The Annals of Applied Statistics

2008, Vol. 2, No. 2, 601–623

DOI: 10.1214/08-AOAS164
c Institute of Mathematical Statistics, 2008

FORECASTING TIME SERIES OF INHOMOGENEOUS POISSON

PROCESSES WITH APPLICATION TO CALL CENTER
WORKFORCE MANAGEMENT
arXiv:0807.4071v1 [stat.AP] 25 Jul 2008

By Haipeng Shen1 and Jianhua Z. Huang2

University of North Carolina at Chapel Hill and Texas A&M University
We consider forecasting the latent rate profiles of a time series
of inhomogeneous Poisson processes. The work is motivated by op-
erations management of queueing systems, in particular, telephone
call centers, where accurate forecasting of call arrival rates is a cru-
cial primitive for efficient staffing of such centers. Our forecasting
approach utilizes dimension reduction through a factor analysis of
Poisson variables, followed by time series modeling of factor score
series. Time series forecasts of factor scores are combined with factor
loadings to yield forecasts of future Poisson rate profiles. Penalized
Poisson regressions on factor loadings guided by time series forecasts
of factor scores are used to generate dynamic within-process rate up-
dating. Methods are also developed to obtain distributional forecasts.
Our methods are illustrated using simulation and real data. The em-
pirical results demonstrate how forecasting and dynamic updating of
call arrival rates can affect the accuracy of call center staffing.

1. Introduction. Queueing models usually assume that arrivals to a queue-

ing system follow an inhomogeneous Poisson process. For efficient manage-
ment of such a system, accurate prediction of the underlying random rate
function of the inhomogeneous Poisson process is of primary importance.
For example, various staffing models need the forecasted rate as one es-
sential input [Gans, Koole and Mandelbaum (2003)]. Both the latency and
uncertainty of the rate makes such a prediction problem interesting and dif-
ficult. The current paper proposes a methodology for forecasting the arrival

Received July 2007; revised January 2008.

1
Supported in part by NSF Grant DMS-06-06577.
2
Supported in part by NSF Grant DMS-06-06580 and National Cancer Institute Grant
CA57030.
Key words and phrases. Dimension reduction, factor model, forecast updating, penal-
ized likelihood, queueing systems, service engineering, singular value decomposition, vector
time series.

This is an electronic reprint of the original article published by the

Institute of Mathematical Statistics in The Annals of Applied Statistics,
2008, Vol. 2, No. 2, 601–623. This reprint differs from the original in pagination
and typographic detail.
1
2 H. SHEN AND J. Z. HUANG

rate functions of a time series of inhomogeneous Poisson processes. The pro-

posal combines the Poisson process assumption with ideas of data-driven
dimension reduction as well as time series forecasting.
Our primary motivation and application come from call centers, where
customers contact their service providers through telephones and wait in
tele queues. The daily call arrivals to a call center form a point process that
can be modeled as an inhomogeneous Poisson process [Brown et al. (2005)].
Motivated by call center operations management, two forecasting problems
will be addressed: to forecast future rate functions days or weeks ahead,
and to update existing forecasts using additional information during a day.
Throughout the paper, we will use call centers to facilitate our presenta-
tion; however, the developed methodology undoubtedly has a much wider
application spectrum.
In this paper a daily call arrival process is modeled as a Poisson process
with a stochastic intensity. In other words, a call arrival process is a Cox
process [Cox (1955)], also known as a doubly stochastic Poisson process.
However, instead of focusing on a single daily arrival process, we consider a
time series of daily arrival processes. What makes our paper different from
the literature of Cox processes is the time series dependency of the daily
intensity functions. Such dependency makes it feasible to perform time series
forecasts of the intensity functions using historical data.
Call centers have become a primary contact point between companies and
their customers in modern business [Mandelbaum (2006),
Aksin, Armony and Mehrotra (2007)]. The forecasting of future call arrival
rates is critical for efficient agent staffing and scheduling, but the current
practice is still in its infancy [Gans, Koole and Mandelbaum (2003)]. To
close the gap, recently more research efforts have been devoted to call cen-
ter forecasting. Among them, Weinberg, Brown and Stroud (2007) propose
a model-based approach where a two-way multiplicative Bayesian model for
inhomogeneous Poisson processes is used to predict the Poisson rate func-
tions; while the data-driven approach of Shen and Huang (2008) combines
dimension reduction through singular value decomposition (SVD) with time
series forecasting, and provides a more robust solution to the problem that
does not rely on the multiplicative model assumption. To the best of our
knowledge, these are the only two papers that have dealt with the problem
of dynamic intraday updating.
However, instead of modeling the arrival counts directly, both approaches
model the square-root transformed counts, which are approximately nor-
mally distributed for Poisson counts [Brown et al. (2007)]. The data-driven
approach in Shen and Huang (2008) does not make use of the Poisson as-
sumption, and aims at forecasting future counts. To the extent that some
stochastic model assumption is reasonable for the counts, such as inhomoge-
neous Poisson processes, incorporating it appropriately should improve the
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 3

statistical efficiency of the method. Note that forecasting the latent arrival
rates of the Poisson processes is of primary interest from the management
point of view (see Section 4.1). Although the point forecasts for counts
and rates are the same, their distributional forecasts are different. Fur-
thermore, neither Weinberg, Brown and Stroud (2007) nor Shen and Huang
(2008) has considered the managerial consequences of different forecasts of
the rates of the Poisson processes.
The purpose of this paper is to develop a forecasting method that utilizes
the widely used Poisson assumption in queueing models. More precisely,
we consider a time series of inhomogeneous Poisson processes and use the
historical data to forecast the future rate functions of the Poisson processes.
In the call center application, for example, each Poisson process models the
customer call arrivals during one day in a sequence of days. Since the rate
function of each Poisson process is approximately constant during short time
intervals (such as quarter hours in our data example in Section 4), the data
for a given day can be aggregated into a vector of Poisson random variables
with varying rates, counting the call arrivals during these intervals. The
collection of such rates for a given day is referred as the rate profile and the
corresponding counts form the count profile. The unobservable rate profiles
form a vector time series, and our goal is to forecast future rate profiles
using the observed historical count profiles.
The dimensionality of the vector time series of count profiles is usually
high. Thus, we propose to first reduce dimension by building a factor model
on the hidden rate profiles. The proposed dimension reduction technique
can be viewed as an extension of singular value decomposition to Poisson
data, or Poisson SVD. According to the factor model, every rate profile can
be expressed as a linear combination of a few factors. Factor loadings and
scores can be computed by fitting generalized linear models (GLM) [Dobson
(2001)]. An alternating maximum likelihood algorithm is employed that fits
marginal Poisson regression models by alternately fixing the factor loadings
and scores. The algorithm is closely related to the alternating least squares
algorithm for computing singular value decomposition [Gabriel and Zamir
(1979)].
After the factor analysis, forecasting future rate profiles reduces to fore-
casting time series of factor scores. Univariate time series models can be
built for each factor score series to produce time series forecasts of the rate
profile. To achieve within-process dynamic rate updating, we propose to ef-
fectively combine the information contained in both the historical processes
and the current process, through maximizing a penalized likelihood crite-
rion. The criterion appropriately balances two measures, the goodness-of-fit
of the factor model to the early segment of the current process and the devi-
ation of the factor scores from their forecasts provided by the interday time
series forecasting.
4 H. SHEN AND J. Z. HUANG

