ARTICLE IN PRESS

Int. J. Production Economics 108 (2007) 74–83

www.elsevier.com/locate/ijpe

A model for selecting the appropriate level of aggregation

in forecasting processes
Giulio Zotteria, Matteo Kalchschmidtb,
a
Dipartimento di Sistemi di Produzione ed Economia Aziendale, Politecnico di Torino, C.so Duca degli Abruzzi 24, Torino, Italy
b
Dipartimento di Ingegneria Gestionale, Università degli Studi di Bergamo, Via Marconi 5, 24044 Dalmine (BG), Italy
Available online 16 December 2006

Abstract

Demand forecasting is a major issue in several industrial sectors. A relevant choice for companies is the proper level of
forecast aggregation. Forecasters need to properly identify what is the object of the forecasting process, in terms of time
bucket (e.g., forecasts are produced on a daily level or on weekly one), set of items the demand refers to (e.g., single item or
group of items), set of locations the demand refers to (e.g., single store or chain of stores). Managers can follow two basic
approaches: on the one hand they can adopt a detailed forecasting approach, i.e., they can forecast demand for the item at
the store by simply looking at the demand for the specific item/store; on the other hand they can adopt an aggregated
forecasting approach.
In this paper, we aim at figuring out what is the balance between the strengths and weaknesses of these two options, and
to identify the contingent variables that might lead managers to adopt one approach rather than the other. In this paper we
study the aggregation across locations by evaluating the components of forecasting error under the assumption of
stationary demand.
Finally, we suggest metrics that one can adopt to support the choice of the appropriate forecasting process, thus
providing help to managers in defining the proper level of aggregation for a specific situation.
r 2007 Elsevier B.V. All rights reserved.

Keywords: Demand forecasting; Aggregation level; Quantitative model

1. Introduction process is defined. In particular, we shall define

demand over three dimensions:
A forecast of final demand is one of the key
inputs for all planning activities as most decisions (a) one shall define the market he/she tries to
aim to make supply meet demand. The forecasting forecast; e.g., one retailer might want to forecast
problem is properly set once the output of such a demand at the single store level, while a
manufacturer might be interested in the demand
for the overall region or country; clearly the
Corresponding author. Tel.: +39 35 205 2360; former forecasting problem is harder to tackle
fax: +39 035 562 779.
than the latter;
E-mail address: [email protected] (b) one shall define the product the demand refers
(M. Kalchschmidt). to; e.g., for a given retailer it might be fairly

0925-5273/$ - see front matter r 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2006.12.030
 ARTICLE IN PRESS
G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83 75

difficult to predict the demand for a given Often practitioners and academicians are led to
product at the style-colour-size-packaging level, believe that an aggregate output (e.g. forecasting
whereas forecasting the total turnover for a total yearly turnover) requires a correspondingly
given product category might not be that hard aggregate process that looks at total yearly sales
(Wacker and Lummus, 2002); over time and vice versa. Empirical evidence shows
(c) finally one needs to define the time frame of the that this is not the case as many companies choose a
forecasting problem, i.e., one shall define the level of aggregation of the forecasting process that
time bucket and the forecasting horizon; indeed differs from the level of aggregation of the
forecasting demand at the day level is much forecasting problem.
more complex than forecasting total yearly In the consumer goods sector, a couple of large
demand; also forecasting what is going to manufacturers in Italy need to plan the inventory
happen tomorrow is simpler than forecasting levels at their warehouses. Each warehouse serves a
what is going to happen in 60 days. set of customers that are large retail chains. To
create a demand forecast for the each sku carried in
the warehouse they look at the demand for each
In the remainder of the paper we will refer to single customer to better understand the effects of
these three dimensions as the level of aggregation of trade promotions as they are retailer-specific. They
the forecasting problem. The smaller the market, the forecast demand for each retail chain-sku combina-
more detailed the definition of the product and the tion and then aggregate them at the warehouse-sku
smaller the time bucket, the more the forecasting level. In other words in the case of these manu-
problem is detailed. facturers the forecasting process is more detailed
One might think that the definition of these than the forecasting problem on the market dimen-
parameters is up to the forecasting manager; on the sion (see Caniato et al., 2002). A major European
contrary, they really depend on the decision-making retailer needs to plan inventory levels at the store-
process forecasting is supporting. One shall set the sku-day level and faces the challenge of managing
level of aggregation of the forecasting problem promotions. This retailer looks at the aggregate
according to the level of aggregation of the decision demand to understand the ‘‘lift’’ factors that a
making one. If the decision-making problem is a promotion creates and then uses such lift factors at
fairly aggregate one, the forecasting is going to be the single store level. In this case, the forecasting
aggregate as well; whereas a detailed decision- process is more aggregate than the forecasting
making problem requires a detailed forecast. E.g., problem on the market dimension (see Zotteri
when one is budgeting total production costs for et al., 2005)1. Finally, in the now classic Sport
next year an aggregate yearly demand (for all Obermeyer Case (Hammond and Raman, 1994),
products and all markets) is enough; on the Wally Obermeyer needs to plan the production of
contrary, when planning inventories, demand fore- parkas at the style-colour-size-level. However, the
cast needs to be very detailed, probably down to the early demand from the Las Vegas fair is analysed at
single store, single Stock Keeping Unit (SKU), and the style–colour level. Only after style–colour
probably for a single week. Also when one predictions are generated they are broken down at
replenishes stores weekly the demand forecast shall the size level. In the case of Obermeyer the
be at the least weekly (if not daily) and a monthly forecasting process is more aggregate than the
demand rate is simply not enough to drive the forecasting problem on the product dimension.
replenishment process. To show that the choice of the level of aggrega-
Thus, there is a fairly tight relationship between tion is a relevant issue we resort to a simple example
the level of aggregation of decision making and the described in Table 1. The company has two stores,
level of aggregation of the forecasting problem. This A and B, and one product. We assume the company
leads some forecasting managers to believe that they
1
have very little latitude over the choice of the proper Literature refers to a forecasting process that is more detailed
level of aggregation of the forecasting process. than the forecasting problem as bottom-up: initially forecasts are
generated at a detailed level and then they are added up (i.e.,
While forecasting managers have little to say on
aggregated). On the contrary, literature refers to situations where
the level of aggregation of the forecasting problem, the forecasting problem is more detailed than then forecasting
they can choose the level of aggregation of the process as top-down: first aggregate forecasts are generated and
forecasting process. then they are broken down at a more detailed level.
 ARTICLE IN PRESS
76 G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83

