Towards “zero abandonments” in call center performance

The call center industry is a big business in today's global economy. Staffing costs account for over half of a call center's total operations costs. Some large call centers, in practice, operate at very close to maximum capacity, believing that such an operations policy is efficient. However, by operating at levels close to 100% utilization, a call center is “living dangerously”. If, for example, call volumes even slightly exceed forecasts, customer calls will queue. As queue lengths and durations increase, customers will tend to abandon their calls. We provide some “rule-of-thumb” formulas that evaluate the cost of abandonments. These formulas may be used to justify an investment in additional agents required to improve the quality of service and reduce abandonments. Standard Erlang-C queueing formulas imply that abandonments can be significantly reduced with a small investment in additional agents. Thus, by improving customer service and hiring additional staff, a call center can improve profitability. We illustrate our analysis with realistic data, based on our work with large-scale customer service centers.     

Quasi-reversible multiclass queues with order independent departure rates

This paper introduces a new class of queues which are quasi-reversible and therefore preserve product form distribution when connected in multinode networks. The essential feature leading to the quasi-reversibility of these queues is the fact that the total departure rate in any queue state is independent of the order of the customers in the queue. We call such queues order independent (OI) queues. The OI class includes a significant part of Kelly's class of symmetric queues, although it does not cover the whole class. A distinguishing feature of the OI class is that, among others, it includes the MSCCC and MSHCC queues but not the LCFS queue. This demonstrates a certain generality of the class of OI queues and shows that the quasi-reversibility of the OI queues derives from causes other than symmetry principles. Finally, we examine OI queues where arrivals to the queue are lost when the number of customers in the queue equals an upper bound. We obtain the stationary distribution for the OI loss queue by normalizing the stationary probabilities of the corresponding OI queue without losses. A teletraffic application for the OI loss queue is presented.     

Retrial queues with balanced call blending: analysis of single-server and multiserver model

In call centers, call blending consists in the mixing of incoming and outgoing call activity, according to some call blending balance. Recently, Artalejo and Phung-Duc have developed an apt model for such a setting, with a two way communication retrial queue. However, by assuming a classical (proportional) retrial rate for the incoming calls, the short-term blending balance is heavily impacted by the number of incoming calls, which may be undesired, especially when the balance between incoming and outgoing calls is vital to the service offered. In this contribution, we consider an alternative to classical call blending, through a retrial queue with constant retrial rate for incoming calls. For the single-server case (one operator), a generating functions approach enables to derive explicit formulas for the joint stationary distribution of the number of incoming calls and the system state, and also for the factorial moments. This is complemented with a stability analysis, expressions for performance measures, and also recursive formulas, allowing reliable numerical calculation. A correlation study enables to study the system’s short-term blending balance, allowing to compare it to that of the system with classical retrial rate. For the multiserver case (multiple operators), we provide a quasi-birth-and-death process formulation, enabling to derive a sufficient and necessary condition for stability in this case (in a simple form), a numerical recipe to obtain the stationary distribution, and a cost model.     

On a Two-Queue Priority System with Impatience and its Application to a Call Center*

We consider an s-server priority system with a protected and an unprotected queue. The arrival rates at the queues and the service rate may depend on the number n of customers being in service or in the protected queue, but the service rate is assumed to be constant for n > s. As soon as any server is idle, a customer from the protected queue will be served according to the FCFS discipline. However, the customers in the protected queue are impatient. If the offered waiting time exceeds a random maximal waiting time I, then the customer leaves the protected queue after time I. If I is less than a given deterministic time, then he leaves the system, else he will be transferred by the system to the unprotected queue. The service of a customer from the unprotected queue will be started if the protected queue is empty and more than a given number of servers become idle. The model is a generalization of the many-server queue with impatient customers. The global balance conditions seem to have no explicit solution. However, the balance conditions for the density of the stationary state process for the subsystem of customers being in service or in the protected queue can be solved. This yields the stability conditions and the probabilities that precisely n customers are in service or in the protected queue. For obtaining performance measures for the unprotected queue, a system approximation based on fitting impatience intensities is constructed. The results are applied to the performance analysis of a call center with an integrated voice-mail-server.     