Our forecasting approach enjoys the benefit of both the model-driven ap-
proach of Weinberg, Brown and Stroud (2007) and the data-driven approach
of Shen and Huang (2008). It directly models and forecasts the rate profiles
of inhomogeneous Poisson processes. [Some earlier analysis is briefly doc-
umented in Shen, Huang and Lee (2007).] Making no rigid model assump-
tions on the rate profiles, our approach is robust, as shown in two simulation
studies. In this paper we also illustrate how rate forecasting and dynamic
updating benefits the staffing decision in a real call center application, which
has not been considered before in the literature.
The data we analyze and use for building forecasting models have an inter-
esting two-way structure: there are both day-to-day variation/dependence
and within-day variation/dependence. This specific structure differentiates
our work with the related literature on longitudinal data analysis and dy-
namic factor models, where only one-way structure is present. Longitudinal
data analysis [Diggle et al. (2002)] concerns repeated measurements from
many subjects. While there are typically within subject correlation, the mea-
surements from different subjects are usually independent. Dynamic factor
models are commonly used in econometrics for analyzing economic time
series [Stock and Watson (2005)]. Since economic variables do not have a
natural ordering as the one present in our within-day call volume profiles,
intraday dynamic updating is not of concern in the dynamic factor models
literature. Moreover, we are unaware of any work on dynamic factor models
for Poisson count data.
The rest of the paper is structured as follows. Section 2 describes the factor
model as well as the alternating maximum likelihood estimation algorithm
for model estimation. Our forecasting approach is described in detail in
Section 3. The proposed method is applied in Section 4 to the call center
arrival data analyzed by Shen and Huang (2008). Some background on call
center staffing is briefly discussed in Section 4.1. Section 4.2 fits the factor
model to the real data. Two simulation studies are presented in Section 4.3.2
to illustrate the robustness of our method and its performance on forecasting
hidden rate profiles. The effect of rate forecasting on staffing level is discussed
in Section 4.3.3. In Section 4.4 several intraday rate updating methods are
used to update the related staffing levels. We conclude in Section 5 with
some discussion.

2. Dimension reduction of Poisson variables.

2.1. A Poisson factor model. Consider a time series of inhomogeneous

Poisson processes that are observed over the same time interval. Following
common practice, we assume that the arrival rate function of each Poisson
process can be well approximated as being piecewise constant over short
time intervals. Let Y = (yij ) be an n × m matrix that records the arrival
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 5

counts from n such processes with each process being aggregated into m time
intervals. Furthermore, we assume that yij is a Poisson random variable with
rate λij . For notational purpose, we let Λ = (λij ) denote the n × m hidden
Poisson rate matrix. The ith row of Λ, denoted as λT(i) = (λi1 , . . . , λim ), is the
rate profile of the ith process. Correspondingly, the ith row of Y, denoted
T = (y , . . . , y ), is the count profile of the ith process. Note that the
as y(i) i1 im
rows of Λ and Y are time ordered.
The rate profiles {λ(i) }ni=1 then form a vector-valued time series taking
values in Rm . It is of interest to forecast future rate profiles λ(n+h) (h > 0).
However, the difficulty is that the rate profiles are unobservable. The idea is
to build a time series forecasting model on the corresponding count profiles
{y(i) }ni=1 , which can then be used to forecast λ(n+h) .
In practice, the dimensionality of the vector time series {y(i) } is usually
so large that it is infeasible to directly apply the classical vector autoregres-
sive and moving average (VARMA) models [Reinsel (1997)]. For example,
the call center application in Section 4 has m = 68, and the application of
Weinberg, Brown and Stroud (2007) has m = 169. This calls for the neces-
sity of dimension reduction. In addition, each y(i) is a vector of Poisson
random variables with a positive rate vector λ(i) . The Poisson nature of the
data needs to be accounted for appropriately.
For dimension reduction of Poisson variables, we consider the following
K-factor model,

y(i) ∼ Poisson(λ(i) ), i = 1, . . . , n,
(2.1)
g(λ(i) ) = βi1 f1 + · · · + βiK fK ≡ Fβ(i) ,
where β (i) = (βi1 , . . . , βiK )T is the K-vector of underlying factor scores for
the ith rate profile, Fm×K = (f1 , . . . , fK ) contains the K factor loading vec-
tors in Rm , and the transformation g is a link function suitable for Poisson
variables such as the logarithmic or square root function. Here and through-
out the paper, application of the link function g to a vector or matrix is un-
derstood as a componentwise operation. See McCullagh and Nelder (1989)
and Dobson (2001) for more discussion on generalized linear models (GLM)
and, in particular, specification of link functions.
Denote the factor matrix as Bn×K = (β T(1) , . . . , β T(n) )T . Then the factor
model for the transformed rate profiles can be written in the following matrix
form:

Y ∼ Poisson(Λ),
(2.2)
g(Λ) = BFT .
The factor model summarizes the transformed rate profiles using K com-
mon factors, denoted as β 1 , . . . , β K , which are the columns of B. Note that
each β k is a time series of factor scores. How to build time series models on
these series for forecasting will be discussed in Section 3. In practice, one
6 H. SHEN AND J. Z. HUANG

would assume K is usually much smaller than n or m, resulting in a consid-

erable amount of dimension reduction. The above model (2.2) may appear
similar to dynamic factor models (DFM) commonly used in econometrics
[Stock and Watson (2005)]. However, there are two major differences. First,
our loadings for the same process are time dependent and correlated, a useful
feature for within-process updating discussed later in Section 3.2, while such
dependence is missing for DFM because economic variables usually have no
natural ordering. Second, our model deals with Poisson variables, which to
the best of our knowledge have not been analyzed in the DFM literature
As it stands, the factor model (2.2) is not identifiable. To achieve identifi-
ability, one can require either fkT fk′ = δkk′ or β Tk β k′ = δkk′ , where δkk′ is the
Kronecker delta that equals 1 for k = k′ and 0 otherwise. See Section 2.2 for
more details.

2.2. An alternating maximum likelihood algorithm. Assuming indepen-

dence of data, one can estimate the factors B and the loadings F using
maximum likelihood. However, since both B and F are unknown, direct
maximization of the likelihood function is complicated. Below, we propose
an iterative procedure that relies on alternating optimization over B and F.
Fixing F, the factor model (2.1) represents n standard Poisson regression
models with the count profiles y(i) as responses and F as the common matrix
of covariates. On the other hand, the same factor model can be written in
terms of the columns of Y (denotes by {yj }) as

yj ∼ Poisson(λj ), j = 1, . . . , m,
(2.3)
g(λj ) = Bf(j) .
Fixing B, the model consists of m Poisson regressions with B as the common
regressor matrix.
The above discussion suggests the following iterative algorithm:
1. Initialize: Apply singular value decomposition (SVD) to g(Y) to obtain
the SVD components {U, V, S}, and set Bold = [s1 u1 , . . . , sK uK ] and
Fold = [v1 , . . . , vK ], where uk is the kth column of the left singular vector
matrix U, vk is the kth column of the right singular vector matrix V
and sk is the kth diagonal element of the ordered diagonal singular value
matrix S.
2. Update:
(a) Fit the Poisson regression model (2.1) with F replaced by Fold , and
obtain an updated factor score matrix Bnew ;
(b) Fit the Poisson regression model (2.3) with B replaced by Bnew , and
obtain an updated factor loading matrix Fnew ;
(c) Apply SVD to the product matrix Bnew FTnew , and denote the ob-
tained SVD components by {Unew , Vnew , Snew }.
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 7