Table 1 aggregating correlated time series can be helpful to

Demand data: differences between aggregate and detailed better estimate seasonality since it can reduce casual
forecasting processes variability. Other works focus on the selection of the
Store Period 1 Period 2 Period 3 Period 4 Period 5 proper level of data aggregation (e.g., Chan, 1993;
Gonzales, 1992; Weiss, 1984). Some authors argue
A 1 0 1 0 1 that the top-down approach can be helpful as it is
B 0 2 2 4 4
more efficient and more accurate in times of stable
Chain 1 2 3 4 5
demand (Theil, 1954; Grunfeld and Griliches, 1960;
Lapide, 1998). Other authors reply that the bottom-
has decided to adopt linear regression to estimate up approach is needed when there are differences
demand and needs to forecast demand for the across time series (Orcutt et al., 1968; Zellner and
product both at the chain level to plan purchases Tobias, 2000; Weatherford et al., 2001).
and at the store level to plan distribution. Finally, a third group of papers (Miller et al.,
In a detailed process we first forecast demand at 1976; Barnea and Lakonishok, 1980; Fliedner,
the store level and then aggregate forecasts at the 1999) seems to take a more contingent approach
chain level. Demand forecast for the next period and shows that the choice between the aggregate
would be 0.5 units for store A and 5.4 units for store and detailed approach depends on correlation
B, adding up to 5.9 units for the chain (this is often among time series.
called the bottom-up approach). If we adopt an Our contribution belongs to this third cluster and
aggregate process we first create a forecast for the tries to highlight demand contingencies that
chain (6 units) and then break it down at the store (should) drive the decision of managers. However,
level. If we break down the forecast according to this paper differs from previous contributions in
the sales rate, the forecast for store A is 1.2 and two ways:
the forecast for store B is 4.8 (this is often called the
top-down approach).  Previous literature is based on data analysis
Thus, not only companies choose different levels rather than modelling; this makes the compar-
of aggregation of the forecasting process, but this ison of papers and findings hard as different
choice matters as, with a given algorithm, the results can be due to different data as well as
outcome may change significantly (Fildes and differences in algorithms or aggregation pro-
Beard, 1992). cesses. These analyses provided the community
Clearly the decisions of a company depend on with empirical evidence of the relevance of the
many factors including the degree of centralization issue but we still lack a model that defines
of the organization, the availability of detailed data, relevant variables and provides a general frame-
algorithms adopted (Mentzer and Cox, 1984; work.
Mentzer and Kahn, 1995), availability and flex-  Previous papers focus on aggregation across
ibility of human resources (Wacker and Lummus, items, whereas this paper provides a model on
2002) and features of the software used to forecast the aggregation across stores, i.e., geographical
demand. While we acknowledge that these issues are locations. However, our model can be applied to
relevant, in our paper we will only investigate how the case of different items as well, as it can be
the nature of demand influences the choice of the used to investigate the aggregation over the
appropriate level of aggregation. market dimension.