A tollbooth tandem queue with heterogeneous servers

We study a tandem queueing system with K servers and no waiting space in between. A customer needs service from one server but can leave the system only if all down-stream servers are unoccupied. Such a system is often observed in toll collection during rush hours in transportation networks, and we call it a tollbooth tandem queue. We apply matrix-analytic methods to study this queueing system, and obtain explicit results for various performance measures. Using these results, we can efficiently compute the mean and variance of the queue lengths, waiting time, sojourn time, and departure delays. Numerical examples are presented to gain insights into the performance and design of the tollbooth tandem queue. In particular, it reveals that the intuitive result of arranging servers in decreasing order of service speed (i.e., arrange faster servers at downstream stations) is not always optimal for minimizing the mean queue length or mean waiting time.     

On the stability of retrial queues

We consider the following Type of problems. Calls arrive at a queue of capacity K (which is called the primary queue), and attempt to get served by a single server. If upon arrival, the queue is full and the server is busy, the new arriving call moves into an infinite capacity orbit, from which it makes new attempts to reach the primary queue, until it finds it non-full (or it finds the server idle). If the queue is not full upon arrival, then the call (customer) waits in line, and will be served according to the FIFO order. If λ is the arrival rate (average number per time unit) of calls and μ is one over the expected service time in the facility, it is well known that μ > λ is not always sufficient for stability. The aim of this paper is to provide general conditions under which it is a sufficient condition. In particular, (i) we derive conditions for Harris ergodicity and obtain bounds for the rate of convergence to the steady state and large deviations results, in the case that the inter-arrival times, retrial times and service times are independent i.i.d. sequences and the retrial times are exponentially distributed; (ii) we establish conditions for strong coupling convergence to a stationary regime when either service times are general stationary ergodic (no independence assumption), and inter-arrival and retrial times are i.i.d. exponentially distributed; or when inter-arrival times are general stationary ergodic, and service and retrial times are i.i.d. exponentially distributed; (iii) we obtain conditions for the existence of uniform exponential bounds of the queue length process under some rather broad conditions on the retrial process. We finally present conditions for boundedness in distribution for the case of nonpatient (or non persistent) customers. This revised version was published online in June 2006 with corrections to the Cover Date.     

Robust connection acceptance control for ATM networks with incomplete source information

Consider an ATM resource where the statistics of each bursty traffic source are not precisely known. How should the resource decide when to accept fresh connections? We address this problem for the case where each source behaves as a simple “On-Off” model and only the peak bandwidth is known. A connection acceptance control algorithm is proposed that relies on the choice of a boundary. The performance of this algorithm is analysed using an approximation based on separation of timescales. The slowly changing call process behaves like the queue process at an infinite server queue, where the arrival process to the queue is dictated by the fast moving burst process. A technique for finding a suitable boundary is also given, based on the Bayesian principle. The algorithm, together with the proposed boundary, could be used in a network to control connection acceptance at each resource along a virtual path.     

Bounds for the tail distribution in a queue with a superposition of general periodic Markov sources: theory and application

An efficient yet accurate estimation of the tail distribution of the queue length has been considered as one of the most important issues in call admission and congestion controls in ATM networks. The arrival process in ATM networks is essentially a superposition of sources which are typically bursty and periodic either due to their origin or their periodic slot occupation after traffic shaping. In this paper, we consider a discrete-time queue where the arrival process is a superposition of general periodic Markov sources. The general periodic Markov source is rather general since it is assumed only to be irreducible, stationary and periodic. Note also that the source model can represent multiple time-scale correlations in arrivals. For this queue, we obtain upper and lower bounds for the asymptotic tail distribution of the queue length by bounding the asymptotic decay constant. The formulas can be applied to a queue having a huge number of states describing the arrival process. To show this, we consider an MPEG-like source which is a special case of general periodic Markov sources. The MPEG-like source has three time-scale correlations: peak rate, frame length and a group of pictures. We then apply our bound formulas to a queue with a superposition of MPEG-like sources, and provide some numerical examples to show the numerical feasibility of our bounds. Note that the number of states in a Markov chain describing the superposed arrival process is more than 1.4 × 10⁸⁸. Even for such a queue, the numerical examples show that the order of the magnitude of the tail distribution can be readily obtained.     

