Bivariate, First-Order Vector -Autoregression Bayesian Model
-
Upload
antonia-caline-petre -
Category
Documents
-
view
226 -
download
0
Transcript of Bivariate, First-Order Vector -Autoregression Bayesian Model
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
1/12
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
2/12
levels. In contrast, when marketing has a persistent effect, the
underlying demand process diverges over time by indenitely
accumulating demand shocks and necessitates a higher level of
safety stock to mitigate demand uncertainty. In this case, a non-
stationary demand model needs to be used instead.
To illustrate, a sales promotion may persuade a thousand
consumers to switch to a product at the promotional price during
each sale period (Dekimpe and Hanssens, 1995a). If these con-
sumers return to their previous purchasing habits once thepromotion has ended, the resulting uctuations are temporary in
nature and have no impact upon the underlying consumer
demand trend. The inventory manager must use a stationary
demand model and be careful not to overinterpret short-run
demand uctuations as an indication of future demand patterns
and set correspondingly high inventory quantities. In contrast, if
two hundred of the promotion-captured customers not only make
an initial purchase but also continue to purchase the product in
future, the demand shocks have a persistent effect and we would
see sales deviating permanently from pre-promotional levels. In
this scenario, the inventory manager is required to set a higher
safety stock level by using a non-stationary demand model in
order to buffer against persistent demand shocks.
In this paper, we assume that the inventory manager operates
with a base stock (order-up-to) policy based on a critical fractile
(e.g., Graves, 1999) and no backorders are assumed. Under this
policy, one orders a variable quantity every xed period of time so
that an inventory position is maintained at a predened base stock
level. We further assume that the underlying demand process is
autocorrelated (Urban, 2005; Charnes et al., 1995; Chen and Blue,
2010). A frequent practice is then to make inventory decisions on
the assumption that the true demand model in response to new
marketing efforts is known with certainty. An overcondent
inventory manager, believing her knowledge of the nature of
demand uctuations to be accurate, chooses either a stationary
or non-stationary demand model to estimate inventory base stock
levels. However, in the absence of a long sampling span, it is
difcult in practice to capture the long-term trend and distinguish
the correct nature of demand shocks. Often, small samples are
falsely thought to represent the properties of the statistical process
that generated them. This is known as the law of small numbers
(Camerer, 1989; Rabin, 2002).
Rather than employing one or other of the demand model
assumptions by default, one may take a step further and use a
statistical test, such as a unit root test, as a formal criterion for
making the distinction between stationary and non-stationary
demand processes (e.g., Nijs et al., 2001; Pauwels et al., 2002).
When sample size is small, it is again difcult to choose a correct
demand model, as conventional unit root tests have low statistical
power in a nite sample (Diebold and Rudebusch, 1991; DeJong
et al., 1992). To sum up, the underlying demand model cannot be
identied with certainty using both contextual expertise and a
statistical rationale in small samples. Furthermore, it may well becostly to wait for more periods to pass and obtain more data in
order to identify the trend more clearly. However, as the required
inventory levels behave much differently for one demand model
compared to the other, the incorrect demand model results in the
under or overestimation of inventory levels, leading to increased
inventory costs.
We propose an inventory policy that directly incorporates the
inherent uncertainty over stationary and non-stationary demand
models in response to new marketing efforts, by using Bayesian
model averaging. Bayesian model averaging is a complete Baye-
sian solution to average over possible models. The concept of
Bayesian model averaging was introduced by Leamer (1978), and
has recently received signicant attention in the statistics and
econometrics literature, in particular from Raftery et al. (1997),
Hoeting et al. (1999), andRaftery and Zheng (2003). We assume
that one of the stationary or non-stationary models is the true
demand model once new marketing efforts are introduced, but
that we do not know which it is. Starting from a prior about which
model is true and observing demand, we compute the posterior
probabilities that each is the true model by applying Bayes'
theorem. We then average over the inventory decisions made by
the two models, weighted by each model's posterior probability. In
this paper, structural results of the proposed inventory model arealso discussed. Specically, the Bayesian model averaging inven-
tory model estimates consistent order quantities based on a
critical fractile and provides better performance, as measured by
a logarithmic scoring rule, than using any single model.
The paper in hand relates to several studies that use a Bayesian
framework to deal with parameter uncertainty for specic
demand models (see, e.g., Azoury, 1985; Azoury and Miyaoka,
2009; Lovejoy, 1990). Using Bayes theorem, the unknown para-
meter is periodically updated based on newly obtained demand
observations. Yet these inventory models cannot address uncer-
tainty about the structure of the underlying demand generating
model, i.e., demand model uncertainty. In this paper, we use the
Bayesian framework to update the belief about candidate demand
models on the basis of past observations to explicitly account for
model uncertainty, which in our case arises from uncertainty
about the nature of demand uctuations after new marketing
efforts. While non-parametric approaches (e.g., Bookbinder and
Lordahl, 1989; Levi et al., 2007) are established to negate the need
to make assumptions about the demand model, they are limited
to independent demand processes and cannot be applied to
serially correlated demand processes. A semi-parametric approach
inLee (2014) provides consistent estimates of the critical fractile
independently of a forecasting model if the demand process
follows a stationary autoregressive demand process and the
forecasting model is within the autoregressive integrated class.
Unlike the non-parametric and semi-parametric inventory models,
our proposed Bayesian model averaging (BMA) inventory model
enables us to deal with model uncertainty in independent, serially
correlated, as well as non-stationary demand processes. As such,
our paper can be seen as a rst step towards suitably modifying
and adapting the recent developments in the Bayesian model
averaging method seen in the statistics and econometrics litera-
ture to the practical problem facing inventory management,
namely that of setting inventory levels in response to new
marketing efforts.
The interaction between two functional areas, marketing and
operations, is recurrently discussed in the literature. See, for
example,Tang (2010),Ma et al. (2013), andMarques et al. (2014).
We contribute to this type of literature focusing on the issue of
marketing efforts and ordering decisions. In particular, our work is
linked to the literature that addresses the classical single-period
inventory problem with advertising, where advertising stimulates
the demand. Khouja and Robbins (2003) assume that the meandemand is both increasing and concave in advertising expenditure
(i.e., the returns of advertising have a diminishing effect on sales)
and demand variance is also a function of advertising expenditure.
They obtain the optimal advertising expenditure and ordering
quantity that maximizes the expected prot or the probability of
achieving a target prot. Their model assumes that the demand
process is independent and the effect of advertising on the under-
lying demand is known. Lee and Hsu (2011) and Guler (2014)
recently extend this model to the distribution-free newsboy pro-
blem. In contrast, our model considers autocorrelated demand and
the effect of marketing actions on the mean and variance of
demand is characterized by the autocorrelation parameter. In most
practical situations, we shall indeed observe autocorrelation in the
demand process, especially when new marketing efforts are made
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289 279
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
3/12
(Dekimpe and Hanssens, 1995b). Also we do not assume that the
effect of marketing is known with certainty. Instead, we assume
that the total long-run impact of marketing is either temporary or
permanent and the managers learn about whether the observed
movements in demand are of a temporary versus a permanent
nature as more demand observations become available.
