremely difficult to estimate accurately [48, 79].
Jewell et al. [37] note that more detailed and complex models may be more sen-
sitive to assumptions regarding the incubation and infectious periods and other
estimates of transmission characteristics. Further, SIR and SEIR models are
highly sensitive to assumptions regarding social movement and the estimated
impacts of containment policies [67]. SIR and SEIR models are known to be
particularly sensitive to assumptions about the distribution of latent and infec-
tious periods [47, 84]. Further, as discussed above, nonidentifiable parameters
can be an issue for compartmental models, and methods for dealing with non-
identifiability of parameters tend to work better for simpler models than for
more complex compartmental models [71].

3 Further Connections Between Hawkes and SEIR

Models
The productivity constant K in the Hawkes model is the obvious analogue of the
reproduction rate R0 in SEIR, with both interpretable as the expected number
of direct transmissions per infected individual. Further, several variations of
the Hawkes process in Equation 1 have deeper connections to SEIR-type com-
partmental models. The point process governed by Equation 2 is a continuous
time analog of a discrete stochastic SIR model when g(t) is specified as expo-
nential [70]. When g(t) is chosen to be gamma distributed, the Hawkes process
also can approximate staged compartment models, like SEIR, if the average
waiting time in each compartment is equal [47]. More complex parametric (or
non-parametric) inter-infection time distributions g(t) may be employed within
the Hawkes process framework in situations where disease dynamics cannot be

9
captured by a SIR or SEIR model. In the early exponential growth stage of an
epidemic, before finite population and social distancing effects play a role, the
linear Hawkes process in Equation 1 can readily be used to model new infections
(see Figure 3).

4500 10 0
SEIR
4000 Hawkes

3500 10 -1

3000
new infections

10 -2
2500

(ti-t j)
2000
10 -3
1500

1000
10 -4
500

0 10 -5
0 20 40 60 80 100 10 1
t t i-t j

Figure 3: Left: (Red) SEIR differential equation dS/dt = −βSI/N , dE/dt =

βSI/N − µE, dI/dt = µE − γI, dR/dt = γI, where β = γR0 ,P γ = .1, R0 = 2,
µ = 1, and N = 5 · 108 . (Blue) linear Hawkes process λt = µ + t>ti Kg(t − ti )
fit to the SEIR curve of new infections using non-parametric expectation-
maximization [54]. Right: Non-parametric histogram estimate for g(t) cor-
responding to the Hawkes process fit.

While the Hawkes process can approximate SEIR in some situations with an
appropriately chosen kernel g(t), queue-Hawkes processes [17] can also be used
to model an exposed latent class of events. Let N be population size, NtE be
the cumulative sum of infections (whether recovered or not) up to time t. Then
we may define a hybrid model incorporating features of both SEIR and Hawkes,
which we call a SEIR-Hawkes process, where the intensity of newly exposed
cases is given by
NtE X
   
E I
λ (t) = 1 − R0 γ exp − γ(t − tj ) , (3)
N I t>tj

and the times of infection are generated via

Z ∞  
P (tIj > tE
j + c) = µ exp − µ(s − t E
j ) ds. (4)
c

Realizations of the SEIR-Hawkes process can be generated via Lewis’ thinning

method for simulation [56, 46]. We first simulate an upper-bounding Hawkes
process with intensity
X  
E I
ν (t) = R0 γ exp − γ(t − sj ) . (5)
t>sIj
Z ∞  
P (sIj > sE
j + c) = µ exp − µ(s − sE
j ) ds. (6)
c

10
Because the Hawkes process in Equation 5 has a branching process representa-
tion [33], the process can be simulated iteratively; for each event pair (sIj , sE
j ),
by
1. Generating a Poisson random variable M with mean R0 .
2. Generating l = 1, ..., M events with inter-event times sE I
l − sj given by an
exponential random variable with parameter γ.
3. Generating l = 1, ..., M events with inter-event times sIl − sE
l given by an
exponential random variable with parameter µ.
Thinning then proceeds sequentially by accepting each event pair (sIj , sE
j ) with
E E E E E
probability λ (sj )/ν (sj ) where λ is computed using only accepted events
in the history and ν E is computed using all simulated events. In Figure 4
we simulate the SEIR-Hawkes process with parameters µ = 1, γ = .1, R0 =
2, N = 1000 and N0E = 10 (tE E
1 = ...t10 = 0) and compare to the forward-
Euler approximate solution (dt = .01) of a SEIR differential equation dS/dt =
−βSI/N , dE/dt = βSI/N −µE, dI/dt = µE−γI, dR/dt = γI, where β = γR0 .