2. Literature review 3. The model

Forecasting has received significant attention 3.1. Assumptions

from academicians but literature on aggregation is
relatively sparse. Several contributions on this issue Our model investigates a single item i, so we
focus on the use of aggregation to estimate assume that there is no relevant relationship (e.g.,
seasonality curves (Dalhart, 1974; Withycombe, no correlation) among products.
1989; Bunn and Vassilopoulos, 1993, 1999; Dekker Product i is sold through a chain of J stores and
et al., 2004). These works provided evidence that the demand process is stationary. Also, for the sake
 ARTICLE IN PRESS
G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83 77

of simplicity we assume there is no correlation differences in selling rates among stores so we

among stores and over time, note that this assump- assume stores to be a priori equal.2
tion could be easily relaxed at the expense of clarity In our model we assume mij to be unknown
and simplicity. parameters we want to estimate. So we will compare
The demand xijt of item i at store j at time t the alternatives we are going to investigate in terms
follows a generic distribution d(mij,sij), where: of ability to properly estimate mij. Indeed, in a
stationary process the expected demand is the
 mij is the expected level of demand for item i at forecast that minimizes the (expected squared)
store j in any time bucket; error. Thus an improved estimate of the expected
 sij is the standard deviation of demand for item i demand leads to a better forecast accuracy on
at store j and it is proportional to the mean average. Given the i.i.d. assumption it is easy to
demand sij ¼ CVmij, where CV is the coefficient show that
of variation of demand that measures variability.
E½ðxijt  x^ ijt Þ2  ¼ E½ðxijt  mij Þ2  þ E½ðmij  x^ it Þ2 
We assume to have observed demand at all J ¼ s2ij þ E½ðmij  x^ it Þ2 , ð1Þ
stores over T periods of time. T is the number of
where x^ ijt is the forecast of demand of item i at store
periods during which we could observe a stationary
j at time t. Thus the total forecasting error, in this
demand process, thus it can be limited for several
context, is equal to the variability of the process plus
reasons:
the estimation error of the parameter. The first
portion of the equation is an exogenous variable for
1. the company/product can be fairly new and thus
the forecasting managers (though it can be to some
the amount of observations on past demand
extent endogenous for the company at large). On
could be limited;
the contrary, forecasting managers can somehow
2. the company might not store the data for a long
improve the second term through better estimates of
period of time and thus a reduced set of
the expected demand. In the remainder of the paper
observations might be available, even if ICTs
we will focus on this term (error of estimate of the
are making data storage cheaper and cheaper;
expected demand) as it is the only portion of the
3. the demand process could experience dramatic
forecasting error that, given our assumptions, can
changes that make (a maybe long) post history
actually be improved through better forecasting
irrelevant. E.g., consider a product that has
techniques.
recently been set on promotion and thus the off
promotion history is somehow irrelevant (and
vice versa) or a store that has re-opened after 3.2. Alternatives
substantial restructuring. Actually, the idea
behind a very traditional forecasting technique We assume the forecasting problem to be
such as moving average is that only the last T detailed, i.e., we assume that managers need to
observations are stationary and thus can be used forecast demand at the sku-store level. E.g.,
to forecast future demand. managers might plan inventories for each single
store in the chain; thus they might need to forecast
Also we assume that the managers have no priors demand at the item/store level. To do so, they can
on the parameters of the distribution that will be follow two basic approaches:
estimated only by looking at past observations. We
consider the forecasting managers to have no  on the one hand, they can adopt a very detailed
information on future demand draws (e.g., no forecasting approach, i.e., they can forecast future
orders are collected before time t) other than past demand for the item at a store by simply looking
observations. at the past demand at the specific store;
In the end, we assume that item i does not sell
2
evenly at all J locations and thus mij is distributed Also one might take store size into account by separating the
estimation of store traffic from the estimation of closing rate
according to f(mj; h  mj) where h is the degree of
(percentage of customers buying the item out of the customers
heterogeneity of stores, i.e., the differences in the that entered the store); when estimating closing rates even very
success of the product among the stores. We assume different stores might have a very similar prior on selling rates of
that the company does not have any prior on the a given product.
 ARTICLE IN PRESS
78 G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83