This paper is organized as follows: Section 2presents a single-
item inventory model. We analyze the key distinction in the
required inventory level when the demand shocks observed afternew marketing efforts are temporary or permanent. In Section 3,
we illustrate that it is difcult in practice to correctly interpret the
nature of demand shocks using both the context expertise and
statistical rationale.Section 4derives the inventory model resulting
from Bayesian model averaging that deals with the uncertainty
about the nature of the demand shocks, and discusses its structural
results.Section 5examines the performance of the BMA inventory
model using numerical studies.Section 6provides a conclusion. The
Appendix contains the proof of the asymptotic analysis inSection 4.
2. A single-item inventory model
This paper considers the problem of setting an inventory level
of a consumer product for which new marketing efforts are made.
Assume that att0, new marketing efforts (single or multiple) are
introduced and these efforts induce a series of unexpected move-
ments (shocks) in the following periods. The total long-run impact
of these movements on underlying demand is assumed to be
either temporary or permanent in nature. The pre-expenditure
demand level is denoted by Y0. For every period tZ0 new
demand observations become available, and the inventory man-
ager updates the demand forecast and inventory decisions accord-
ing to a base stock (order-up-to) policy based on a critical fractile.
In the following, we describe the demand process and inven-
tory control policy in more detail and highlight the key distinction
in the required inventory level between the cases where the
effects of the marketing efforts are temporary or permanent.
Given the well-established dynamic nature of marketing effectresponse in the marketing literature, we adopt the univariate rst-
order autoregressive demand process used in Dekimpe and
Hanssens (1995b) and make a distinction between stationary
and non-stationary demand processes arising from new marketing
efforts. We then show a multivariate extension to endogenize the
marketing efforts (such as measured by marketing expenditure) in
the demand model.
2.1. Demand process
The demand process is a univariate rst-order autoregressive
(AR(1)) demand and is given as follows:
Yt cYt 1ut; 1where c is a nonnegative constant, is an autocorrelationparameter and ut are random shocks that are independent and
identically distributed from N0; 2u. For ease of exposition, wehave omitted any deterministic components from the model
except a constant term c. In general, deterministic trends and/or
seasonal dummy variables can be added to (1) (Dekimpe and
Hanssens, 1999). Specically, we can describe the mean of a time
seriestas the sum of the level componentCt, trend componentTtand seasonal component St. For example, we can write
t Ct Tt St where Tt t, St si 1iDti. In this case, is a
linear trend parameter, t is a time index, is a seasonal trendparameter, andDtiare the seasonal dummies. InSection 6, we discuss
the generalization of this demand process by including higher auto-
regressive lags (AR(p)).
Applying successive backward substitutions (Box et al., 1994),
we can rewrite(1) as in innite moving average form:
Yt c=1 utut 12
ut 2: 2
Present demand Yt is therefore explained as a weighted sum of
present and previous random shocks, u t; ut 1; ut 2;.
It is the unknown value ofthat denes the nature of demandshocks, either temporary or permanent due to the marketing
efforts. If jjo1, the effect of past uctuations wanes andeventually becomes negligible. The observed uctuations are
temporary in nature and do not alter the underlying long-term
demand trend. In this case, the demand process is stationary as it
reverts to a long-term xed mean, EYt c=1, and it has anite variance, VarYt 2u=1
2. Also note that the speed of
mean reversion is faster for the low value of . When 0, theunderlying process is simply independent and identically
distributed.
If 1,(2) becomes
Yt ctut ut 1ut 2:
In this case, the impact of observed shocks is permanent and
indenitely accumulated over time. The long-term mean does not
exist and is unbounded. The variance depends on tand increaseswith time without bound, i.e., VarYt 2ut. Such a demand seriesis non-stationary and contains a unit root. It moves freely in one
direction or another without any reversion to a xed trend.
2.2. Inventory control policy
The inventory manager operates with a base stock policy based
on a critical fractile (e.g., Graves, 1999), in which the inventory
position is brought to a prescribed level at the beginning of each
period to meet a desired consumer service level K. No backorders
are assumed. Therefore, (1 K) is the probability of stockout
during a replenishment cycle, denoted by L. Note thatKis dened
by K s=s h where h and s are the proportional holding and
shortage costs, respectively, following the well-known newsven-
dor result (Porteus, 2002).
LetQnbe the required base stock level to meet a desired service
level Kat tn, i.e., n historical demands are observed since new
marketing efforts are introduced:
QnLnz
Ln; 3
where nL and n
L are the mean and standard deviation of LTD,
respectively. z is a constant chosen to meet the desired service
levelK, i.e.,z 1Kwhere is a standard normal distributionfunction. The base stock level in each period is therefore set to
support the expected demand for replenishment lead time nL plus
a safety stock SSLnzLn to mitigate the risk of stockouts due to
demand uncertainty, measured by nL .
2.3. Base stock levels under stationary and non-stationary demand
processes
In this section, we demonstrate that the required base stock
levels behave much differently for the case of stationary demand
than for non-stationary demand, which is driven by the nature of
random shocks observed after new marketing efforts. We demon-
strate that the nature of demand signals can affect the base stock
levelQnin two ways: rst, by directly altering the expected LTD nL
(mean effect); and second, by altering the safety stock level
requirement SSLnzLn (variance effect).
When the demand process follows a rst-order autoregressive
process, as described in(1), the LTD, which is the sum of demand
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289280
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
4/12
over periods n 1; n2;; n L, is expressed as follows:
Yn 1Yn 2 Yn L
cYnun 1 c1 2
Ynun 1 un 2
c1 2
L 1
L
YnL 1
un 1 L 2
un 2 un L:
4
The expected LTD is then found by setting the expected values of
the future random terms in(4), un 1; un 2;; un L, to zero:
Ln c1 1 1 2
L 1
Yn2
L
c L 1
i 0
i
j 0
j" #
Yn L
i 1
i: 5
First, note that the expected LTD is dependent on the current
demand state in period n, i.e., Yn. When jjo1, the currentdemand state is discounted and its effect on Yn L eventually becomes
negligible as L increases, i.e., EYn L c12
L 1
LYn . Therefore, future demands Yn L, especially for large L, tendto revert back to the long-term mean.
When 1, the expected LTD becomes
Ln c L 1
i 0
1i" #LYn L L 12
cYn : 6Under a non-stationary process, the current demand state Ynis an
anchor for all future demands and in each period a constant drift c
is added to this anchor point during the replenishment lead time,
i.e., EYn L cL Yn. Therefore, the expected LTD under a non-
stationary process may be signicantly different from under a
stationary process, depending on the value of, and the differenceincreases as lead time increases.