Figure 4: SEIR differential equation simulation (red) and 50 realizations of the

SEIR-Hawkes process. Parameters for the SEIR model are R0 = 2, µ = 1 for
the E → I rate, γ = .1 for the I → R rate, and population size N = 1000.

4 Conclusion
The SEIR model is currently far more widely used to model epidemic diseases
such as Covid-19 than the Hawkes model, and its parameterization is phys-
ically plausible, with parameters that are readily interpeted in the epidemi-
ological community. The SEIR model also appears to forecast epidemics ade-
quately in most cases, especially in the early spread of the disease. However, the
Hawkes model seems to offer more accurate forecasts, with approximately 20-
30% smaller errors on average in most cases. Among the several reasons listed in
Section 2 for this discrepancy, the most significant seem to be mis-specification

11
in the SEIR model and its sensitivity to errors in estimates of latent quantities
such as the number of asymptomatic individuals and the distribution of incuba-
tion times. In general, when maximal accuracy is desired, models for forecasting
observations should typically be only as complex as necessary to represent the
main features of interest in the data, with minimal dependence on unobserved
or noisy data [39].

There are close connections between SEIR and Hawkes models, and indeed
the two types of models can be constructed to be equivalent or to converge
to one another in special cases. The SEIR-Hawkes model described here may
provide further linkage between the two paradigms in cases where one seeks the
accuracy of point process modeling without sacrificing the physical plausibility
and interpretation of SEIR parameters, and the model is shown here to emulate
characteristics of SEIR models closely.

5 Acknowledgements
This research was supported by NSF grants SCC-1737585 and ATD-1737996.

12
References
[1] Linda JS Allen. Some discrete-time si, sir, and sis epidemic models. Math-
ematical biosciences, 124(1):83–105, 1994.
[2] Sina F Ardabili, Amir Mosavi, Pedram Ghamisi, Filip Ferdinand, An-
namaria R Varkonyi-Koczy, Uwe Reuter, Timon Rabczuk, and Peter M
Atkinson. Covid-19 outbreak prediction with machine learning. Available
at SSRN 3580188, 2020.
[3] Jesus R Artalejo, Antonis Economou, and Maria Jesus Lopez-Herrero. The
stochastic seir model before extinction: Computational approaches. Applied
mathematics and computation, 265:1026–1043, 2015.
[4] David Baud, Xiaolong Qi, Karin Nielsen-Saines, Didier Musso, Léo Po-
mar, and Guillaume Favre. Real estimates of mortality following covid-19
infection. The Lancet infectious diseases, 2020.
[5] Eran Bendavid, Bianca Mulaney, Neeraj Sood, Soleil Shah, Emilia Ling,
Rebecca Bromley-Dulfano, Cara Lai, Zoe Weissberg, Rodrigo Saavedra,
James Tedrow, et al. Covid-19 antibody seroprevalence in santa clara
county, california. MedRxiv, 2020.
[6] Andrea L Bertozzi, Elisa Franco, George Mohler, Martin B Short, and
Daniel Sledge. The challenges of modeling and forecasting the spread of
covid-19. arXiv preprint arXiv:2004.04741, 2020.
[7] Dimitris Bertsimas. Mit covidanalytics, May 2020. covidanalytics.io [On-
line; accessed 24-May-2020].
[8] Andrew Bray, Ka Wong, Christopher D Barr, Frederic Paik Schoenberg,
et al. Voronoi residual analysis of spatial point process models with appli-
cations to california earthquake forecasts. The Annals of Applied Statistics,
8(4):2247–2267, 2014.
[9] Simon Cauchemez, Pierre-Yves Boëlle, Christl A Donnelly, Neil M Fergu-
son, Guy Thomas, Gabriel M Leung, Anthony J Hedley, Roy M Anderson,
and Alain-Jacques Valleron. Real-time estimates in early detection of sars.
Emerging infectious diseases, 12(1):110, 2006.
[10] Wen-Hao Chiang, Xueying Liu, and George Mohler. Hawkes process mod-
eling of covid-19 with mobility leading indicators and spatial covariates.
medRxiv, 2020.
[11] Gerardo Chowell and Hiroshi Nishiura. Transmission dynamics and control
of ebola virus disease (evd): a review. BMC medicine, 12(1):196, 2014.
[12] Damian Clancy, Philip D O’Neill, et al. Bayesian estimation of the basic
reproduction number in stochastic epidemic models. Bayesian Analysis,
3(4):737–757, 2008.
[13] Robert Alan Clements, Frederic Paik Schoenberg, and Danijel Schorlem-
mer. Residual analysis methods for space-time point processes with appli-
cations to earthquake forecast models in california. The Annals of applied
statistics, pages 2549–2571, 2011.