3. Repeat Step 2 with Bold and Fold replaced by the first K columns of
Unew Snew and Vnew respectively until convergence.
The algorithm is presented in such a way that the factor loadings are
orthonormal in that FT F = I. Alternatively, we can modify the algorithm
to make the score vectors satisfy BT B = I. To achieve that, we can set B to
be the first K columns of the left singular vector matrix U, and F to be the
corresponding columns of SV. This alternative formulation turns out to be
more useful for the penalized dynamic updating approach to be discussed
in Section 3.2.
Our Poisson factor model extends singular value decomposition (SVD) to
Poisson count data. The above alternating maximum likelihood algorithm is
an extension of the alternating least squares algorithm for computing SVD
[Gabriel and Zamir (1979)].

2.3. Selection of the number of factors. In practice, one needs to decide

on K, the number of underlying factors. Since we are in a forecasting con-
text, one possibility is to compare out-of-sample forecasting performance
for different choices of K. Note that the forecasting performance measure
can only be calculated using the counts, as the rates are not observed. Our
experience suggests that increasing K tends to improve forecasting perfor-
mance, but the improvement becomes minimal beyond a particular value of
K, denoted as K ∗ , as additional factors start to fit the noise in the data.
This K ∗ can be used as the “true” number of factors.
Alternatively, one can extend the idea of scree plots from principal com-
ponents analysis (PCA) to the current context. In PCA, the eigenvalues of
the sample covariance matrix are ordered decreasingly and plotted against
their indices in a scree plot. An “elbow” in the scree plot, at which the slopes
of lines joining the plotted points change from “steep” to “shallow,” is usu-
ally used to locate the number of significant principal components (PCs).
To extend scree plots to Poisson variables, note that the eigenvalues equal
to the reductions of residual sum of squares in reconstruction of the data
using PCs as more PCs are added.
In GLM, a commonly used goodness-of-fit measure is the so called de-
viance [Dobson (2001)]. Consider the K-factor model (2.1), and let B b and
b denote the maximum likelihood estimates of the parameters. Denote the
F
corresponding maximized log-likelihood as
b
l(Y; Λ = Λ) b = g −1 (B
with Λ bFb T ).

In addition, consider the saturated model where the rate matrix Λ is esti-
mated as the count matrix Y, and denote the corresponding log-likelihood
as l(Y; Λ = Y). The deviance of the model (2.1) is then defined as
b
2{l(Y; Λ = Y) − l(Y; Λ = Λ)}.
8 H. SHEN AND J. Z. HUANG

Similar to the residual sum of squares, the deviance decreases as K increases.

We propose to calculate the reduction of deviance of the model (2.1) for an
increasing sequence of K, and plot it against K in a deviance reduction plot.
The number of necessary factors can then be suggested by an “elbow” in the
plot. See Figure 1 for an illustration of this idea, which shows that the steep
drop of the deviance slows down significantly beyond the elbow around four
or five factors.

3. Forecasting.

3.1. Forecasting future rate profile. Given the historical count data y(i) ,
i = 1, . . . , n, and assuming the latent Poisson rates satisfy (2.1), consider
forecasting the future rate profile λ(n+h) (h > 0). According to (2.1),
λ(n+h) = g−1 (βn+h,1 f1 + · · · + βn+h,K fK ).
Since the factor loading vectors f1 , . . . , fK can be obtained from historical
data using the alternating maximum likelihood (AML) algorithm, forecast-
ing the m-dimensional rate profile λ(n+h) reduces to forecasting the K-
dimensional vector of factor scores β (n+h) = {βn+h,1 , . . . , βn+h,K }T . Suppose
we can obtain such a forecast β̂ TS
(n+h) , then a forecast of λ(n+h) follows as

λ̂TS −1 TS TS
(n+h) = g (β̂n+h,1 f1 + · · · + β̂n+h,K fK ),

TS
where β̂n+h,k is a time series forecast of βn+h,k , 1 ≤ k ≤ K. Because the
count profile y(n+h) has a Poisson distribution with rate λ(n+h) , the point
forecast of y(n+h) is the same as λ̂TS(n+h) .
In addition to the factors, the AML algorithm also produces the matrix
B of factor scores, whose row vectors β (1) , . . . , β (n) form a K-dimension
time series. This vector time series is subject to time series modeling to
generate desired forecast β̂ TS
(n+h) . Note that the dimension K of this vector
time series is much smaller than the dimension m of the original observa-
tion y(i) . However, we propose to model the score time series for each factor
separately using univariate time series models such as exponential smooth-
ing, ARIMA models [Box, Jenkins and Reinsel (1997)] or state space models
[Harvey (1990)]. An appropriate model can be decided on by analyzing his-
torical data. The rational for this separate univariate time series modeling
lies in the alternating estimation algorithm (Section 2.2). We observe that
the score series β k is orthogonal to score series β ′k for k 6= k′ . This lack of
contemporaneous correlation suggests that the cross-correlations at nonzero
lags are likely to be small. Hence, it suffices to forecast each β k separately.
To generate interval and distributional forecasts, we modify the bootstrap
approach described in Shen and Huang (2008). For a fixed k, we use the
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 9

fitted time series model for the series β k recursively, and bootstrap the
model errors from the fitted model to generate a sample of B forecasts
TS,b
{β̂n+h,k }, 1 ≤ b ≤ B. Then B forecasts of λ(n+h) can be obtained as
TS,b −1 TS,b TS,b
b
λ (n+h) = g (β̂n+h,1 f1 + · · · + β̂n+h,K fK ), b = 1, . . . , B,
which provides a distributional forecast for λ(n+h) . To obtain a distributional
forecast of the count profile, for any b, we randomly sample one count profile
forecast y TS,b
b (n+h) from Poisson distributions with rate λb TS,b . The sample of B
(n+h)
forecasts of y(n+h) then leads to its distributional forecast. Interval forecasts
can be obtained easily using quantiles of the distributional forecasts.