Next, we demonstrate how the safety stock level under non-
stationary demand will be more conservative than under station-
ary demand; that is to say, an inventory manager will hold higher
levels of safety stock under the non-stationary process to achieve
the same service level as in the stationary process. Using (4), we
can express the total LTD uncertainty, denoted by EnL , as a linear
combination of future random terms un 1; un 2;
; un L:
ELn un L1un L 11 2
L 1
un 1:
Since the random errorsutare assumed to be uncorrelated in (1),
the standard deviation of LTD is given by
Ln u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi11 2 1
2L
12
q
u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
L 1
i 0
i
j 0
j !2vuut : 7
For a given lead time, the standard deviation of LTD is an
increasing function of the autocorrelation parameter jj. When1 and the underlying demand process is non-stationary, (7)
becomes
Ln u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
L 1
i 0
1 i2
s u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiLL 12L 1
6
r : 8
This implies that the relationship between nL and L becomes
convex for the non-stationary demand process (Graves, 1999).
Since the required safety stock level is proportional toLnfor a pre-specied service level, i.e., SSLnz
Ln where z
1K and is a
constant, we require dramatically more safety stock for non-
stationary demand compared with fast mean reverting-
stationary demand. The difference in safety stock requirements
is larger for longer lead times. We note that, unlike the expected
LTD, the standard deviation of LTD is independent of the current
demand state Yn. For a given demand process and a xed lead
time, the required safety stock is constant over different periods.
2.4. A multivariate extension
In practice, companies respond to perceived changes in the
market environment or in their goals by adjusting their marketing
mix. In the empirical study conducted by Dekimpe and Hanssens
(1995a), the adjustments in the marketing mix are shown to be
either temporary, in that the company abandons the change in
favor of the previous level after some periods, or permanent, in
that there is no return to the previous level. It is therefore ofimportance to examine to what extent the persistence of demand
can be related to an unexpected change in marketing activity. For
this purpose, we use vector autoregressive (VAR) models to model
the dynamic interactions between performance variable (in our
case demand) and marketing variables (such as measured in
marketing expenditure) (Dekimpe and Hanssens, 1995b). Both
demand and marketing actions are endogenous so that they are
explained by their own past and by the past of the other
endogenous variables. Specically, the vector autoregressive
model estimates the baseline of each endogenous variable and
forecasts its future values on the basis of the dynamic interactions
of all jointly endogenous variables.
For ease of exposition, consider the bivariate, rst-order vector
autoregressive model:
Y1;t
Y2;t
" #
c1
c2
" #
11 1221 22
" # Y1;t 1
Y2;t 1
" #
u1;t
u2;t
" #
whereY1;tand Y2;tdenote the underlying demand and marketing
action at time t, respectively, and u1;t and u2;tare white noises in
the respective processes. Equivalently, we write the vector auto-
regressive model in the following form:
Yt C Yt 1Ut; 9
whereYt,CandUtare 2 1 vectors and is a 2 2 matrix. This
specication directly captures purchase reinforcement (11),delayed response (12), performance feedback (21) and inertia
in marketing decision making (22) (Dekimpe and Hanssens,1999). When12 21 0, the underlying demand and marketingaction are independent of each other.
Using (9), we can express the mean and variance of demand
and marketing mix for the lead time as
Ln L 1
i 0
i
j 0
j
AC L
i 1
iYn
and
Ln
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
L 1
i 0
i
j 0
j
A
!2vuut uwhere A
0 1
1 0 and
u is a vector of standard deviations of
u1;t; u2;t0.
When dealing with evolving variables, Yt in(9) is replaced by
YtYtYt 1. In this case, the mean and variance of demand
and marketing mix for the lead time are
Ln L L 1
2 C Yn
and
Ln
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiLL 12L 1
6
r u:
Since the results from the univariate model can be easily
extended to the multivariate model using (3), the rest of this
paper focuses on the exposition of the univariate model.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289 281
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
5/12
3. Uncertainty about the effect of new marketing efforts on
product demand
In this section, we discuss the inherent uncertainty in deter-
mining the nature of demand uctuations arising from new
marketing efforts, which leads to demand model uncertainty.
As illustrated inSection 2(due to both mean and variance effect),
stationary and non-stationary demand processes require much
different base stock levels and it is crucial for efcient inventoryplanning that we distinguish between the two processes. An
incorrect demand model assumption leads to under or overestim-
ated base stock levels and subsequently increased inventory costs.
Let us imagine that a company launches a new advertising
campaign for a product at t0 and observes the effect this has on
product demand over a sustained period. Fig. 1 plots simulated
autoregressive demand processes (as given by (1)) to illustrate
instances where the demand shocks resulting from this marketing
action are temporary or permanent in nature. We have constructed
these series using the same random process, i.e., the same simu-
lated errors utare used. The only difference being the value of the
autocorrelation parameter. If the marketing effect is found to betemporaryo1, the demand process tends to revert to the long-term mean and follows the dotted line. The higher values of the
autocorrelation parameter correspond to the slower rate of the
mean reversion rate. When the demand is highly autocorrelated
and the mean reversion rate is slow, we call the underlying demand
process to be nearly non-stationary. In contrast, if advertising has a
permanent effect (1), the demand process diverges over timewithout mean reversion and follows the solid line. Note that the
gure is constructed based on one simulated random process,
therefore illustrates one possible scenario.
As observed inFig. 1, the stationary demand process with the
slow mean reversion rate follows a similar path to the non-
stationary demand process for up to 60 periods. It would therefore
have been very difcult to interpret whether or not the observed
movements in demand during these periods were temporary
uctuations around a xed mean. This difculty is also known as
the law of small numbers (Camerer, 1989; Rabin, 2002). How-
ever, observe that the consequence of an incorrect demand model
by under or overinterpreting short-run uctuations as indicators
of future demand patterns is very signicant in the long run. As we
review more data, the distinction between the two processes
becomes clearer.
Rather than employing one or other demand model by sub-
jective judgment, one may use a statistical test such as a unit root
test as a formal criterion for differentiating stationary models from
non-stationary models (e.g.,Nijs et al., 2001; Pauwels et al., 2002).
One popular unit root test, a DickeyFuller test (Dickey and Fuller,
1979), hypothesizes whether or not there is a unit root in the rst-
order autoregressive process. The test is based on the t-ratio
DFn 1
s ; 10
where is the ordinary least squares estimator of the autocorre-lation coefcient ands is its standard error. IfDFnrrwhererisa critical value, the null hypothesis of a unit root is rejected and a
stationary demand model is chosen. Otherwise, we assume a non-
stationary demand model.
This formal statistical test is also known to have low power
against nearly non-stationary processes in small samples (DeJong
et al., 1992; Diebold and Rudebusch, 1991). Fig. 2 illustrates
the empirical power of the DickeyFuller test against the station-
ary alternative with a 5% signicance level. We observe that the
test has particularly low statistical power when the value of isvery close to unity and the sample size is small. For example, when
0.9 and the sample size is 40, the likelihood of specifyingthe correct stationary model is as low as 10% and there is an
approximate 90% probability ofnding a seemingly non-stationarypattern (Type II error). When the underlying process is non-
stationary, the test correctly detects a unit root with a 95%
probability, which is consistent with the chosen signicance level.
4. Bayesian model averaging inventory model
We propose an inventory model that explicitly acknowledges
uncertainty over stationary and non-stationary demand models as
a result of new marketing efforts, using Bayesian model averaging.