13
[14] Robert Alan Clements, Frederic Paik Schoenberg, and Alejandro Veen.
Evaluation of space–time point process models using super-thinning. En-
vironmetrics, 23(7):606–616, 2012.

[15] Coronavirus COVID. Global cases by the center for systems science and en-
gineering (csse) at johns hopkins university. Coronavirus Resource Center.
March, 17:2020, 19.
[16] Daryl J Daley and D Vere Jones. An Introduction to the Theory of Point
Processes: Elementary Theory of Point Processes. Springer, 2003.
[17] Andrew Daw and Jamol Pender. The queue-hawkes process: Ephemeral
self-excitement. arXiv preprint arXiv:1811.04282, 2018.
[18] Michael Day. Covid-19: identifying and isolating asymptomatic people
helped eliminate virus in italian village. BMJ, 368:m1165, 2020.

[19] Chris Dye and Nigel Gay. Modeling the sars epidemic. Science,
300(5627):1884–1885, 2003.
[20] Neil D Evans, Lisa J White, Michael J Chapman, Keith R Godfrey, and
Michael J Chappell. The structural identifiability of the susceptible in-
fected recovered model with seasonal forcing. Mathematical biosciences,
194(2):175–197, 2005.
[21] CP Farrington, MN Kanaan, and NJ Gay. Branching process models for
surveillance of infectious diseases controlled by mass vaccination. Biostatis-
tics, 4(2):279–295, 2003.

[22] Neil Ferguson, Daniel Laydon, Gemma Nedjati Gilani, Natsuko Imai, Kylie
Ainslie, Marc Baguelin, Sangeeta Bhatia, Adhiratha Boonyasiri, ZULMA
Cucunuba Perez, Gina Cuomo-Dannenburg, et al. Report 9: Impact of
non-pharmaceutical interventions (npis) to reduce covid19 mortality and
healthcare demand. 2020.

[23] Seth Flaxman, Swapnil Mishra, Axel Gandy, H Juliette T Unwin, He-
len Coupland, Thomas A Mellan, Harrison Zhu, Tresnia Berah, Jeffrey W
Eaton, Pablo NP Guzman, et al. Estimating the number of infections and
the impact of non-pharmaceutical interventions on covid-19 in european
countries: technical description update. arXiv preprint arXiv:2004.11342,
2020.

[24] Institute for Health Metrics and Evaluation (IHME). Imhe covid-19 predic-
tions, May 2020. covid19.healthdata.org [Online; accessed 24-May-2020].
[25] Gianluca Frasso and Philippe Lambert. Bayesian inference in an extended
seir model with nonparametric disease transmission rate: an application to
the ebola epidemic in sierra leone. Biostatistics, 17(4):779–792, 2016.
[26] Keith R Godfrey and Michael J Chapman. Identifiability and indistin-
guishability of linear compartmental models. Mathematics and Computers
in Simulation, 32(3):273–295, 1990.