 on the other hand, they can adopt a more Thus, the standard error of estimate of the
aggregated forecasting approach. One might detailed process (seed) is
assume that the demand rate for the item is
CV mij
constant across all locations (maybe adjusting for seedmij ¼ pffiffiffiffi .
the size of the stores in such a way that one only T
needs to assume that the customers’ tastes are If the aggregate process is followed then the
constant across locations) and thus look at the demand rate estimate is m ~ ij
overall demand for the chain to get a more PJ PT  
reliable estimate. j¼1 t¼1 xijt sij
~ ij ¼
m g mi ; pffiffiffipffiffiffiffi
TJ J T
pffiffiffiffiffiffiffiffiffiffiffiffiffi!
As literature suggests, these two approaches have CV mi 1 þ h2
contrasting pros and cons. The first approach can ¼ g mi ; pffiffiffipffiffiffiffi .
J T
very well capture the uniqueness and specificity of
the demand at a given store; the flipside however is In this case it can be shown that standard error of
the inability to enjoy large samples to draw estimate of the aggregate process (seea) is (see
statistically significant conclusions (sample size is Appendix A for all relevant proofs)
T). The model faces a very small specification error sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
(i.e., will be very adequate conceptually) but faces a CV2 ð1 þ h2 Þm2i
substantial sampling error as the number of para- seeamij ¼ þ ðmij  mi Þ2 .
JT
meters is very large (J). The second approach, on
the other hand, can be considered to be fairly crude, The above formulas3 highlight the basic trade off
since often different stores have different customers we are facing: in the detailed approach we have only
that tend to like different products and create the sampling error that might be relatively large, as
different demand patterns. However, the second data are limited. On the contrary, the aggregated
approach has the advantage of enjoying a larger approach will enjoy a relatively smaller sampling
sample (sample size is JT). So conclusions might be error (on average J times smaller because we sum
conceptually crude as the specification error will be the demand from all the J stores); however, in the
substantial, but the sampling error will be limited. aggregated approach we also face a specification
Under our assumptions, managers need to set error ((mijmi)2), i.e., the model we are building is
inventory levels for each single store and thus need an oversimplification of reality and assumes that the
to forecast demand at this fairly detailed level. Hence, demand is constant across all stores. Indeed, this
we compare the ability of the aggregated and detailed error does not vanish as we increase the number of
forecasting process to create an accurate forecast at observations T.
this level of detail. Also notice that under our Clearly these are very detailed errors at the item/
assumptions both alternatives provide the same store level and one approach might work better for
aggregate forecast, thus in this specific context some stores while the second might work better for
comparing the two approaches in terms of their ability others. Actually when choosing the aggregate rather
to forecast aggregate demand does not make sense. than the detailed approach one should not look at
Given the assumptions we have made, the average whether one method performs better at a specific
of the past T observations is the best estimator store but rather at how accurate the two alternatives
available (unbiased minimum variance estimator— are on the average.
UMVE) of the average demand (i.e., the estimator We suggest that one might be willing to minimize
that minimizes the right-hand size in Eq. (1)). Thus, the sum of all errors across all stores. Indeed the
we deploy this estimator both in the aggregate and sum of all errors is likely to correlate with costs due
in the detailed processes to check which process to as safety stock and stock outs. In particular, we
performs best. suggest to measure the sum of all quadratic errors
If the detailed forecasting process is used then the that is consistent with the adoption of standard
estimate of the average demand for item i at store j 3
Note that this is the UMVE of the average (over the store
is m ^ ij
locations) expected demand mi; so other estimators would
PT     increase the variance ðCV2 ð1 þ h2 Þm2i Þ=J T of the estimate or
xijt sij CV  mij
m^ ij ¼ t¼1 g mij ; pffiffiffiffi ¼ g mij ; pffiffiffiffiffi . the average (over all locations J) of the difference between the
T T T expected demand and the expected value of the estimator .
 ARTICLE IN PRESS
G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83 79

error of estimate at the single store level (see In the case of heterogeneity, the detailed model
Syntetos et al., 2005, for a similar approach in a shall be used if
different forecasting problem). sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Thus, one might compare the total error for the CV2 J  1 1
if o1 and h4 2  1.
detailed approach (ted) T J 1 CV J1
T J