We rst introduce the general concept of Bayesian model aver-
aging and apply this to make robust inventory decisions where
there is uncertainty about the persistence effect of marketing
efforts on the product demand.
4.1. Bayesian model averaging
Bayesian model averaging is an overall concept that directly
incorporates model uncertainty in the estimation process. Con-
sider a set of m models M fM1;; Mmg. We know that one of
these models is the true model, but do not know which one it is.
We have prior beliefs about the probability that the ith model is
Time
Simulateddeman
d
0 20 40 60 80 100
200
250
300
=1
=0.9
=0.8
=0.7
Fig. 1. Simulated rst-order autoregressive demand processes where 1 (solid
line), 0.9, 0.8, 0.7 (dotted lines), and u 6.
20 40 60 80 100 120 140 160
0
20
40
60
80
100
Sample size
Likelihoodofcorrectmodelsp
ecification(%)
=0.9
=0.8
=0.7
Fig. 2. Empirical power of the DickeyFuller test against the stationary alternative
with a 5% signicance level u 1.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289282
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
6/12
the true model, which is denoted by PrMi. Given the observed
data D, we then update our beliefs to compute the posterior
probabilities:
PrMijD PrDjMiPrMi
mj 1 PrDjMjPrMj;
where
PrDjMi Z
PrDji;MiPrijMidi
is the integrated likelihood of model Mi, i is the vector ofparameters of model Mi, PrijMi is the prior density ofi undermodel Mi, and PrDji;Mi is the likelihood.
Denote as the observable to be predicted; this could be afuture observation (in our case LTD) or the utility of a course of
action. Each model implies an estimate of. The estimates offrom all models are then weighted by their posterior probabilities
to compute the posterior distribution of given the data D:
PrjD m
i 1
PrjMi; DPrMijD: 11
4.2. Bayesian model averaging (BMA) inventory model
The model uncertainty we study in this paper results from
uncertainty about the nature of demand uctuations resulting
from new marketing efforts. The two natural candidate demand
models are, therefore, the stationary model MS jjo1 and thenon-stationary model MNS(1).
Given n demand observations, let Qn;S and Qn;NS be the base
stock levels that are based on the stationary and non-stationary
demand models, respectively. When the stationary model is
applied, we derive Qn;Susing(3), (5) and (7) as
Qn;S c L 1
i 0
i
j 0
j" #
Yn L
i 1
i z u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
L 1
i 0
i
j 0
j !2vuu
t : 12
When the non-stationary model is applied, we derive Qn;NSusing(3), (6) and (8)as
Qn;NS c L 1
i 0
1 i
" #LYnz u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
L 1
i 0
1i2
s : 13
Since Qn represents the K% quantile of LTD, the Bayesian model
averaging base stock level is calculated using (11) (Raftery and
Zheng, 2003):
Qn;BMA Qn;SPrMSjD Qn;NSPrMNSjD: 14
This implies that the proposed Bayesian model averaging (BMA)
inventory model weighs each of the two models by their posterior
probabilities to estimate the base stock level.
In order to employ the BMA inventory model, we need to
specify the model priors PrMiand the parameter priors PrijMi.Specication of PrMi, the prior distribution over competing
models, can be based on the subjective knowledge of the modeler.
When there is little prior information about the relative plausi-
bility of the models under consideration, the assumption that all
models are equally likely a priori is a reasonable neutral choice,
i.e., PrMS PrMNS 0:5. We assume equal model priors in our
simulation studies inSection 5.
The BMA inventory model not only updates beliefs about
model assumption but also updates model parameter values. For
this we need to set the parameter priors PrijMi and it is usefulto express the rst-order autoregressive demand model in(1) in a
vector form:
Y cX1u X u
whereYis a vector of observed demandYn; Yn 1;; Y10 andX1is
a vector of lagged demand, Yn 1 ; Yn 2;; Y00 since the introduc-
tion of the new marketing efforts at t0. We consider the (2 1)
individual parameters c;0 and u2 to be unknown. We then
use the standard normal gamma conjugate class of priors (Raftery
et al., 1997),
N; 2uV
and
2uX2:
Therefore,, , the 2 2 matrix V, and the 2 1 vector arehyperparameters to be selected. Based on the chosen priors, we
can calculate the required marginal likelihood of model Mianalytically, followingMadigan and Raftery (1994):
PrYi;Vi; Xi; Mi
n
2
=2
n=2
2
jIXiViX
ti j
1=2 YXii
t I XiViXti
1Y Xii n=2:
Next we discuss the choice of the hyperparameters (Raftery et al.,
1997). For the stationary model, we choose S c; 0, where c isthe ordinary least squares estimate of c. The prior covariance
matrix for Sis equal to
CovS 2uVS
2u
s2Y 0
0 2s 2X1
!;
wheres2Ydenotes the sample variance ofY,s2
X1denotes the sample
variance ofX1, andis another hyperparameter to be selected. Theprior variance ofcis chosen conservatively and that of is chosento reect increasing precision about this parameter as the variance
ofX1 increases. For the non-stationary model, we need to impose
the unit root restriction, i.e., 1. Therefore,NS c; 1 and
CovNS 2
uVNS 2
u
s2Y 0
0 0 !
:
In the simulation study that follows, we choose the remaining
hyperparameters , and as suggested by Raftery et al. (1997),i.e., we set2.58,0.28, and2.85.
Next, we show that the posterior probability of the true model
tends to 1 as the sample size goes innity (Theorem 1). This
implies that the proposed BMA inventory model estimates con-
sistent order quantities based on a critical fractile (Corollary 1).
Also,Theorem 2shows that in the presence of model uncertainty,
using the BMA inventory model provides better performance in
estimating inventory levels, as measured by a logarithmic scoring
rule, compared to by using any single modelMSorMNS.Theorem 3
(Theorem 4) shows the relative size of inventory levels made by
the stationary, BMA and non-stationary models when the desiredservice level is higher (lower) than 50%. See the Appendix for
proofs.
Theorem 1. If MS is the true model and PrMS PrMNS 0:5,
limn-1PrMSjD 1 and vice versa.
Corollary 1. If MS is the true model and PrMS PrMNS 0:5,
limn-1Qn;BMA Qn;Sand vice versa.
Theorem 2. Assume that Qn is the critical fractile of the true demand
model. Then, we have Elog fPrQnjDgrElog fPrQn;BMAjDgr
Elog fPrQn;jjDg where j S; NS.
Theorem 3. Assume c;; Yn40. For a desired service level KZ0:5,
Qn;SoQn;BMAoQn;NS.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289 283
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
7/12
Theorem 4. Assume c;; Yn40. For a desired service level Ko0:5,Qn;SoQn;BMAoQn;NS if
cL 1i 0 f1i i
j 0j
g YnL Li 1
j
uffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiL 1i 0f1i
2 ij 0j
2g
q 4 11K:Otherwise, Qn;NSoQn;BMAoQn;S.