14
[27] Joshua Seth Gordon, Robert Alan Clements, Frederic Paik Schoenberg,
and Danijel Schorlemmer. Voronoi residuals and other residual analyses
applied to csep earthquake forecasts. Spatial Statistics, 14:133–150, 2015.
[28] Nicholas C Grassly and Christophe Fraser. Mathematical models of in-
fectious disease transmission. Nature Reviews Microbiology, 6(6):477–487,
2008.
[29] Jia Gu, Han Yan, Ya Huang, Yu Zhu, Hao Sun, Xin Zhang, Yu Wang,
Yumou Qiu, and Song Chen. Better strategies for containing covid-19
epidemics—a study of 25 countries via an extended varying coefficient seir
model. medRxiv, 2020.
[30] Quanquan Gu. Ucla statistical machine learning lab, June 2020.
covid19.uclaml.org [Online; accessed 19-June-2020].
[31] Youyang Gu. Covid-19 projections using machine learning, May 2020.
covid19-projections.com [Online; accessed 24-May-2020].
[32] Alex Perkins Guido Espana. Notre dame mobility, June 2020.
github.com/TAlexPerkins/covid19 NDmobility forecasting [Online; ac-
cessed 19-June-2020].
[33] Alan G Hawkes and David Oakes. A cluster process representation of a
self-exciting process. Journal of Applied Probability, 11(3):493–503, 1974.
[34] Nicolas Hengartner and Paul Fenimore. Quantifying model form uncer-
tainty of epidemic forecasting models from incidence data. Online Journal
of Public Health Informatics, 10(1), 2018.
[35] P Hernández, C Pena, A Ramos, and JJ Gómez-Cadenas. A simple formu-
lation of non-markovian seir. arXiv preprint arXiv:2005.09975, 2020.
[36] New Jersey Covid-19 Information Hub. How is the state using data to
make decisions and slow thespread of covid-19, May 2020. covid19.nj.gov
[Online; accessed 16-May-2020].
[37] Nicholas P Jewell, Joseph A Lewnard, and Britta L Jewell. Predictive
mathematical models of the covid-19 pandemic: underlying principles and
value of projections. Jama, 323(19):1893–1894, 2020.
[38] Park J. Kaplan A.M. and Schoenberg F. Nonparametric estimation of
recursive point processes with application to mumps in pennsylvania, 2020.
Submitted May 20, 2020.
[39] J Daniel Kelly, Junhyung Park, Ryan J Harrigan, Nicole A Hoff, Sarita D
Lee, Rae Wannier, Bernice Selo, Mathias Mossoko, Bathe Njoloko, Emile
Okitolonda-Wemakoy, et al. Real-time predictions of the 2018–2019 ebola
virus disease outbreak in the democratic republic of the congo using hawkes
point process models. Epidemics, 28:100354, 2019.
[40] William Ogilvy Kermack and Anderson G McKendrick. A contribution
to the mathematical theory of epidemics. Proceedings of the royal society
of london. Series A, Containing papers of a mathematical and physical
character, 115(772):700–721, 1927.

15
[41] Simas Kucinskas. Tracking r of covid-19. Available at SSRN 3581633, 2020.
[42] Los Alamos National Laboratory. Covid-19 confirmed and forecasted case
data, May 2020. covid-19.bsvgateway.org [Online; accessed 24-May-2020].
[43] Phenyo E Lekone and Bärbel F Finkenstädt. Statistical inference in a
stochastic epidemic seir model with control intervention: Ebola as a case
study. Biometrics, 62(4):1170–1177, 2006.
[44] Joseph Chadi Lemaitre, Kyra H Grantz, Joshua Kaminsky, Hannah R
Meredith, Shaun A Truelove, Stephen A Lauer, Lindsay T Keegan, Sam
Shah, Josh Wills, Kathryn Kaminsky, et al. A scenario modeling pipeline
for covid-19 emergency planning. medRxiv, 2020.
[45] Simon A Levin and Viggo Andreasen. Mathematical models of infectious
diseases. Forefronts, 2(8):4–6, 1986.
[46] PA W Lewis and Gerald S Shedler. Simulation of nonhomogeneous poisson
processes by thinning. Naval research logistics quarterly, 26(3):403–413,
1979.
[47] Alun L Lloyd. Destabilization of epidemic models with the inclusion of
realistic distributions of infectious periods. Proceedings of the Royal Society
of London. Series B: Biological Sciences, 268(1470):985–993, 2001.
[48] Thomas Lumley. Counting rare things is hard, April 2020. statschat.org.nz
[Online; accessed 16-May-2020].
[49] David Marsan and Olivier Lengline. Extending earthquakes’ reach through
cascading. Science, 319(5866):1076–1079, 2008.
[50] Stefano Merler, Marco Ajelli, Laura Fumanelli, and Alessandro Vespig-
nani. Containing the accidental laboratory escape of potential pandemic
influenza viruses. BMC medicine, 11(1):252, 2013.
[51] Sebastian Meyer, Leonhard Held, and Michael Höhle. Spatio-temporal
analysis of epidemic phenomena using the r package surveillance. arXiv
preprint arXiv:1411.0416, 2014.
[52] Matthew D. Parno Ian D. Detwiller Matthew W. Farthing William
P. England Glover E. George Michael L. Mayo, Michael A. Row-
land. Us army engineer research and development center, June 2020.
github.com/reichlab/covid19-forecast-hub [Online; accessed 19-June-2020].
[53] George Mohler et al. Modeling and estimation of multi-source clustering in
crime and security data. The Annals of Applied Statistics, 7(3):1525–1539,
2013.
[54] George Mohler, Frederic Schoenberg, Martin B Short, and Daniel Sledge.
Analyzing the impacts of public policy on covid-19 transmission in indiana:
The role of model and dataset selection.
[55] Pierre Montagnon. A stochastic sir model on a graph with epidemiological
and population dynamics occurring over the same time scale. Journal of
mathematical biology, 79(1):31–62, 2019.