3.2. Dynamic within-process rate updating. In addition to the historical

count profiles y(1) , . . . , y(n) , sometimes one also observes the counts of the
(n + 1)th process during the first m0 time intervals. It is of interest to use
TS b
them to update λ (n+h) , the time series forecast obtained previously. Such up-
dating can reduce forecasting errors substantially [Weinberg, Brown and Stroud
(2007), Shen and Huang (2008)]. For call centers, the ability to change agent
schedules in response to the updating then yields operational benefits. For
example, to adjust for a revised forecast, call center managers may send
people home early or have them perform alternative tasks, or they may call
in part-time or work-from-home agents.
Denote the new observations from the (n + 1)th process collectively as
e
y(n+1) and the corresponding Poisson rate vector as λe(n+1) . Let y(n+1) l =
T
(yn+1,m0 +1 , . . . , yn+1,m) be the latter count profile for the (n + 1)th process
with the rate vector being λl(n+1) . For notational simplicity, we suppress
the dependence of y(n+1) e /λe(n+1) and y(n+1)
l /λl(n+1) on m0 in the following
discussion. There are two sets of information, {y(1) , . . . , y(n) } and y(n+1) e .
We are interested in making use of both to obtain an updated forecast for
λl(n+1) .
Using historical data, the alternating maximum likelihood algorithm in
Section 2.2 can be run to get estimates of the factor loading vectors f1 , . . . , fK
and the factor score series β 1 , . . . , β K . Time series models are then built on
TS , . . . , β̂ TS
β 1 , . . . , β K to generate forecasts β̂n+1,1 n+1,K . Therefore, just using the
historical count profiles, the time series forecast of λl(n+1) is given by

(3.1) λ̂l,TS −1 TS l TS l
n+1 = g (β̂n+1,1 f1 + · · · + β̂n+1,K fK ),

where fkl is the latter segment of fk after the initial m0 time intervals. How-
ever, this forecast does not utilize any new information contained in yn+1 e .
The new information can be incorporated using a Poisson regression as
we describe now. Suppose the factor model (2.1) holds for the rate profile
10 H. SHEN AND J. Z. HUANG

λ(n+1) , then we have the following expression,

(3.2) g(λn+1,j ) = βn+1,1 fj1 + · · · + βn+1,K fjK , j = 1, . . . , m.

Let Fe be an m0 × K matrix whose (j, k)th entry is fjk , 1 ≤ j ≤ m0 ,

1 ≤ k ≤ K, and let β (n+1) denote the vector (βn+1,1 , . . . , βn+1,K )T of factor
scores. The following Poisson regression model can then be built on y(n+1) e ,
( e
y(n+1) ∼ Poisson(λe(n+1) ),
(3.3)
g(λe(n+1) ) = Fe β (n+1) .

This suggests that we can forecast β (n+1) by maximum likelihood (ML),

solving the following optimization problem,
m0
X
min {λn+1,j − yn+1,j log(λn+1,j )} subject to g(λe(n+1) ) = Fe β (n+1) ,
β(n+1)
j=1

which minimizes the negative log-likelihood function. Note that the regressor
Fe in the Poisson regression is obtained using historical data.
However, the above Poisson regression approach ignores the time series
dependence among the rate profiles, present in the factor score series. To im-
prove the ML forecast, we propose to combine it with the time series forecast
of β (n+1) using a penalized Poisson regression. Specifically, we minimize the
following penalized likelihood criterion with respect to β (n+1) ,
m0
X 2
{λn+1,j − yn+1,j log(λn+1,j )} + ωkβ (n+1) − β̂ TS
(n+1) k
j=1
(3.4)
subject to g(λe(n+1) ) = Fe β (n+1) ,

where β̂ TS
(n+1) is a time series forecast based on the historical count pro-
files, and ω > 0 is a penalty parameter. To simplify our procedure, only one
penalty parameter is used. Thus, it is desirable for the K time series {β k } to
be roughly on the same scale. This can be achieved by requiring them to be
orthonormal in the alternating maximum likelihood algorithm (Section 2.2).
The penalized criterion (3.4) involves two terms: the first term measures
e
the goodness-of-fit of the model to y(n+1) in terms of likelihood, while the
second term penalizes a large departure from the time series forecast. Its
solution is a compromise between the two terms based on the size of ω, the
penalty parameter. In practice, ω can be selected based on the forecasting
performance on a rolling hold-out sample [Shen and Huang (2008)].
Below in Section 3.3, we describe an iterative re-weighted least squares
(IRLS) algorithm for minimizing (3.4) based on a quadratic approximation
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 11

of the objective function. The minimizer then gives us the penalized maxi-
mum likelihood (PML) forecast of β (n+1) . The PML forecast of λl(n+1) (and
l
y(n+1) ) is then given by

(3.5) λ̂l,PML −1 PML l PML l

n+1 = g (β̂n+1,1 f1 + · · · + β̂n+1,K fK ).

For distributional updates of λl(n+1) , we propose to use a bootstrap pro-

cedure similar to the one aforementioned in Section 3.1. First, by bootstrap-
ping the errors from the time series forecasting models for β (n+1) , we obtain
B time series forecasts β̂ TS,b PML,b
(n+1) , which in turn lead to B PML forecasts β̂ (n+1)
by minimizing the criterion (3.4). The derivation of the B PML forecasts can
be sped up considerably via a one-step updating approximation as discussed
at the end of Section 3.3. The B PML updates of λl(n+1) are constructed as
follows:
λ̂PML,b −1 PML,b PML,b
n+1,j = g (β̂n+1,1 fj1 + · · · + β̂n+1,K fjK ),

j = m0 + 1, . . . , m; b = 1, . . . , B.
The interval and density forecasts of λn+1,j are obtained using the empirical
distribution of λ̂PML,b
n+1,j , b = 1, . . . , B. To obtain a distributional update of
PML,b
the count profile, for b = 1, . . . , B, we randomly sample one forecast ŷn+1,j
of yn+1,j from the Poisson distribution with rate λ̂PML,b
n+1,j . The sample of B
forecasts of yn+1,j then leads to its interval and density forecasts.

3.3. An IRLS algorithm for penalized Poisson regression. Suppose we

have some initial estimate of β (n+1) denoted as β 0(n+1) , and the correspond-
ing estimate of λ(n+1) denoted as λ0(n+1) . The minimizing criterion (3.4) is
equivalent to
m0
X
C(β (n+1) ) ≡ {λn+1,j − yn+1,j log(λn+1,j )}
j=1
(3.6)
K
X
TS
+ω (βn+1,k − β̂n+1,k )2 .
k=1

Introducing λ0(n+1) , the criterion (3.6) can be re-written as

m0
X
C(β (n+1) ) = [(λn+1,j − λ0n+1,j )
j=1

− yn+1,j {log(λn+1,j ) − log(λ0n+1,j )}]

(3.7)
12 H. SHEN AND J. Z. HUANG

m0
X
+ {λ0n+1,j − yn+1,j log(λ0n+1,j )}
j=1

K
X
TS
+ω (βn+1,k − β̂n+1,k )2 ,
k=1

where λn+1,j = g−1 (fjeT β (n+1) ) and λ0n+1,j = g−1 (fjeT β 0(n+1) ). Dropping the
sub/super-scripts, the summand in the first term of (3.7) is
(3.8) (λ − λ0 ) − y{log(λ) − log(λ0 )},
where λ = g−1 (f T β) and λ0 = g−1 (f T β 0 ). It follows from a second-order
Taylor expansion that
(λ − λ0 ) − y{log(λ) − log(λ0 )} = w(y, f T β 0 ){f T β − y ∗ (y, f T β 0 )}2 + c(β 0 ),
where c(β 0 ) is a constant given the initial estimate β 0 , while the specific
expressions of the functions w(·, ·) and y ∗ (·, ·) depend on the link function
g.
The exact expressions of w(·, ·) and y ∗ (·, ·) for several commonly used link
functions are listed below. For the identity link,
y (f T β 0 )2 − yf T β 0
w(y, f T β 0 ) = , y ∗ (y, f T β 0 ) = f T β 0 − ;
2(f β 0 )2
T y
for the logarithmic link,
T β0
w(y, f T β 0 ) = 12 f T β 0 , y ∗ (y, f T β 0 ) = f T β 0 − (1 − ye−f );
and for the square-root link,
y (f T β 0 )3 − yf T β 0
w(y, f T β 0 ) = 1 + , y ∗ (y, f T β 0 ) = f T β 0 − .
(f β 0 )2
T (f T β 0 )2 + y
Note that, for all three links, the induced response variable y ∗ is a shifted
version of f T β 0 .
Plugging back the sub/super-scripts, barring any additive constant inde-
pendent of β (n+1) , the quadratic approximation of the minimizing criterion
C(β (n+1) ) is equivalent to
m0
X
w(yn+1,j , fjeT β 0(n+1) ){fjeT β (n+1) − y ∗ (yn+1,j , fjeT β 0(n+1) )}2
j=1