5. Numerical results
5.1. Experimental design
In this section we investigate the nite sample performance of
the BMA inventory model using Monte Carlo experiments, where
there is uncertainty about the persistence of new marketing efforts
on customer demand. We benchmark its performance against four
alternative inventory models:
1. The stationary inventory model (S): This approach employs a
stationary demand model by default and ignores demand
model uncertainty. Given n demand observations, it estimates
base stock levels Qn;S using (12) with point estimates of the
parameter vector, c; ; 2
u0 where c and are the uncon-
strained least squares estimators, and 2u is the estimatedvariance of the demand model residuals. Since the underlying
process is assumed to be stationary, these model parameters
are estimated using the level of the demand variable rather
than its rst difference.
2. The non-stationary inventory model (NS): This approach
employs a non-stationary demand model by default and also
ignores model uncertainty. It imposes a unit root restriction
( 1) and estimates base stock levels Qn;NS using (13) withpoint estimates of the parameter vector, c; 2u
0 where c is the
constrained least squares estimator, and 2u is the estimatedvariance of the demand model residuals. Since the underlying
process is assumed to be non-stationary, these model para-
meters are estimated using the rst difference of the demand
variable.3. The pre-testing inventory model(PT): Rather than employing one
or other demand model by default, this model uses a unit root
test as a formal criterion for selecting a demand model. We use
the DickeyFuller test, which is based on the t-ratio DFn as
dened in(10). The pre-testing inventory model estimates base
stock levels by Qn;PT Qn;SIDFnrr Qn;NSIDFn4r, where r
is a critical value and I is an indicator function. We select the
value ofrfor a 5% signicance level.
4. The Bayesian inventory model (B): This approach uses a Bayesian
framework to allow for parameter uncertainty in the stationary
demand model. It uses a Bayesian formulation to update
parameters of the stationary demand model based on the
normal gamma conjugate class of priors from standard Baye-
sian theory (Rossi et al., 2005). Given n demand observations,the Bayesian inventory model estimates base stock levels Qn;Busing(12) with the updated parameter values ~c; ~; ~2u
0 using
Bayes' theorem. It acknowledges errors arising from parameter
estimation in the stationary inventory model, yet disregards
demand model uncertainty.
Note that the BMA inventory model nests the Bayesian model
that updates base stock levels with respect to parameter uncer-
tainty. The BMA model learns about not only the correct model
specication (stationary versus non-stationary), but also the
parameter values of each model as the sample size grows.
A simple averaging of the stationary and non-stationary models
by setting the posterior probabilities PrMSjD PrMSjD 0:5 in
(14)was also considered, but the results of this are omitted as the
process underperforms that of the BMA inventory model. In
addition, we considered using different prior probabilities in the
BMA inventory model, such as by setting the prior probabilities
based on the p-values of the DickeyFuller unit root test. The
results of this are also omitted as the process underperforms that
by setting equal prior probabilities (i.e., PrMS PrMNS 0:5).
As discussed in Section 3, we focus on the nearly non-
stationary demand process, for which it is difcult to identify
the correct nature of observed demand shocks. The nearly non-stationary process is designed with a set of high-value autocorre-
lation parameters, Af0:70; 0:75; 0:80; 0:85; 0:90; 0:95; 1g. Recallthat the value of this autocorrelation parameter determines the
nature of demand process (o1 and 1 correspond to astationary and non-stationary process, respectively). In the simu-
lations, this value is treated to be unknown and thus there is
0.70 0.75 0.80 0.85 0.90 0.95 1.00
5
0
5
10
Autocorrelation parameter,
Averagebias(%)
S
NS
PT
B
BMA
Fig. 3. Average single-period bias in base stock levels (%) compared to the full-
information policy.
0.70 0.75 0.80 0.85 0.90 0.95 1.00
0
20
40
60
80
100
Autocorrelation parameter,
Averagecost
error(%)
S
NS
PT
B
BMA
Fig. 4. Average single-period cost error (%) compared to the full-information policy.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289284
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
8/12
uncertainty about the persistence of new marketing efforts on
demand. The performance of each inventory model is evaluated in
terms of how far the estimated base stock levels and one-period
inventory costs are from the optimal value Qnn and its cost CQn
n,
which are calculated using the true demand model, i.e.,cand areknown with certainty (full-information policy).
The parameters kept constant for all simulations are u 4,K0.95 andL 10. Ifo1, the constantcin the demand model is
set by c EYt1 with EYt 100. This ensures that thedemand series with different values of revert to the same meanEYt. When1, we simply set c0.
Given these parameters, we rst generate a demand path for
n30 consecutive time periods. For each sample path, we com-
pute the base stock level Qn using each inventory model for the
upcoming replenishment time (from t n to t n L). We then
calculate the bias of the estimated base stock level QnQn
n=Qn
n
and cost error C Qn CQn
n=CQn
n of each inventory model. We
report the average bias of base stock level and average cost error,
which are calculated over 1000 sample paths. We also present a
sensitivity analysis in Section 5.3 to examine the impact of the
assumed parameters, such as sample size, lead time and service
level, on inventory cost error.
5.2. Results
Figs. 3 and 4 compare the BMA inventory model to the alter-
native models (stationary (S), non-stationary (NS), pre-testing
(PT), or Bayesian (B)). The average bias in base stock levels and
average cost error of each inventory model are plotted as a function
of the autocorrelation parameter . The observed bias and costerrors are calculated in reference to the full-information case and
therefore are due to parameter and/or model uncertainty about the
demand process when estimating inventory policies.
As shown in Fig. 3, the stationary inventory model (S) under-
estimates base stock levels. This is due to the downward bias in safety
stock levels as the least squares estimators of are biased towardzero, especially for the value of close to unity (Shaman and Stine,1988). This inventory model (S) therefore assumes the unnecessarily
stronger reversion of demand and set a lower level of safety stock. We
also observe that the bias created in the stationary inventory model isreduced by the Bayesian approach (B), which integrates uncertainty
over parameter values into the base stock level calculation.
The non-stationary inventory model (NS) overestimates safety
stock levels for the underlying stationary demand processes
o1. This implies that the non-stationary inventory modelgenerates unnecessary safety stocks by assuming higher demand
uncertainty, especially when the underlying stationary demand
process has relatively low autocorrelation. The pre-testing inven-
tory model (PT) noticeably dominates the non-stationary inven-
tory model, as it avoids safety stock overestimation by incorrectly
selecting the non-stationary demand model when the unit root
test correctly rejects the null of the non-stationary process.
Distortions of the pre-testing inventory model still exist as a result
of the low power of the DickeyFuller test in small samples.
The BMA inventory model (BMA) sets the base stock levels in
between the non-stationary model (NS) and the stationary model
(S) as Theorem 3predicts. Fig. 3 suggests that both the Bayesian
inventory model (B) and BMA inventory models (BMA) are close to
the full-information case on average. For lower values of, theBayesian inventory model outperforms the BMA model, but for close to 1, the reverse is true.