X
J
CV2 In other words, the solution suggests to adopt the
tedi ¼ seed2mij ¼ Jð1 þ h2 Þm2i aggregate model only in case of homogeneous
j¼1
T
chains, as in other cases the error one makes by
with the error of the aggregate one (tea) (see providing the same estimate for all stores is too
Appendix A for proofs) large. Also we shall notice that the detailed
alternative is not even an option in case of variable
X
J
CV2 ð1 þ h2 Þ 2 demand, small samples and large chains.
teai ¼ seea2mij ¼ mi þ Jðh mi Þ2 . Eq. (2) also suggests to select the detailed model
j¼1
T
in case the product has been around (and/or
Finally, the approach with the smaller total demand has been stable) for a relatively long period
(expected) error is selected. Thus, the detailed of time
approach is going to be selected if
J  1 h2 þ 1
T4CV2 .
CV2 CV2 ð1 þ h2 Þ J h2
Jð1 þ h2 Þo þ J h2 . (2)
T T This adds a product life cycle flavour to the
model. Indeed, for a given product we suggest to use
an aggregate demand model early on in the product
4. Analysis of results lifecycle (or just after a sudden change in demand
such as the start of a promotion) and then use a
The above formulas can be used to investigate the detailed process once enough evidence to support
performance of the two alternative approaches for store-specific forecasts is gathered. This result is
various combinations of the contingent variables consistent with results provided by previous works
CV, h, J, T. Also, Eq. (2) enables to identify (e.g., Roberts, 1998). Also this finding is very
threshold values for the above variables. consistent with the practise of a European grocery
Eq. (2) suggests that the detailed approach shall retailer that early on in the promotion looks at the
be chosen in cases of low demand variability, since general demand trend but, as more sales data are
Eq. (2) is verified when collected, looks at the selling rate of each single
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi store independently. Indeed, early in the product (or
J h2 promotion) life cycle, what matters the most is to
CVo T.
J  1 h2 þ 1 gain a large enough sample to reduce the sampling
Indeed, variability makes the estimation of error. Thus, the aggregate process can work better
expected demand for each store hard; a high degree than the detailed one. Later on, the sampling error
of variability forces the company to aggregate data is reduced: while this leads the error of the detailed
at the chain level to gain a large enough sample. process to zero, the aggregate process is still left
In the case of the chain size J, Eq. (2) suggests to with the specification error that does not disappear
adopt the detailed solution if with larger samples, as shown in Fig. 1.
Finally, we shall notice that in all our derivations
CV2 1 þ h2 of the thresholds on the four contingent variables
if p1 8J,
T h2 we introduced three functions:
CV2 1 þ h2 1
if 1 þ h2 41; Jo1 þ .  (CV2/T) ¼ v measures the extent to which the
T h2 CV2 1þh2
1
T h2 sample size enables to provide accurate estimates,
In other words, we suggest to use the detailed given the variability of demand; in other words
process where the chains are relatively small as the this function captures the sampling error;
advantage one gains by aggregating at the chain  (J1/J) ¼ J1, 0.5pJ1o1 (JX2) captures the
level is limited. effect of the chain size, and in particular this
 ARTICLE IN PRESS
80 G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83

14
12
10
8
error

detailed
6 J −1 . h2 + 1
T = CV 2 . aggregate
4 J h2

2 specification
error
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37
time

Fig. 1. Errors as a function of the number of periods T observed (case of J ¼ 10, CV ¼ 1, h ¼ 0.5).

measures the advantage that the aggregate A simple example will make the issue more clear.
process has over the detailed one as it can Let us assume that all the demands at the J stores
increase the sample size and thus reduce the are equal but the demand is rather variable and the
sampling error; amount of information available is rather limited
 h2/(h2+1) ¼ h1, 0ph1o1 is the factor that (i.e., large CV and small T). In this case m ^ ij are
captures the effect of heterogeneity on the likely to be fairly different and thus one could be led
specification error of the aggregate approach. to believe that the demand is heterogeneous. On the
contrary, we shall tell the effect of variability from
Indeed, all above equations can be re-written as a the effect of heterogeneity by looking at the variance
function of v, J1 and h1. across m ^ ij , to estimate whether the parameters
(rather than their estimates) are actually hetero-
5. Estimation of parameters geneous.
In other words, one might be tempted to estimate
The model presented in this paper provides h as
insights into the variables that underlie the choice
of the level of aggregation of the forecasting P
J
^ i Þ2
^ ij  m
ðm
process. However, when one wants to move from 2 j¼1
the interpretative and modelling stage to the h^ ¼ .
^ 2i
Jm
prescriptive one, he/she needs to estimate the
parameters of the model. However, in Appendix B we show that
While the estimation of parameter J is self !
X
J  
explanatory, and for T we refer to the third section 2 ð1 þ h2 ÞðJ  1ÞCV2
E ^ ij  m
ðm ^ iÞ ¼ m2i þ J h2
of this paper, other variables deserve further j¼1
T
discussion. In particular, the estimate of h and CV
(3)
is somehow tricky. Clearly when one wants to
estimate the degree of difference across the stores and thus what seems to be an intuitive estimate of h
he/she needs to look at demand at the store/item is actually biased because of variability.
level, i.e., estimate different values of mij with a Thus, Eq. (3) can be used to estimate h, once an
detailed process. However, in case different values estimate of CV is derived.
m^ ij are observed, one shall wonder whether such It is fairly easy to estimate CV by looking at the
differences are due to actual differences in mij across demand for a given product at a given location over
stores (and thus shall be interpreted as heterogene- time as follows:
ity) or might be due to the variability of demand. In vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uT
the latter case, differences in mij across stores might uP
J u
u ðxijt  m ^ ij Þ2
simply be due to sampling errors as the estimates 1 X
^ ¼
CV t t¼1
.
might be different while the parameters are equal
J 1 ðT  1Þm ^ 2i
(or have little differences).
 ARTICLE IN PRESS
G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83 81