Online scheduling policies for multiclass call centers with impatient customers

We consider a call center with two classes of impatient customers: premium and regular classes. Modeling our call center as a multiclass GI/GI/s+M$\mathit{GI}/\mathit{GI}/s+M$ queue, we focus on developing scheduling policies that satisfy a target ratio constraint on the abandonment probabilities of premium customers to regular ones. The problem is inspired by a real call center application in which we want to reach some predefined preference between customer classes for any workload condition. The motivation for this constraint comes from the difficulty of predicting in a quite satisfying way the workload. In such a case, the traditional routing problem formulation with differentiated service levels for different customer classes would be useless. For this new problem formulation, we propose two families of online scheduling policies: queue joining and call selection policies. The principle of our policies is that we adjust their routing rules by dynamically changing their parameters. We then evaluate the performance of these policies through a numerical study. The policies are characterized by simplicity and ease of implementation.     

On the interaction between retrials and sizing of call centers

This paper models a call center as a Markovian queue with multiple servers, where customer impatience, and retrials are modeled explicitly. The model is analyzed as a continuous time Markov chain. The retrial phenomenon is explored numerically using a real example, to demonstrate the magnitude it can take and to understand its sensitivity to various system parameters. The model is then used to assess the impact of disregarding existing retrials in the staffing of a call center. It is shown that ignoring retrials can lead to under-staffing or over-staffing with respect to the optimal, depending on the forecasting assumptions being made.     

Managing uncertainty in call centres using Poisson mixtures

We model a call centre as a queueing model with Poisson arrivals having an unknown varying arrival rate. We show how to compute prediction intervals for the arrival rate, and use the Erlang formula for the waiting time to compute the consequences for the occupancy level of the call centre. We compare it to the current practice of using a point estimate of the arrival rate (assumed constant) as forecast. Copyright © 2001 John Wiley & Sons, Ltd.     

The impact of special days in call arrivals forecasting: A neural network approach to modelling special days

A key challenge for call centres remains the forecasting of high frequency call arrivals collected in hourly or shorter time buckets. In addition to the complex intraday, intraweek and intrayear seasonal cycles, call arrival data typically contain a large number of anomalous days, driven by the occurrence of holidays, special events, promotional activities and system failures. This study evaluates the use of a variety of univariate time series forecasting methods for forecasting intraday call arrivals in the presence of such outliers. Apart from established, statistical methods, we consider artificial neural networks (ANNs). Based on the modelling flexibility of the latter, we introduce and evaluate different methods to encode the outlying periods. Using intraday arrival series from a call centre operated by one of Europe’s leading entertainment companies, we provide new insights on the impact of outliers on the performance of established forecasting methods. Results show that ANNs forecast call centre data accurately, and are capable of modelling complex outliers using relatively simple outlier modelling approaches. We argue that the relative complexity of ANNs over standard statistical models is offset by the simplicity of coding multiple and unknown effects during outlying periods.     

Forecasting call frequency at a financial services call centre

A forecasting model is developed for the number of daily applications for loans at a financial services telephone call centre. The purpose of the forecasts and the associated prediction intervals is to provide effective staffing policies within the call centre. The model building process is constrained by the availability of only 2 years and 7 months of data. The distinctive feature of the data is that demand is driven in the main by advertising. The analysis given focuses on applications stimulated by press advertising. Unlike previous analyses of broadly similar data, where ARIMA models were used, a model with a dynamic level, multiplicative calendar effects and a multiplicative advertising response is developed and shown to be effective.     

Analysis of call centre arrival data using singular value decomposition

We consider the general problem of analysing and modelling call centre arrival data. A method is described for analysing such data using singular value decomposition (SVD). We illustrate that the outcome from the SVD can be used for data visualization, detection of anomalies (outliers), and extraction of significant features from noisy data. The SVD can also be employed as a data reduction tool. Its application usually results in a parsimonious representation of the original data without losing much information. We describe how one can use the reduced data for some further, more formal statistical analysis. For example, a short‐term forecasting model for call volumes is developed, which is multiplicative with a time series component that depends on day of the week. We report empirical results from applying the proposed method to some real data collected at a call centre of a large‐scale U.S. financial organization. Some issues about forecasting call volumes are also discussed. Copyright © 2005 John Wiley & Sons, Ltd.     