Fig. 4 presents the average cost errors of the ve inventory
models. It is consistent with the results in Fig. 3 as the smaller
20 40 60 80 100 120 140
0
20
60
100
140
=0.7
Sample size, n
Averagecosterror(%
)
PT B BMA
20 40 60 80 100 120 140
0
20
60
100
140
=0.8
Sample size, n
Averagecosterror(%
)
PT B BMA
20 40 60 80 100 120 140
0
20
60
100
140
=0.9
Sample size, n
Averagecosterror(%)
PT B BMA
20 40 60 80 100 120 140
0
20
60
100
140
=1
Sample size, n
Averagecosterror(%)
PT B BMA
Fig. 5. Impact of sample size n on average cost error (%) compared to the full-information policy.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289 285
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
9/12
absolute value of bias corresponds to the lower cost error. This
gure demonstrates the robustness of the BMA inventory model
against the nature of demand shocks, i.e., different values of. Itsmaximum value loss is 45% in all cases examined. The perfor-
mance edge of the BMA model is attributable to incorporating
model uncertainty directly into the estimation process. In contrast,
the performance of alternative models can deteriorate signicantly
depending on the value of. For the stationary demand process
with small values of , the stationary and Bayesian inventorymodels perform well. However, their performance deteriorates as increases, as they tend to underestimate demand uncertaintyand stockouts are frequent. In contrast, the performance of the
non-stationary and pre-testing inventory models is poor for small
values of because of overestimated demand uncertainty, result-ing in high holding costs; however their performance improves as
increases. This is consistent with the ndings in Fig. 3, as theconstant upward and downward biases in the base stock estimates
increase overall inventory costs.
5.3. Sensitivity analysis
We examine the effectiveness of the BMA inventory model
across values of sample size, lead time and service level. Forbrevity, we use the Bayesian and pre-testing inventory models as
benchmarks, as they tend to be more robust than the stationary
and non-stationary models as indicated in Figs. 3 and 4.
Fig. 5shows the cost error of the BMA model as a function of
sample size compared with the Bayesian and pre-testing models.
There are four panels for different values of{0.7,0.8,0.9,1}. Eachpanel, therefore, represents a different mean reverting rate in the
demand process. We observe that the cost error of all three
inventory models decreases to zero as the sample size increases.
The cost error of the BMA inventory model converges to zero as
the posterior distributions of the model probabilities and model
parameters are more accurately updated with a larger sample size
(Corollary1). Similarly, the model parameter values in the Baye-
sian inventory model are better updated in larger samples. As
discussed in Section 3, the DickeyFuller test has a higher
statistical power and correctly distinguishes between the station-
ary and non-stationary processes as the sample size increases.
Subsequently, the performance of the pre-testing model improveswith sample size.
Fig. 5also suggests that cost savings from the BMA inventory
model may be substantial for smaller samples no50. The BMA
approach is therefore useful in improving the efciency of an
inventory system where there is only a limited amount of available
historical demand data and it is prohibitively costly to wait for
time to pass in order to make a more informed decision.
Fig. 6 illustrates that the cost errors of the inventory models
increase as the lead time increases. This is because errors in
forecasting LTD, resulting from parameter uncertainty and model
uncertainty, accumulate over lead time. A reduction in lead time
will enable cost performance to be less sensitive to the additional
uncertainty. If a reduction is not achievable in practice, the BMA
inventory model is preferred among those considered due to its
robustness. Fig. 7 illustrates the impact of service level K on
inventory costs. This explains the general tendency observed for
cost error to be greater at higher service levels. When Kis high, an
inventory manager may benet from substantial cost savings by
using the BMA inventory model.
We also ran simulations to test for the impact of demand
volatility, measured byu, on cost sub-optimality. We found littleeffect on cost error as a function of u for any of the threeinventory models.
5 10 15 20
0
50
100
150
200
=0.7
Lead time, L
Averagecosterror(%
)
PT B BMA
5 10 15 20
0
50
100
150
200
=0.8
Lead time, L
Averagecosterror(%
)
PT B BMA
5 10 15 20
0
50
100
150
200
=0.9
Lead time, L
Averagecosterror(%)
PT B BMA
5 10 15 20
0
50
100
150
200
=1
Lead time, L
Averagecosterror(%)
PT B BMA
Fig. 6. Impact of lead time L on average cost error (%) compared to the full-information policy.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289286
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
10/12
6. Conclusion
In practice, new marketing efforts introduce random uctua-
tions in the underlying demand process. False interpretation of the
nature of demand signals leads to under or overestimation of
inventory levels, and subsequently increased inventory costs. In
this paper, we discussed the inherent uncertainty in determining
the nature of demand uctuations, i.e., demand model, when a
reversion to a long-term trend in the underlying demand process
is very gradual and there is insufcient historical data since new
marketing efforts are introduced. Under these conditions, there is
a high probability of nding a spurious non-stationary pattern
using both contextual expertise and prevailing statistical ratio-
nales. We therefore face the risk of choosing an incorrect demand
model when making inventory decisions.
We proposed an inventory model based on Bayesian modelaveraging to better address uncertainty about the demand model
when estimating required base stock levels based on a critical
fractile. In particular, we consider the case where demand
becomes either stationary or non-stationary in response to new
marketing efforts. We show that the proposed BMA inventory
model estimates consistent base stock levels and it provides better
predictive performance, as measured by a logarithmic scoring rule,
than by using any single demand model (either stationary or non-
stationary model). The simulation results conrm the structural
results and suggest that the BMA inventory model carries impress-
ively low risk and is the most robust option among the feasible
inventory models we considered. The BMA model can bring a
signicant improvement to an inventory system, particularly for a
system characterized by highly autocorrelated demand, small
sample sizes, long replenishment lead times and high service
levels.
This paper applies Bayesian model averaging to address uncer-
tainty about the nature of demand signals in the rst-order
autoregressive (AR(1)) demand process. We are able to generalize
the proposed inventory model in more complex demand processes
that include other autoregressive lags (AR(p)):
Yt c1Yt 12Yt 2pYtput;
wherec is a constant and p is the lag order of the autoregressive
process. As the stationary/evolving character of a series is still
determined by its level of integration, we can derive stationary
and non-stationary models based on whether or not the autore-
gressive polynominalL 11LpLp has a root on the
unit circle. We can then average over the inventory decisions made
by the two demand models to make more robust inventorydecisions. Note that any deterministic components of the model,
such as deterministic trends and seasonal dummy variables, can be
removed prior to applying the proposed approach and added back
again to the resultant order quantities. More generally still,
Bayesian model averaging can be applied to address other types
of demand model uncertainty when estimating base stock levels,
such as uncertainty about potential predictors in a linear regres-
sion demand model (e.g., Azoury and Miyaoka, 2009).
There are some limitations of using Bayesian model averaging.