Thus, once the estimation of the variability is error is not reduced by larger samples (i.e.,
derived from time series we can tell the contribution aggregation).
of variability from the contribution of heterogeneity Moreover, attention has been paid here on
in the variance of the estimates of mean selling rates aggregation on the market dimension (e.g., forecast
at stores (see Eq. (3)). at store or chain level). It would be important to
add product dimension, so more options are
available and we can aggregate over products and/
6. Conclusions or stores.
Also the current model assumes no priors on
This work evaluates the trade-off between parameters and namely heterogeneity and thus the
specification and sampling error when choosing expected level of demand at different stores.
the level of aggregation of the forecasting pro- Actually this seems to be a fairly strong assumption
cess. This paper identifies the contingent vari- as even before a product is launched the managers
ables that might lead managers to adopt a detailed might have some ideas on both the absolute level of
rather than an aggregate approach. This is ac- demand (mi) and its distribution across stores.
hieved by evaluating the components of forecasting Indeed, store size or traffic could be a significant
error under the assumption of non-correlated (over (and only partially unknown) drivers of demand at
time and across products or markets), stationary the store level (mij).
demand and by considering aggregation only on Finally, it would be interesting to model cluster-
locations. ing of stores (or items) as an intermediate alter-
The contribution may be significant in real native between a very detailed and a very aggregate
applications, since it may help companies to under- model (see Zotteri et al., 2005; Caniato et al., 2005).
stand what the best approach to forecast future This would help companies to identify the level of
demand is. Thus we also suggest metrics that can be aggregation that better solves the trade-off between
adopted to support the choice of the appropriate specification and sampling error. This would
forecasting process. However the current model is probably entail the modelling of both intra-cluster
limited from several perspectives and thus future and inter-cluster heterogeneity.
development is needed to properly investigate this Also the current research requires some addi-
issue. tional efforts on the empirical side to test the current
A first set of directions for further developments (and eventually future) theory.
regards the simplistic assumptions of current model, The robustness of the current model shall be
so additional efforts should be devoted to building tested by checking the sensitivity to both the fairly
the theory. strong assumption made and errors in estimate of
This work assumes demand to be stationary, parameters. This can be achieved by performing
however in many different contexts this is not the sensitivity analyses through Monte Carlo simula-
case (e.g., spare parts, fashion goods, and so on). tion. Also the actual ability to select the appropriate
Thus it would be useful to model the case of non- aggregation of the forecasting process should be
stationary demand. A first step could be to assume tested with real data that are currently being
linear trend and use linear regression to generate collected.
future forecasts. Also such a change might lead to
differences in aggregate forecasts between the
detailed and the aggregate forecasting process (see Acknowledgments
Zotteri et al., 2005); thus one might have to
compare the performance of various alternatives Authors have contributed jointly to the present
both at the chain and store level. work; however Giulio Zotteri has edited the
A second interesting issue is to explore the following paragraphs: ‘‘The model’’, ‘‘Analysis of
case of demand correlated over time and/or results’’, ‘‘Estimation of parameters’’ and ‘‘Conclu-
products and/or across locations; nevertheless we sions’’. Matteo Kalchschmidt has written the
argue that we would probably get similar results. remaining paragraphs ‘‘Introduction’’ and ‘‘Litera-
Indeed, correlation across stores simply reduces ture review’’. Also we wish to thank three anon-
the benefits of aggregation as the behaviour of ymous referees for their helpful guidance and
stores over time is similar and thus the sampling insightful comments and suggestions.
 ARTICLE IN PRESS
82 G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83