K
X
TS
+ω (βn+1,k − β̂n+1,k )2 .
k=1
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 13

In matrix form, the above quadratic approximation can be written as

∗
C(β (n+1) ) ≈ (y(n+1) − Fe β (n+1) )T W(y(n+1)
∗
− Fe β (n+1) )
T
+ ω(β (n+1) − β̂ TS TS
(n+1) ) (β (n+1) − β̂ (n+1) ),
∗
where the induced response vector y(n+1) has y ∗ (yn+1,j , fjeT β 0(n+1) ) as the
jth entry, while the diagonal weighting matrix W has w(yn+1,j , fjeT β 0(n+1) )
as the jth diagonal entry. Hence, the minimizer of quadratic approximation
of C(β (n+1) ) is

(3.9) β̂ (n+1) = (FeT WFe + ωI)−1 (FeT Wy(n+1)

∗
+ ω β̂ TS
(n+1) ).

The above derivation suggests the following iterative re-weighted least

squares (IRLS) algorithm to solve the minimization problem (3.4):
e
1. Initialize: Fit the Poisson regression model (3.3) on y(n+1) and use the
0
maximum likelihood estimate of β (n+1) as β (n+1) .
2. Update: Calculate the updated estimate β̂ (n+1) according to (3.9) for a
particular time series forecast β̂ TS
(n+1) .
3. Repeat Step 2 with β 0(n+1) replaced by newly updated β̂ (n+1) until con-
vergence.
When using the above IRLS algorithm to derive a distributional forecast
for β (n+1) , the algorithm needs to be iterated until convergence for each
bootstrapped time series forecast β̂ TS,b
(n+1) . In our implementation, we initial-
ize the algorithm by setting β 0(n+1) to be the penalized point forecast β PML
(n+1) ,
and iterate the algorithm one step toward the bootstrapped forecast β̂ TS,b
(n+1) .
This simplification helps reduce the computational effort and works well in
our examples.

4. Application to call center data. In this section we illustrate the pro-

posed methods using the call center data analyzed in Shen and Huang (2008).
For efficient staffing of the call center, we are interested in forecasting future
daily arrival rate profiles (Section 4.3), and dynamically updating existing
forecasts using new information within a day (Section 4.4). Shen and Huang
(2008) focused on forecasting future call volumes instead of rates. In addi-
tion to extending their work to deal with Poisson rate forecasting, we go
one step beyond comparing various rate forecasts—the forecasted rates are
used to calculate the required staffing level, that is, the number of agents
necessary to staff the call center in order to achieve some service level con-
straint (Sections 4.3.3 and 4.4). The effect of rate forecasting and updating
on determining the staffing level has not been considered before.
14 H. SHEN AND J. Z. HUANG

The call center data record the incoming call volumes to a northeastern
US bank call center during every quarter hour within the normal business
hour (7:00AM–midnight). The data cover 210 weekdays between January
6th and October 24th, 2003. Among them, ten abnormal days are excluded
from the analysis, including six holidays where the call volumes are very
low and four days where the data are missing [Shen and Huang (2008)].
Empirical research has recently suggested that it is appropriate to model
the arrival process of customer calls to a call center as an inhomogeneous
Poisson process [Brown et al. (2005)].
This section is organized as follows. We first provide some background on
call center agent staffing in Section 4.1 to facilitate our later calculation of
staffing levels. We fit the factor model (2.1) in Section 4.2 and develop time
series models for the factor score series. In Section 4.3 simulation studies
are first performed to investigate the rate forecasting performance of our
methods. The effect of interday rate forecasting on staffing level is then
illustrated using the real data. Various intraday updating methods are then
compared in Section 4.4 in terms of the related staffing level forecasts.

4.1. Background on call center agent staffing. Service systems such as

call centers are usually analyzed as queueing systems. Queueing theoretic
models, such as Erlang-C (or M/M/N queue), are used to balance agent
utilization and quality of service (QoS) according to some pre-specified QoS
measure [Gans, Koole and Mandelbaum (2003)]. The capacity planning and
workforce scheduling process in such systems usually takes place in three
steps. First, historical arrival data are used to generate forecasts of arrival
rates in each time period over some planning horizon. Second, queueing
models are used to determine staffing levels for each of the time periods.
For example, one simplistic yet previously commonly-used staffing strategy
is the 80-20 rule, which finds the minimum number of agents to staff a
system such that “80% of the customers will be served while their waiting
(or delay) time is less than 20 seconds.” Finally, the staffing level obtained
in the second step is used as inputs to some mathematical programming to
generate a set of agent schedules.
Below we focus on the second step of deciding the staffing level, and illus-
trate how improved rate forecasting, especially dynamic intraday updating,
can improve the staffing accuracy. Suppose the arrival rate during a time
period is λ, and an agent can handle µ calls per unit of time. Then, the
offered load is defined as R = λ/µ. Intuitively, this gives the minimum num-
ber of agents that a call center needs in order for the system to be stable.
Otherwise, more calls will arrive than the agents can process, and the queue
will eventually explode.
One well-known heavy traffic queueing-theoretic result [Halfin and Whitt
(1981)] is that for highly utilized moderate to large systems such as the call
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 15

center we analyze here, the required staffing level N is approximately equal

to
√
(4.1) N = R + θ R,
where θ > 0 is a service level parameter that is decided by the steady-state
probability of calls being delayed, α. To be exact,
 −1
θΦ(θ)
(4.2) α= 1+ ,
φ(θ)
where Φ and φ are the standard normal distribution and density functions.
For example, θ = 1 corresponds to 22% of the calls being delayed in steady
state.
This simple
√ staffing rule is known as the square-root safety staffing princi-
ple, where θ R is the safety staffing needed above the minimum requirement
R in order to achieve some desired service level. The reason that the safety
staffing is in the order of the square-root of the offered load is nontriv-
ial. Recently, Garnett, Mandelbaum and Reiman (2002) extend the above
result to queueing systems with abandonment by allowing the service level
parameter θ to be negative. In call centers, some impatient customers end up
abandoning the tele-queue before getting served. Square-root safety staffing
is awfully simple. However, in order for it to work well in practice, one relies
on accurate prediction of future offered load R, in particular, the arrival rate
λ, which is the focus of the current research.