Specically, there is a lack of consensus regarding how to use
subjective prior information and, subsequently, how to choose the
set of candidate models, how to assign prior probabilities to each
candidate model, and how to select prior distributions for para-
meters. This is a general concern in the growing body of Bayesian
0.80 0.85 0.90 0.95
0
50
100
150
=0.7
Service level, K
Averag
ecosterror(%)
PT B BMA
0.80 0.85 0.90 0.95
0
50
100
150
=0.8
Service level, K
Averag
ecosterror(%)
PT B BMA
0.80 0.85 0.90 0.95
0
50
100
150
=0.9
Service level, K
Averagecosterror(%)
PT B BMA
0.80 0.85 0.90 0.95
0
50
100
150
=1
Service level, K
Averagecosterror(%)
PT B BMA
Fig. 7. Impact of service level Kon average cost error (%) compared to the full-information policy.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289 287
-
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
11/12
model averaging literature and calls for further research. Despite
these limitations, it is often shown that Bayesian model averaging
provides better average predictive performance than any single
model in a range of practical applications (Hoeting et al., 1999). In
addition, there is the computational burden involved in the
construction of posterior distributions where there is no closed-
form solution to the marginal likelihood of the model. With the
development of computational power, the process of calculating
model likelihoods should not be an obstacle to the practicalimplementation of the BMA inventory model.
There are a number of directions in which it would be of
interest to extend the proposed Bayesian model averaging
approach to. These could include extensions to a multi-period
setting and dealing effectively with censored demand data. Finally,
it is important to examine the empirical validity of the Bayesian
model averaging model using real demand data in the future
research.
Acknowledgments
The author thanks editor Edwin Cheng and an anonymous
referee for their thoughtful, clear, and constructive feedback. She is
also thankful for many valuable comments made by Stefan
Scholtes, Danny Ralph and James Taylor.
Proofs
Proof of Theorem 1. Assume that MS is the true model. When
prior probabilities PrMS PrMNS 0:5 are used, the posterior
probability for MSis
PrMSjD PrDjMSPrMS
PrDjMSPrMS PrDjMNSPrMNS
1
1 PrDjMNS
PrDjMS
:
15
SinceMSandMNSare nested models and proper priors are used for
the model parameter, the Bayes factor ofMNSwhen compared with
the true model,MSin this case, tends in probability to zero as the
sample size tends to innity (O'Hagan and Forster, 2004):
limn-1
PrDjMNS
PrDjMS 0:
By substituting this into (15), we have limn-1PrMSjD 1.
Similarly, limn-1PrMNSjD 1 can be shown when MNS is the
true model.
Proof of Corollary 1. This immediately follows from Theorem 1
and(14).
Proof of Theorem 2. From the non-negativity of the Kullback
Leibler information divergence, Elogfmi 1 PrjMi; DPrMijDgrElogfPrjMj; Dg forj 1;; m (Madigan and Raftery, 1994).Therefore, we have Elog fPrQn;BMAjDgrElogfPrQn;jjDg for
j S;NS. Since a logarithmic scoring is proper, with a negative sign the
logarithmic scoring is minimized when reporting the true probability,
PrQnjD. Therefore, we have ElogfPrQnjDgrElogfPrQn;BMAjDg.
Proof of Theorem 3. Qn;SoQn;NS holds only if Ln;NS
Ln;S
zLn;NSLn;S40. By rearranging, we need to satisfy the condition
Ln;NSLn;S
Ln;NSLn;S
4 z:
First note that Ln;SoLn;NS and
Ln;So
Ln;NS hold since n
L andnL in
(5) and (7) are increasing functions of for non-negative values of
c;; Yn. Secondly, for KZ0:5, the value ofz 1
K is positive or
equal to zero. Therefore, the above condition holds trivially and we
have Qn;SoQn;NS. Also, 0oPrMSjDo1 and PrMNSjD 1
PrMSjD. Since Qn;BMA is dened by (14), we have Qn;SoQn;BMAoQn;NS.
Proof of Theorem 4. Qn;SoQn;NSholds only if
Ln;NSLn;S
Ln;NSLn;S4z:
For Ko0:5, the value of z is negative. Therefore, this condition
does not hold trivially. Since z 11 K and from(12) and(13), we can rewrite this condition as
cL 1i 0 f1 i i
j 0j
g YnL Li 1
j
u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiL 1
i 0f1i2 ij 0
j2g
q 4 11 K:We can then use the same argument as above to show
Qn;SoQn;BMAoQn;NS. Similarly, we can show the condition for
Qn;NSoQn;BMAoQn;S.
References
Azoury, K.S., 1985. Bayes solution to dynamic inventory models under unknowndemand distribution. Manag. Sci. 31, 11501160.
Azoury, K.S., Miyaoka, J., 2009. Optimal policies and approximations for a Bayesianlinear regression inventory model. Manag. Sci. 55, 813826.
Bookbinder, J.H., Lordahl, A.E., 1989. Estimation of inventory re-order levels usingthe bootstrap statistical procedure. IIE Trans. 21, 302312.
Box, G.E.P., Jenkins, G.M., Reinsel, G.C., 1994. Time Series Analysis: Forecasting andControl. Prentice-Hall, Englewood Cliffs, NJ.
Camerer, C.F., 1989. Does the basketball market believe in the hot hand? Am.Econ. Rev. 79, 12571261.
Charnes, J.M., Marmorstein, H., Zinn, W., 1995. Safety stock determination withserially correlated demand in a periodic review inventory system. J. Oper. Res.Soc. 46, 10061013.
Chen, A., Blue, J., 2010. Performance analysis of demand planning approaches foraggregating forecasting and disaggregating interrelated demands. Int. J. Prod.Econ. 128, 586602.
DeJong, D.N., Nankervis, J.C., Savin, N.E., Whiteman, C.H., 1992. Integration versustrend stationary in time series. Econometrica 60, 423433.
Dekimpe, M.G., Hanssens, D.M., 1995a. Empirical generalizations about marketevolution and stationarity. Market. Sci. 14, G109G121.
Dekimpe, M.G., Hanssens, D.M., 1995b. The persistence of marketing effects onsales. Market. Sci. 14, 121.
Dekimpe, M.G., Hanssens, D.M., 1999. Sustained spending and persistent response:a new look at long-term marketing protability. J. Market. Res. 36, 397412.
Dickey, D.A., Fuller, W.A., 1979. Distribution of the estimators for autoregressivetime series with a unit root. Am. Stat. Assoc. 74, 427 431.
Diebold, F.X., Rudebusch, G.D., 1991. On the power of DickeyFuller tests againstfractional alternatives. Econ. Lett. 35, 155160.
Graves, S.C., 1999. A single-item inventory model for a nonstationary demandprocess. Manuf. Serv. Oper. Manag. 1, 5061.
Guler, M.G., 2014. A note on the effect of optimal advertising on the distribution-free newsboy problem. Int. J. Prod. Econ. 148, 9092.
Hanssens, D.M., 1998. Order forecasts, retail sales, and the marketing mix forconsumer durables. J. Forecast. 17, 327346.
Hoeting, J.A., Madigan, D., Raftery, A.E., Volinsky, C.T., 1999. Bayesian modelaveraging: a tutorial. Stat. Sci. 14, 382401.
KantarMedia, 2011. Kantar Media Reports U.S. Advertising Expenditures Increased3.2% in the First Half of 2011. Retrieved Feburary 1, 2012. http://www.kantarmedia.com/sites/default/ les/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdf.