Appendix A. Proofs X
J
tedi ¼ seed2ij
j¼1
Given assumptions we know that
J CV2 m2
X
PJ ij CV2 X
J

j¼1 mij
¼ ¼ m2
¼ mi , j¼1
T T j¼1 ij
J
CV2
PJ ¼ Jð1 þ h2 Þm2i .
 mi Þ2 T
j¼1 ðmij
¼ h2 m2i , And can compute the total error one makes by
J
using the aggregate approach as
hence we can derive that
X
J
PJ 2 PJ 2 PJ 2 teai ¼ seea2mj
j¼1 ðmij  mi Þ j¼1 ðmij Þ  j¼1 ðmi Þ
¼ j¼1
J J J  
PJ 2 2 X CV2 m2 ð1 þ h2 Þi 2
j¼1 ðmij Þ  Jðmi Þ ¼ þ ðmi  mij Þ
¼ ¼ h2 m2i j¼1
TJ
J
J 
X  X
J
and thus CV2 m2 ð1 þ h2 Þ
¼ i
þ ðmi  mij Þ2
j¼1
TJ j¼1
X
J
2
m2ij ¼J m2i ð1 þ h Þ. 2
CV ð1 þ h Þ 2 2
j¼1 ¼ mi þ Jðh mi Þ2 ,
T
Also we can derive the properties of the estimator X
J J CV2 m2
X ij
~ ij
m tedi ¼ seed2mij ¼
T
PJ PT j¼1 j¼1
t¼1 Eðxijt Þ
CV2 X
j¼1 J
Eðm ~ ij Þ ¼ ¼ mi , CV2 2 2
TJ ¼ m2ij ¼ mi ðh þ 1Þ.
T j¼1 T

s2m~ ij ¼ E½ðm~ ij  mi Þ2 
PJ PT 2 PJ PT 2 Appendix B. Estimation proofs
j¼1 t¼1 sxijt j¼1 t¼1 ðCV mij Þ
¼ ¼
T2 J2 T2 J2
2 PJ PT P
CV j¼1 t¼1 mij CV2 T Jj¼1 m2ij
2 !
¼ ¼ X
J
^ i Þ2
^ ij  m
ðm
T2 J2 T2 J2 E
J
CV J m2i ð1 þ h Þ CV m2i ð1 þ h2 Þ
2 2 2 j¼1
¼ ¼ .
T J2 TJ 1X J
¼ Ef½ðm ^ i Þ2 g
^ ij  mij Þ þ ðmij  m
In the derivation process we first used the J j¼1
assumption on independence of demands across
1X J
stores and over time and then the distributive ¼ ^ ij  mij Þ2 þ Eðmij  m
fEðm ^ i Þ2
J j¼1
assumption on the variance of mij.
^ ij  mij Þðmij  m
þ 2E½ðm ^ i Þg,
~ ij  mij Þ2 
seeamij ¼ E½ðm
where by definition
~ ij  mi Þ þ ðmi  mij ÞÞ2 
¼ E½ððm
~ ij  mi Þ2  þ E½ðmi  mij Þ2 
¼ E½ðm CV2 m2ij
^ ij  mij Þ2 ¼
Eðm
þ 2E½ðm ~ ij  mi Þðmi  mij Þ T
and
CV2 m2i ð1 þ h2 Þ
¼ þ ðmi  mij Þ2 .
TJ ^ i Þ2 ¼ Eðmij  mi þ mi  m
Eðmij  m ^ i Þ2
We can compute the total error one makes by CV2 ð1 þ h2 Þm2i
¼ ðmij  mi Þ2 þ þ 0,
using the detailed approach as JT
 ARTICLE IN PRESS
G. Zotteri, M. Kalchschmidt / Int. J. Production Economics 108 (2007) 74–83 83