16
[56] Yosihiko Ogata. On lewis’ simulation method for point processes. IEEE
transactions on information theory, 27(1):23–31, 1981.

[57] Yosihiko Ogata. Statistical models for earthquake occurrences and residual
analysis for point processes. Journal of the American Statistical association,
83(401):9–27, 1988.
[58] Yosihiko Ogata. Space-time point-process models for earthquake occur-
rences. Annals of the Institute of Statistical Mathematics, 50(2):379–402,
1998.
[59] Yosihiko Ogata et al. The asymptotic behaviour of maximum likelihood es-
timators for stationary point processes. Annals of the Institute of Statistical
Mathematics, 30(1):243–261, 1978.
[60] Philip D O’Neill. Introduction and snapshot review: relating infectious
disease transmission models to data. Statistics in medicine, 29(20):2069–
2077, 2010.
[61] Dave Osthus, Kyle S Hickmann, Petruţa C Caragea, Dave Higdon, and
Sara Y Del Valle. Forecasting seasonal influenza with a state-space sir
model. The annals of applied statistics, 11(1):202, 2017.

[62] Marie V Ozanne, Grant D Brown, Jacob J Oleson, Iraci D Lima, Jose W
Queiroz, Selma MB Jeronimo, Christine A Petersen, and Mary E Wil-
son. Bayesian compartmental model for an infectious disease with dynamic
states of infection. Journal of applied statistics, 46(6):1043–1065, 2019.

[63] An Pan, Li Liu, Chaolong Wang, Huan Guo, Xingjie Hao, Qi Wang, Jiao
Huang, Na He, Hongjie Yu, Xihong Lin, et al. Association of public health
interventions with the epidemiology of the covid-19 outbreak in wuhan,
china. Jama, 2020.
[64] Junhyung Park, Adam W Chaffee, Ryan J Harrigan, and Frederic Paik
Schoenberg. A non-parametric hawkes model of the spread of ebola in west
africa, 2018.
[65] Junhyung Park, Frederic Paik Schoenberg, Andrea L Bertozzi, and P Jef-
frey Brantingham. Investigating clustering and violence interruption in
gang-related violent crime data using spatial-temporal point processes with
covariates. 2019.

[66] Spencer Carran Sarah Cobey Katelyn Gostic Lauren McGough Sylvia Ran-
jeva Frank Wen Phil Arevalo, Ed Baskerville. Forecasting sars-cov-2 dy-
namics for the state of illinois, June 2020. github.com/cobeylab/covid IL
[Online; accessed 19-June-2020].
[67] Gergo Pinter, Imre Felde, Amir Mosavi, Pedram Ghamisi, and Richard
Gloaguen. Covid-19 pandemic prediction for hungary; a hybrid machine
learning approach. A Hybrid Machine Learning Approach (May 2, 2020),
2020.
[68] Jakob Gulddahl Rasmussen. Bayesian inference for hawkes processes.
Methodology and Computing in Applied Probability, 15(3):623–642, 2013.

17
[69] Alex Reinhart et al. A review of self-exciting spatio-temporal point pro-
cesses and their applications. Statistical Science, 33(3):299–318, 2018.