Khouja, M., Robbins, S.S., 2003. Linking advertising and quantity decisions in single-period inventory model. Int. J. Prod. Econ. 86, 93105.
Leamer, E.E., 1978. Specication Searches. Wiley, New York.Lee, C.M., Hsu, S.L., 2011. The effect of advertising on the distribution-free newsboy
problem. Int. J. Prod. Econ. 129, 217224.Lee, Y.S., 2014. A semi-parametric approach for estimating critical fractiles under
autocorrelated demand. Eur. J. Oper. Res. 234, 163173.Levi, R., Roundy, R.O., Shmoys, D.B., 2007. Provably near-optimal sampling-based
policies for stochastic inventory control models. Math. Oper. Res. 32, 821839.Lovejoy, W.S., 1990. Myopic policies for some inventory models with uncertain
demand distributions. Manag. Sci. 36, 724738.Ma, P., Wang, H., Shang, J., 2013. Supply chain channel strategies with quality and
marketing effort-dependent demand. Int. J. Prod. Econ. 144, 572581.Madigan, D., Raftery, A.E., 1994. Model selection and accounting for model
uncertainty in graphical models using occam's window. J. Am. Stat. Assoc. 89,
15351546.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289288
http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref1http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref1http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref1http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref1http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref1http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref2http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref2http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref2http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref2http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref2http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref3http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref3http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref3http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref3http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref3http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref4http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref4http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref4http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref8http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref8http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref8http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref8http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref8http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref9http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref9http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref9http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref9http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref9http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref10http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref10http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref10http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref10http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref10http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref12http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref12http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref12http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref12http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref12http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref14http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref14http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref14http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref14http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref14http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref16http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref16http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref16http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref16http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref16http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref17http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref17http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref17http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref17http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref17http://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://refhub.elsevier.com/S0925-5273(14)00274-6/sbref19http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref19http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref19http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref19http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref19http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref20http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref20http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref20http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref20http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref21http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref21http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref21http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref21http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref21http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref22http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref22http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref22http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref22http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref22http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref23http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref23http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref23http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref23http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref23http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref24http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref24http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref24http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref24http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref24http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref25http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref25http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref25http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref25http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref25http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref26http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref25http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref25http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref24http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref24http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref23http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref23http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref22http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref22http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref21http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref21http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref20http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref19http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref19http://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://www.kantarmedia.com/sites/default/files/press/Kantar_Media_Q2_2011_US_Ad_Spend.pdfhttp://refhub.elsevier.com/S0925-5273(14)00274-6/sbref17http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref17http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref16http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref16http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref15http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref14http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref14http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref13http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref12http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref12http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref11http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref10http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref10http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref9http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref9http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref8http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref8http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref7http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref6http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref5http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref4http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref4http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref3http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref3http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref2http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref2http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref1http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref1 -
8/10/2019 Bivariate, First-Order Vector -Autoregression Bayesian Model
12/12
Marques, A., Lacerda, D.P., Camargo, L.F.R., Teixeira, R., 2014. Exploring the relation-ship between marketing and operations: neural network analysis of marketingdecision impacts on delivery performance. Int. J. Prod. Econ. 153, 178190.
Nijs, V.R., Dekimpe, M.G., Steenkamp, J.E.M., Hanssens, D.M., 2001. The category-demand effects of price promotions. Market. Sci. 20, 122.
O'Hagan, A., Forster, J., 2004. Kendall's Advanced Theory of Statistics: BayesianInference, vol. 2B. Arnold, London.
Pauwels, K., Hanssens, D.M., Siddarth, S., 2002. The long-term effects of pricepromotions on category incidence, brand choice, and purchase quantity.
J. Market. Res. 39, 421439.Porteus, E.L., 2002. Foundations of Stochastic Inventory Theory. Stanford Business
Books, Stanford, CA.Rabin, M., 2002. Inference by believers in the law of small numbers. Quart. J. Econ.
117, 775816.Raftery, A.E., Madigan, D., Hoeting, J.A., 1997. Bayesian model averaging for linear
regression models. J. Am. Stat. Assoc. 92, 179191.
Raftery, A.E., Zheng, Y.Y., 2003. Discussion: performance of Bayesian model
averaging. J. Am. Stat. Assoc. 98, 931938.Retailgazette, 2011. Promotions Now Represent 39% of Grocery Sales. Retrieved
on March 1, 2012. http://www.retailgazette.co.uk/articles/23231-promotions-
now-represent-39-of-grocery-sales .Rossi, P.E., Allenby, G.M., McCulloch, R., 2005. Bayesian Statistics and Marketing,
Wiley Series in Probability and Statistics, Wiley, Hoboken, NJ.Shaman, P., Stine, R.A., 1988. The bias of autoregressive coefcient estimators. J. Am.
Stat. Assoc. 83, 842848.Srinivasan, S., Pauwels, K., Hanssens, D.M., Dekimpe, M.G., 2004. Do promotions
bene
t manufacturers, retailers, or both? Manag. Sci. 50, 617
629.Tang, C.S., 2010. A review of marketing-operations interface models: from co-
existence to coordination and collaboration. Int. J. Prod. Econ. 125, 2240.Urban, T.L., 2005. A periodic-review model with serially-correlated, inventory-
level-dependent demand. Int. J. Prod. Econ. 95, 287295.
Y.S. Lee / Int. J. Production Economics 158 (2014) 278289 289
http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref28http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref28http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref28http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref28http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref28http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref29http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref29http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref29http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref31http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref31http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref31http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref32http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref32http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref32http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref32http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref32http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref33http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref33http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref33http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref33http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref33http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref34http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref34http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref34http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref34http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref34http://www.retailgazette.co.uk/articles/23231-promotions-now-represent-39-of-grocery-saleshttp://www.retailgazette.co.uk/articles/23231-promotions-now-represent-39-of-grocery-saleshttp://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref39http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref39http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref39http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref39http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref39http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref40http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref40http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref40http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref40http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref40http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref40http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref40http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref39http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref39http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref38http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref37http://www.retailgazette.co.uk/articles/23231-promotions-now-represent-39-of-grocery-saleshttp://www.retailgazette.co.uk/articles/23231-promotions-now-represent-39-of-grocery-saleshttp://refhub.elsevier.com/S0925-5273(14)00274-6/sbref34http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref34http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref33http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref33http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref32http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref32http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref31http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref31http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref30http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref29http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref29http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref28http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref28http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27http://refhub.elsevier.com/S0925-5273(14)00274-6/sbref27