^ ij  mij Þðmij  m
E½ðm ^ i Þ Dalhart, G., 1974. Class seasonality—a new approach. In:
" P ! PJ PT !# Proceedings of 1974 Conference of American Production
T
t¼1 xijt j¼1 t¼1 xijt and Inventory Control Society, Reprinted in Forecasting, 2nd
¼E  mij mij 
T JT ed. APICS, Washington, DC, pp. 11–16.
" P ! P PT PT !# Dekker, M., Van Donselaar, K., Ouwehand, P., 2004. How to use
T
t¼1 xijt kaj t¼1 xijt xijt aggregation and combed forecasting to improve seasonal
¼E  mij mij   t¼1
T JT JT demand forecast. International Journal of Production Eco-
" P ! PT !# nomics 90, 151–167.
T
t¼1 xijt xijt Fildes, R., Beard, C., 1992. Forecasting systems for production
¼0þE  mij  t¼1
T JT and inventory control. International Journal of Operations
" P # "P PT # and Production Management 12 (5), 4–27.
T T
t¼1 xijt t¼1 xijt t¼1 xijt Fliedner, G., 1999. An investigation of aggregate variable time
¼ E mij E
JT T JT series forecast strategies with specific subaggregate time series
"PT P T P # statistical correlation. Computers & Operations Research 26,
m2ij 2
t¼1 xijt t¼1 xijt uat xiju
¼ E þ 1133–1149.
J J T2 J T2 Gonzales, P., 1992. Temporal aggregation and systemic sampling
Eðxijt Þ2 T Eðx2ijt Þ TðT  1ÞEðxijt Þ2 in structural time-series models. Journal of Forecasting 11,
¼   271–281.
J J T2 J T2 Grunfeld, Y., Griliches, Z., 1960. Is aggregation necessarily bad?
2
Eðxijt Þ 2 2
Eðxijt Þ sx CV2 m2ij Review of Economics and Statistics 42, 1–13.
¼  ¼  ijt ¼  .
JT JT JT JT Hammond, J., Raman, A., 1994. Sport Obermeyer Ltd., Case
number 9-695-022, HBS Publishing, Boston, MA.
Thus Lapide, L., 1998. New developments in business forecasting. The
" 2 Journal of Business Forecasting Summer, 28–29.
XJ CV m2ij
þ ðmij  mi Þ2 Mentzer, J.T., Cox Jr., J., 1984. Familiarity, application, and
j¼1
T performance of sales forecasting techniques. Journal of
# Forecasting 3, 27–36.
CV2 ð1 þ h2 Þm2i CV2 m2ij Mentzer, J.T., Kahn, K.B., 1995. Forecasting technique famil-
þ 2 iarity, satisfaction, usage, and application. Journal of Fore-
JT JT
casting 14, 465–476.
J CV2 ð1 þ h2 Þm2i Miller, J.G., Berry, W., Lai, C-Y.F., 1976. A comparison of
¼ þ J h2 m2i alternative forecasting strategies for multi-stage production
T inventory systems. Decision Sciences 7, 714–724.
CV2 ð1 þ h2 Þm2i CV2 ð1 þ h2 Þm2i Orcutt, G., Watts, H.W., Edwards, J.B., 1968. Data aggregation
þ 2 and information loss. American Economic Review 58,
T T
773–787.
ðJ  1ÞCV2 ð1 þ h2 Þm2i Roberts, J.H., 1998. Marketing approaches to forecasting
¼ þ J h2 m2i .
T problems. Journal of Forecasting 17, 169–174.
Syntetos, A.A., Boylan, J.E., Croston, J.D., 2005. On the
categorisation of demand patterns. Journal of the Operational
Research Society 56 (5), 495–503.
References Theil, H., 1954. Linear Aggregation of Economic Relations.
North-Holland, Amsterdam.
Barnea, A., Lakonishok, J., 1980. An analysis of the usefulness of Wacker, J.G., Lummus, R., 2002. Sales forecasting for strategic
disaggregated accounting data for forecasts of corporate resource planning: practical implications and research direc-
performance. Decision Sciences 11, 17–26. tions. International Journal of Production and Operations
Bunn, D.W., Vassilopoulos, A.I., 1993. Using group seasonal Management 22 (9), 1014–1031.
indices in multi-item short-term forecasting. International Weiss, A.A., 1984. Systematic sampling and temporal aggrega-
Journal of Forecasting 9, 517–526. tion in time-series models. Journal of Econometrics 12,
Bunn, D.W., Vassilopoulos, A.I., 1999. Comparison of seasonal 547–552.
estimation methods in multi-item short-term forecasting. Weatherford, L.R., Kimes, S.E., Scott, D.A., 2001. Forecasting
International Journal of Forecasting 15, 431–443. for hotel revenue management—testing aggregation against
Caniato, F., Kalchschmidt, M., Ronchi, S., Verganti, R., Zotteri, disaggregation. Cornell Hotel and Restaurant Administration
G., 2002. Forecasting demand fluctuations due to promo- Quarterly August, 53–64.
tional activities: a case in the fresh food industry. In: Withycombe, R., 1989. Forecasting with combined seasonal
Proceedings of POMS Conference, San Francisco. indices. International Journal of Forecasting 5, 547–552.
Caniato, F., Kalchschmidt, M., Ronchi, S., Verganti, R., Zotteri, Zellner, A., Tobias, J., 2000. A note on aggregation, disaggrega-
G., 2005. Clustering customers to forecast demand. Produc- tion and forecasting performance. Journal of Forecasting 19,
tion Planning and Control 16, 32–43. 457–469.
Chan, W., 1993. Disaggregation of annual time-series with Zotteri, G., Kalchschmidt, M., Caniato, F., 2005. The impact of
common seasonal patterns. Journal of Econometrics 55, aggregation level on forecasting performance. International
173–200. Journal of Production Economics 93–94, 479–491.