[70] Marian-Andrei Rizoiu, Swapnil Mishra, Quyu Kong, Mark Carman, and
Lexing Xie. Sir-hawkes: linking epidemic models and hawkes processes to
model diffusions in finite populations. In Proceedings of the 2018 World
Wide Web Conference, pages 419–428, 2018.
[71] Kimberlyn Roosa and Gerardo Chowell. Assessing parameter identifiability
in compartmental dynamic models using a computational approach: appli-
cation to infectious disease transmission models. Theoretical Biology and
Medical Modelling, 16(1):1, 2019.
[72] Jay Boice Ryan Best. Where the latest covid-19 models think we’re headed
— and why they disagree, May 2020. fivethirtyeight.com [Online; posted
24-May-2020].
[73] Timothy Sauer, Tyrus Berry, Donald Ebeigbe, Michael M Norton, Andrew
Whalen, and Steven J Schiff. Identifiability of infection model parameters
early in an epidemic. medRxiv, 2020.
[74] Frederic Paik Schoenberg. A note on the consistent estimation of spatial-
temporal point process parameters. Statistica Sinica, pages 861–879, 2016.
[75] Frederic Paik Schoenberg. Nonparametric estimation of variable produc-
tivity hawkes processes. arXiv preprint arXiv:2003.08858, 2020.
[76] Frederic Paik Schoenberg, Marc Hoffmann, and Ryan J Harrigan. A recur-
sive point process model for infectious diseases. Annals of the Institute of
Statistical Mathematics, 71(5):1271–1287, 2019.
[77] Danijel Schorlemmer, Maximilian J Werner, Warner Marzocchi, Thomas H
Jordan, Yosihiko Ogata, David D Jackson, Sum Mak, David A Rhoades,
Matthew C Gerstenberger, Naoshi Hirata, et al. The collaboratory for the
study of earthquake predictability: achievements and priorities. Seismolog-
ical Research Letters, 89(4):1305–1313, 2018.
[78] Neeraj Sood, Paul Simon, Peggy Ebner, Daniel Eichner, Jeffrey Reynolds,
Eran Bendavid, and Jay Bhattacharya. Seroprevalence of sars-cov-2–
specific antibodies among adults in los angeles county, california, on april
10-11, 2020. JAMA, 2020.

[79] Balaji Srinivasan. Peer review of “covid-19 antibodyseroprevalence in santa

clara county, california”, April 2020. medium.com/@balajis [Online; ac-
cessed 16-May-2020].
[80] Gabriele Stabile and Giovanni Luca Torrisi. Risk processes with non-
stationary hawkes claims arrivals. Methodology and Computing in Applied
Probability, 12(3):415–429, 2010.
[81] Chandini Jain Vishal Tomar. Auquan data science, June 2020. covid19-
infection-model.auquan.com [Online; accessed 19-June-2020].

18
[82] Jacco Wallinga and Peter Teunis. Different epidemic curves for severe acute
respiratory syndrome reveal similar impacts of control measures. American
Journal of epidemiology, 160(6):509–516, 2004.

[83] Eric Walter and Yves Lecourtier. Unidentifiable compartmental models:

what to do? Mathematical biosciences, 56(1-2):1–25, 1981.
[84] Helen J Wearing, Pejman Rohani, and Matt J Keeling. Appropriate models
for the management of infectious diseases. PLoS medicine, 2(7), 2005.

[85] Lee Worden, Rae Wannier, Nicole A Hoff, Kamy Musene, Bernice Selo,
Mathias Mossoko, Emile Okitolonda-Wemakoy, Jean Jacques Muyembe-
Tamfum, George W Rutherford, Thomas M Lietman, et al. Real-time
projections of epidemic transmission and estimation of vaccination impact
during an ebola virus disease outbreak in the eastern region of the demo-
cratic republic of congo. arXiv preprint arXiv:1811.01175, 2018.
[86] Teresa Yamana, Sen Pei, and Jeffrey Shaman. Projection of covid-19 cases
and deaths in the us as individual states re-open may 4, 2020. medRxiv,
2020.
[87] Ah Sung Yang. Modeling the Transmission Dynamics of Pertussis Using
Recursive Point Process and SEIR model. PhD thesis, UCLA, 2019.
[88] Chong You, Yuhao Deng, Wenjie Hu, Jiarui Sun, Qiushi Lin, Feng Zhou,
Cheng Heng Pang, Yuan Zhang, Zhengchao Chen, and Xiao-Hua Zhou.
Estimation of the time-varying reproduction number of covid-19 outbreak
in china. International Journal of Hygiene and Environmental Health, page
113555, 2020.
[89] Baichuan Yuan. Multivariate hawkes processes for real-time covid-19 death
forecasting, 2020. Preprint submitted to Journal of LATEX Templates.
[90] J Douglas Zechar, Danijel Schorlemmer, Maximilian J Werner, Matthew C
Gerstenberger, David A Rhoades, and Thomas H Jordan. Regional earth-
quake likelihood models i: First-order results. Bulletin of the Seismological
Society of America, 103(2A):787–798, 2013.
[91] Jiancang Zhuang, Yosihiko Ogata, and David Vere-Jones. Analyzing earth-
quake clustering features by using stochastic reconstruction. Journal of
Geophysical Research: Solid Earth, 109(B5), 2004.

19