4.2. Model building. We considered K = 1, . . . , 10 and fit the correspond-

ing factor model (2.1) to the call center data. The square-root link function
was employed. The left panel of Figure 1 gives the deviance reduction plot,
which suggests that five factors are sufficient as the reduction of the de-
viance becomes flat afterward. The right panel plots the first five factor

Fig. 1. The deviance reduction plot, suggesting five significant factors whose loadings are
plotted.
16 H. SHEN AND J. Z. HUANG

Fig. 2. Time series plot of the first score series and its lag-one scatter plot, suggesting
an AR(1) time series model with a day-of-the-week dependent slope.

loading vectors, which are normalized to achieve comparable scales. The

first loading vector roughly summarizes the average daily rate profile. The
additional loading vectors capture various contrasts among different time pe-
riods within a day. The findings are consistent with those in Shen and Huang
(2008) except theirs are about the count (instead of rate) profiles.
To perform interday forecasting and intraday updating, we first need to
develop a forecasting model for each score time series. The left panel of
Figure 2 plots the first score series β 1 . Different symbols indicate different
weekdays, revealing a strong day-of-the-week effect. The right panel shows
that the factor scores of two consecutive weekdays are linearly related, con-
ditioning on day-of-the-week. This motivated us to consider the following
varying-coefficient AR(1) model,

βi1 = a1 (di−1 ) + b1 βi−1,1 + εi1 ,

where di−1 denotes the day-of-the-week of day i − 1, and the varying in-
tercept a1 depends on di−1 . A more complicated model was also considered
where the slope is allowed to depend on day-of-the-week. However, an F -test
for the nested models returns a p-value of 0.7724, which suggests that the
larger model does not provide any significant improvement.
Further exploratory data analysis suggested similar varying intercept AR(1)
models for the other score series. Hence, we opt to use such models to fore-
cast the future scores, which are denoted as

(4.3) βik = ak (di−1 ) + bk βi−1,k + εik , i = 2, . . . , n, k = 1, . . . , 5.

 FORECASTING INHOMOGENEOUS POISSON PROCESSES 17

4.3. One-day-ahead rate forecasting. We first describe in Section 4.3.1

two useful parametric models motivated by the call center data. The models
are used in Section 4.3.2 to generate data in simulation studies to test our
interday forecasting methods. By considering the effect of forecasting on
staffing level, our forecasting methods will be further tested in Section 4.3.3
using the real data. For our methods, we consider dimension reduction by the
factor model (2.1) with K = 1, . . . , 5 underlying factors respectively, followed
by time series modeling with (4.3) on each series of factor scores. We shall
denote our methods as TS1 to TS5 depending on the number of factors used.

4.3.1. Two parametric models. We consider two parametric models, where

the data yij ∼ Poisson(λij ), i = 1, . . . , n, j = 1, . . . , m, and the underlying
Poisson rate is assumed to have a two-way structure characterizing re-
spectively the day-by-day variation and time-of-day variation. These two
kinds of variations are typical for call center data [Brown et al. (2005),
Weinberg, Brown and Stroud (2007)]. By taking into count the Poisson na-
ture of the observations, the two models extend the corresponding models
used in Shen and Huang (2008), which instead assume the square-root trans-
formed counts as Gaussian observations.
A multiplicative model. The multiplicative (MUL) model assumes that the
Poisson rates satisfy
p


λij = αi γdi j , di = 1, 2, 3, 4, 5,
 α − a = b(α ηi ∼ N (0, φ2 ),
i di i−1 − adi−1 ) + ηi ,
(4.4) X

 γdi j = 1.

j

Under this model, the square-root of the rate has a two-way multiplicative
structure: the day-to-day variation (αi ) is auto-regressive of order one with
mean depending on day-of-the-week (di ); the time-of-day variation (γdi j )
also depends on day-of-the-week. This model can be viewed as
a non-Bayesian version of the Bayesian model proposed by
Weinberg, Brown and Stroud (2007), except that they also require the in-
traday variation γdi j to be smooth.
An additive model. The additive (ADD) model assumes that the Poisson
rates satisfy
q

 λij = µ + αi + βj + γdi j , di = 1, 2, 3, 4, 5,



(4.5) αi − adi = b(αi−1 − adi−1 ) + ηi , ηi ∼ N (0, φ2 ),
 X X X X X

 α i = βj = γd j = γ d j = γdi j = 0.

 i i
i j i j ij

Under this model, the square-root-transformed rate has a two-way additive

structure: the day-to-day variation follows an AR(1) process adjusting for
18 H. SHEN AND J. Z. HUANG

day-of-the-week (adi ) effect; it also has an interaction with the time-of-day

variation (γdi j ). q
As mentioned earlier, when yij ∼ Poisson(λij ), xij = yij + 1/4 is ap-
p
proximately N ( λij , 1/4). Hence, the estimation and forecasting methods
described in Sections A.3 and A.4 of Shen and Huang (2008) are applicable
for the above two models.
Both Weinberg, Brown and Stroud (2007) and Shen and Huang (2008)
considered an industry standard as their benchmark, which uses historical
averages of the same day-of-the-week as forecasts. They both showed that
the approach works poorly; hence, we choose not to consider it in our studies.

4.3.2. Simulation studies of rate forecasting performance. Because the

Poisson rate profiles are unobservable in practice, in order to evaluate the
rate forecasting performance of our methods, we carry out two simulation
studies, using the multiplicative model (4.4) and the additive model (4.5)
respectively to generate data. The model parameters are decided on by fit-
ting the models to the real data described at the beginning of Section 4 and
rounding the parameter estimates properly. We generate 100 data sets from
each model for use in our simulation studies. The simulation results (de-
tailed below) show that, without a specific parametric model assumption,
our approach can always perform almost as well as the method using the
true data generating model. As a comparison, a method based on a specific
model assumption performs worse if the model assumption is wrong.
Seven forecasting methods are considered in our simulation studies, in-
cluding forecasting assuming the multiplicative model (4.4), forecasting as-
suming the additive model (4.5), and our methods TS1 to TS5. For all
methods, the true square-root link function is used. For each data set, a
rolling out-of-sample forecast exercise is performed with the last 50 days as
the forecasting set; and for each day in the forecasting set, the preceding
150 days are used to derive the forecast. The parameters involved in the
methods are re-estimated for each day in the forecasting set.
To compare the performance of forecasting the underlying rate, two per-
formance measures are calculated, the root mean squared error (RMSE) and
mean relative error (MRE). For day i in each data set, we define
v
u X
u1 m m
100 X |λ̂ij − λij |
RMSEi = t (λ̂ij − λij )2 and MREi = ,
m j=1 m j=1 λij

where λ̂ij is the forecast for λij .

The Mean RMSE and Mean MRE (%) of the forecasted rates are then
calculated for each simulated data set. For each forecasting method, the
two panels of Figure 3 plot the empirical cumulative distribution functions
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 19

Fig. 3. Comparison of empirical CDF of Mean RMSE for rate forecasting. The true
model is highlighted in the legend using an asterisk. TS4 performs comparably to the true
model, while the wrong model performs worse.