Analytical Distribution of Waiting Time in the M/{iD}/1 Queue

We give an analytical formula for the steady-state distribution of queue-wait in the M/G/1 queue, where the service time for each customer is a positive integer multiple of a constant D > 0. We call this an M/{iD}/1 queue. We give numerical algorithms to calculate the distribution. In addition, in the case that the service distribution is sparse, we give revised algorithms that can compute the distribution more quickly.AMS subject classification: 60K25, 90B22     

Large deviations and the generalized processor sharing scheduling for a two-queue system

We establish the optimal asymptotic decay rate of per-session queue length tail distributions for a two-queue system where a single constant rate server serves the two queues using the Generalized Processor Sharing (GPS) scheduling discipline. The result is obtained using the sample-path large deviation principle and has implications in call admission control for high-speed communication networks. This revised version was published online in June 2006 with corrections to the Cover Date.     

A Mathematical Programming Approach to Optimal Markovian Switching of Poisson Arrival Streams to Queueing Systems

Motivated by certain situations in manufacturing systems and communication networks, we look into the problem of maximizing the profit in a queueing system with linear reward and cost structure and having a choice of selecting the streams of Poisson arrivals according to an independent Markov chain. We view the system as a MMPP/GI/1 queue and seek to maximize the profits by optimally choosing the stationary probabilities of the modulating Markov chain. We consider two formulations of the optimization problem. The first one (which we call the PUT problem) seeks to maximize the profit per unit time whereas the second one considers the maximization of the profit per accepted customer (the PAC problem). In each of these formulations, we explore three separate problems. In the first one, the constraints come from bounding the utilization of an infinite capacity server; in the second one the constraints arise from bounding the mean queue length of the same queue; and in the third one the finite capacity of the buffer reflect as a set of constraints. In the problems bounding the utilization factor of the queue, the solutions are given by essentially linear programs, while the problems with mean queue length constraints are linear programs if the service is exponentially distributed. The problems modeling the finite capacity queue are non-convex programs for which global maxima can be found. There is a rich relationship between the solutions of the PUT and PAC problems. In particular, the PUT solutions always make the server work at a utilization factor that is no less than that of the PAC solutions.     

Adaptive server staffing in the presence of time-varying arrivals: a feed-forward control approach

We study the short-term staffing problem of systems that experience random, non-stationary demand. The typical method to accommodate changes in arrival rate is to use historical data to identify peak periods and associated forecasting for upcoming time windows. In this paper, we develop a method that instead detects change as it happens. Motivated by an automatic call distributor system in a call centre with time-varying arrivals, we propose a change detection algorithm based upon non-homogeneous Poisson processes. The proposed method is general and may be thought of as a feed-forward strategy, in which we detect a change in the arrival process, estimate the new magnitude of the arrival rate, and assign an appropriate number of servers to the tasks. The generalized likelihood ratio statistic is used and a recommendation for its decision limit is developed. Simulation results are used to evaluate the performance of the detector in the context of a telephone call centre.     

A combined splitting—cross entropy method for rare-event probability estimation of queueing networks

We present a fast algorithm for the efficient estimation of rare-event (buffer overflow) probabilities in queueing networks. Our algorithm presents a combined version of two well known methods: the splitting and the cross-entropy (CE) method. We call the new method SPLITCE. In this method, the optimal change of measure (importance sampling) is determined adaptively by using the CE method. Simulation results for a single queue and queueing networks of the ATM-type are presented. Our numerical results demonstrate higher efficiency of the proposed method as compared to the original splitting and CE methods. In particular, for a single server queue example we demonstrate numerically that both the splitting and the SPLITCE methods can handle our buffer overflow example problems with both light and heavy tails efficiently. Further research must show the full potential of the proposed method.     

Designing a call center with an IVR (Interactive Voice Response)

A call center is a service operation that caters to customer needs via the telephone. Call centers typically consist of agents that serve customers, telephone lines, an Interactive Voice Response (IVR) unit, and a switch that routes calls to agents. In this paper we study a Markovian model for a call center with an IVR. We calculate operational performance measures, such as the probability for a busy signal and the average wait time for an agent. Exact calculations of these measures are cumbersome and they lack insight. We thus approximate the measures in an asymptotic regime known as QED (Quality and Efficiency Driven) or the Halfin–Whitt regime, which accommodates moderate to large call centers. The approximations are both insightful and easy to apply (for up to 1000’s of agents). They yield, as special cases, known and novel approximations for the M/M/N/N (Erlang-B), M/M/S (Erlang-C) and M/M/S/N queue.