(CDF) of the Mean RMSE calculated on the 100 data sets simulated using
the multiplicative and additive models respectively. The true model is indi-
cated in the legend using an asterisk. When the data are generated using the
multiplicative model, the additive model is a misspecified model and vice
versa. In both cases, our TS methods do not use the model assumptions on
the form of the underlying Poisson rate. The improvement of TS5 over TS4
is minimal; hence, the CDF for TS5 is omitted in both panels.
Several observations can be made from the plots. For the TS methods,
when increasing the number of factors K, the performance measure gets
stochastically smaller; TS4 is competitively close to the true model; when
K increases beyond five, the forecasting improvement remains minimal. This
suggests that four underlying factors are sufficient to approximate the true
models. In addition, the wrong parametric model results in stochastically
inferior forecasts than TS4. The uniformly good performance of our method
shows its robustness against model assumptions, which is important in prac-
tice where the information of the true underlying model is hardly ever avail-
able. The results for Mean MRE (%) are similar and are not shown.

4.3.3. The effect of rate forecasting on staffing level. Below we apply the
seven forecasting methods considered previously in Section 4.3.2 to the call
center data, and illustrate how the rate forecasting affects staffing level. This
important fundamental question has not been investigated previously in the
literature. To calculate the staffing level, we apply the square-root safety
staffing rule discussed in Section 4.1.
We assume an average service time of 5 minutes per call, which corre-
sponds to an agent service rate of 12 calls per hour, or 3 calls per 15-minute
interval. This figure is chosen subjectively for illustration purposes based
on the empirical analysis in Brown et al. (2005). Suppose the arrival rate is
20 H. SHEN AND J. Z. HUANG

Table 1
Summary statistics (mean, median, lower quartile Q1, upper quartile Q3) of RMSE and
MRE of the forecasted staffing level in a rolling forecast exercise. The forecasting set
contains 50 days. MUL, TS4 and TS5 perform comparable

RMSE MRE (%)

Q1 Median Mean Q3 Q1 Median Mean Q3

ADD 14.28 18.03 20.31 21.80 4.63 5.40 6.29 6.68

MUL 14.27 16.65 19.69 20.38 4.46 5.06 5.91 6.30
TS1 14.91 18.55 20.63 21.79 4.96 6.28 7.07 8.31
TS2 13.54 17.51 19.88 20.89 4.73 5.41 6.32 6.88
TS3 14.74 17.80 19.77 19.90 4.63 5.27 6.12 6.62
TS4 14.32 16.55 19.50 20.27 4.45 5.06 6.00 6.42
TS5 13.98 16.55 19.46 20.57 4.44 5.06 5.98 6.35

forecasted to be λj for the jth interval. Then the offered load is Rj = λj /3.
To figure out the additional safety staffing, we assume that the call center
expects a 22% delay probability in steady state, which then determines the
safety staffing factor θ to be 1 according to (4.2). Hence, it follows that the
required staffing level for interval j is
q
λj /3 + λj /3.

We perform, on our data set, the same rolling one-day-ahead forecasting

exercise as the one described in the simulation studies of Section 4.3.2. As
the benchmark, we use the “oracle” staffing level decided by assuming the
actual call volumes as the rate forecasts. Note that we wouldn’t observe
the actual volumes when actually planning the staffing in practice. We then
treat the “oracle” staffing level as the truth when calculating performance
measures of the forecasted staffing level decided by the various rate forecasts.
Table 1 compares summary statistics of the RMSE and MRE of the
staffing level forecasts from the seven methods. For the TS methods, the
results from using the square-root link function are presented. In general,
the forecasting accuracy improves as one increases K, the number of fac-
tors. TS4 and TS5 give comparable results. On the ground of parsimony,
TS4 may be preferred to TS5. The ADD model performs similar to TS2,
while MUL similar to TS4/TS5.
In practice, out-of-sample forecasting performance can be used to pick the
link function. Our choice of the square-root link in Table 1 is primarily for
comparison purpose, because both ADD and MUL models use the square-
root link. In addition, we also tried the identity link and the logarithmic
link for our methods in this example, which generated inferior results than
the square-root link.
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 21

4.4. Dynamic intraday updating of rate and staffing level. The call cen-
ter data are used to illustrate the benefit of dynamic updating. We focus
on the 10:00AM updating and the 12:00PM updating, which means that
intraday dynamic updating is performed at 10:00AM and 12:00PM, respec-
tively. The benchmark is the TS4 method which performs no updating. We
shall use the square-root link and K = 4 based on the forecasting results in
Section 4.3.3. Correspondingly, we term our penalized maximum likelihood
intraday updating approach as PML4.
For comparison purpose, we also consider two alternative intraday up-
dating approaches that combine the MUL/ADD forecasts with historical
proportions (HP). Suppose the MUL point forecast for the rate profile of
day n + 1 is λ̂MUL
(n+1) , which also gives the point forecast for the count profile
y(n+1) . For an updating point m0 :
• calculate the ratio R between the total number of calls arrived and the
cumulative forecasted rate up to the time period m0 ,
Pm0
j=1 yn+1,j
R = Pm0 MUL
;
j=1 λ̂n+1,j

• update the MUL forecasts for the remainder of the day as

λ̂HPM MUL
n+1,j = Rλ̂n+1,j , j = m0 + 1, . . . , m.
We refer to the above updating method as HPM. One can replace the MUL
forecast with the ADD forecast in the above algorithm, which leads to the
HPA updating. The HPM and HPA updating methods given here are ex-
tensions of similar methods for updating counts described in Section 4.2.2
of Shen and Huang (2008).
Below we apply the aforementioned updating methods to the call center
data, and illustrate the effect of dynamic intraday updating on staffing call
centers. The same rolling forecasting exercise as in Section 4.3.3 is performed,
and we again use the square-root staffing rule to determine staffing level.
For PML4, the value of the penalty parameter ω needs to be decided at
each updating point. To this end, we consider a set of candidate values, and
choose the one that minimizes some forecasting performance measure based
on the call volumes (rather than rates), such as RMSE. Specifically, we use
the beginning 150 days in our data set as the training set, and perform
an out-of-sample rolling forecast exercise on this training set. Note that this
training set does not overlap with the forecasting set we initially hold out for
out-of-sample forecast evaluation. For the purpose of selecting ω, the last one
third (i.e., 50 days) of the training set is used as a hold-out sample. For each
day in this sample, its preceding 100 days are used to fit the Poisson factor
model and to generate the PML4 updating for each given ω. Compute some
22 H. SHEN AND J. Z. HUANG

Table 2
Summary statistics (mean, median, lower quartile Q1 and upper quartile Q3) of RMSE
of the updated staffing level for the 10:00AM and 12:00PM updatings. PML4 outperforms
the other methods

10:00AM updating 12:00PM updating

Q1 Median Mean Q3 Q1 Median Mean Q3

TS4 12.42 14.29 18.03 18.82 12.42 14.29 18.03 18.82

HPA 12.80 17.04 19.33 21.46 12.20 13.23 16.67 16.33
HPM 12.71 15.49 19.02 21.03 11.98 13.04 16.17 16.02
PML4 12.62 14.81 17.19 18.10 11.46 12.87 15.41 14.54

performance measure (e.g., RMSE) for every day in the hold-out sample and
take the average. The ω that minimizes this average performance measure
will be used for all days in the forecasting set. In this study, we consider ω
from {0, 10, . . . , 109 }, and ω = 103 is chosen for both updatings.
Table 2 presents summary statistics of the RMSE of the forecasted staffing
levels from TS4, HPA, HPM and PML4. The averages are calculated over
the 50 days in the forecasting set. For a fair comparison, only data after
12:00PM are used when calculating the RMSE. The superior performance
of PML4 over the other methods is quite clear, which improves over TS4
by 14.5% on average. We also observe that, for every intraday updating
method, updating later always improves the forecasting accuracy. In addi-
tion, intraday updating reduces the prediction interval width as illustrated
in the right panel of Figure 4.
Figure 4 provides a graphical illustration of the forecasted staffing lev-
els for September 2nd, 2003. The left panel shows the required number of
agents for every 15-minute interval between 10:00AM and 17:00PM, when

Fig. 4. Comparison between the required number of agents per 15-minute interval for
September 2nd based on the forecasted rates from various methods. The PML4 10:00AM
updating gives the most accurate staffing.
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 23

the majority of the calls get connected to the center. We observe that the
10:00AM updating results in a dramatic upward shift in the related staffing
level, which is very close to the oracle staffing level. On the other hand, the
TS4 forecast leads to uniformly under-staffing throughout the day. What
happens is that September 2nd corresponds to the day after Labor Day.
Consequently, the call center experienced an unusably high call volume that
day, which was not accounted for by the TS4 method. The HP updating
methods can somehow adjust for the upward shift; however, their perfor-
mance is worse than the PML4 method.
In the right panel of Figure 4, we superimpose the 95% prediction inter-
vals for the staffing level resulting from the corresponding intervals for the
forecasted arrival rates. Only PML4 and TS4 are compared. The number of
bootstrapped forecasts is chosen to be B = 1000. We observe that the in-
tervals from the intraday updating (PML4) almost always cover the “true”
staffing level, while the intervals without the updating miss for more than
half of the time periods. In addition, intraday updating results in uniformly
narrower prediction intervals, which suggests that the prediction is more
precise. The average interval width for TS4 and PML4 is 84.86 and 49.65,
respectively; hence, a 41.5% average reduction results from the intraday up-
dating. The managerial benefit is that the call center manager can work
with a much narrower range of staffing level in the presence of arrival rate
uncertainty.

5. Conclusion and discussion. In this paper we develop methods for fore-

casting and dynamic updating of the latent and uncertain rate profiles of a
time series of inhomogeneous Poisson processes. Our new approach extends
the model-driven approach of Weinberg, Brown and Stroud (2007) and the
data-driven approach of Shen and Huang (2008). The latter does not explic-
itly consider forecasting the latent Poisson rates. Instead, the authors con-
sider forecasting of call volumes that can be regarded as a noisy version of
the rates. We propose a Poisson factor model that combines the distribution
assumption with data-driven dimension reduction. An iterative algorithm
is proposed for parameter estimation. Besides showing the competitiveness
and robustness of our forecasting methods, we also illustrate how intraday
updating can improve the accuracy of call center staffing, which is the first
in the relevant literature and provides some important managerial insights.
In terms of model fitting, our approach involves two stages: first fit a Pois-
son factor model for dimension reduction, then fit time series models on the
factor score series. If the time series models could be specified together with
the factor model prior to data analysis, a full likelihood approach could be
statistically more efficient. However, the time series models are usually not
easy to specify beforehand, so our two stage procedure is natural for model
building. Moreover, numerical calculation for the full likelihood approach
24 H. SHEN AND J. Z. HUANG

is much more complicated than our procedure, although it is conceptually

simple. In our call center application, fast computation is of great concern,
especially when performing dynamic within-day updating. Nevertheless, it
is still interesting to explore the full likelihood approach. Whether statis-
tical efficiency of using full likelihood translates into benefits in accurate
forecasting deserves further research.

Acknowledgments. The authors want to extend grateful thanks to the

Editor, the Associate Editor and two reviewers, whose comments have greatly
improved the scope and presentation of the paper.

REFERENCES
Aksin, Z., Armony, M. and Mehrotra, V. (2007). The modern call-center: A multi-
disciplinary perspective on operations management research. Production and Operations
Management 16 665–688.
Box, G. E. P., Jenkins, G. M. and Reinsel, G. C. (1994). Time Series Analysis: Fore-
casting and Control, 3rd ed. Prentice Hall, Englewood Cliffs, New Jersey. MR1312604
Brown, L. D., Cai, T., Zhang, R., Zhao, L. and Zhou, H. (2007). The root-unroot
algorithm for density estimation as implemented via wavelet block thresholding. Tech-
nical report.
Brown, L. D., Gans, N., Mandelbaum, A., Sakov, A., Shen, H., Zeltyn, S. and
Zhao, L. (2005). Statistical analysis of a telephone call center: A queueing-science
perspective. J. Amer. Statist. Assoc. 100 36–50. MR2166068
Cox, D. R. (1955). Some statistical methods connected with series of events. J. Roy.
Statist. Soc. Ser. B 17 129–164. MR0092301
Diggle, P., Heagerty, P., Liang, K. Y. and Zeger, S. (2002). Analysis of Longitudinal
Data, 2nd ed. Oxford Univ. Press, New York. MR2049007
Dobson, A. J. (2001). An Introduction to Generalized Linear Models, 2nd ed. Chapman
and Hall/CRC, London. MR1070981
Gabriel, K. R. and Zamir, S. (1979). Lower rank approximation of matrices by least
squares with any choice of weights. Technometrics 21 489–498.
Gans, N., Koole, G. M. and Mandelbaum, A. (2003). Telephone call centers: Tutorial,
review, and research prospects. Manufacturing & Service Operations Management 5
79–141.
Garnett, O., Mandelbaum, A. and Reiman, M. (2002). Designing a call center with
impatient customers. Manufacturing & Service Operations Management 4 208–227.
Halfin, S. and Whitt, W. (1981). Heavy-traffic limits for queues with many exponential
servers. Operations Research 29 567–588. MR0629195
Harvey, A. C. (1990). Forecasting, Structural Time Series Models and the Kalman Filter.
Cambridge Univ. Press, New York. MR1085719
Mandelbaum, A. (2006). Call centers. Research bibliography with abstracts. Technical
report, Technion, Israel.
McCullagh, P. and Nelder, J. A. (1989). Generalized Linear Models, 2nd ed. Chapman
and Hall, London. MR0727836
Reinsel, G. C. (1997). Elements of Multivariate Time Series Analysis, 2nd ed. Springer,
New York. MR1451875
Shen, H. and Huang, J. Z. (2008). Interday forecasting and intraday updating of call
center arrivals. Manufacturing & Service Operations Management. To appear.
 FORECASTING INHOMOGENEOUS POISSON PROCESSES 25

Shen, H., Huang, J. Z. and Lee, C. (2007). Forecasting and dynamic updating of un-
certain arrival rates to a call center. Proceedings of the IEEE/INFORMS International
Conference on Service Operations and Logistics, and Informatics 50–55.
Stock, J. H. and Watson, M. W. (2005). Implications of dynamic factor models for
VAR analysis. NBER Working Paper No. 11467.
Weinberg, J., Brown, L. D. and Stroud, J. R. (2007). Bayesian forecasting of an
inhomogeneous Poisson process with applications to call center data. J. Amer. Statist.
Assoc. 102 1185–1199.

Department of Statistics and Operations Research Department of Statistics

University of North Carolina at Chapel Hill Texas A&M University
Chapel Hill, North Carolina 27599 College Station, Texas 77843
USA USA
E-mail: [email protected] E-mail: [email protected]