Académique Documents
Professionnel Documents
Culture Documents
summer 2004
system and a computer-aided ordering
Integrated Demand and Supply Chain system with simulation facilities are
Management available.
A key objective of management in retail
vol. 4, no. 1
is the optimisation of the supply chain • The profit lost by losing a customer
from the manufacturer to the store to completely if a customer cannot find his
fulfil the individual customer’s demand. or her favourite item in the store and
journal
Sales and Order Forecasts in CPFR
© ecr
55
another retailer. This effect can usually manufacturer such as number of SKUs in
not be measured directly and may only be a layer or on a pallet or the number of
approximated by market research or by pallets in a truck.
educated guesses. • Bracketed conditions where the
advantage of a lower price has to be
Typical elements of the cost for stores compared with the additional capital
from the supply side are : binding, stock-keeping and handling
• The ordering and capital binding cost per costs.
SKU per delivery period in the store.
• The delivery rhythm or more generally On the DC level, the non-availability of
the delivery plan from the manufacturer goods corresponds to the out of stock
or the DC to the store generating necessary situations in the store.
lot sizes for ordering.
• The minimal number of SKUs to be The manufacturer plays an important
ordered by the store from the DC or the role in decreasing the costs on the demand
manufacturer. and on the supply side. The manufacturer
summer 2004
Typical elements of the cost for a • lower the capital binding costs by
distribution centre from the supply side adapting to the retailer’s needs for cost-
are : optimal minimal ordering sizes,
journal
• The ordering and capital binding cost per packaging, delivery plans and changing
SKU per delivery period in the DC. the bracketed conditions.
• The delivery rhythm or more generally
© ecr
the delivery plan from the manufacturer These effects can only be achieved by
56 to the DC generating necessary lot sizes for collaboration of the retailer with the
ordering. manufacturer by sharing information
about the specification of items,
• The minimal number of SKUs to be packaging, delivery plans, discount
ordered by the DC or the manufacturer. systems but most importantly by sharing
• Logistic order restrictions on the number information about future sales and orders
Gerhard Arminger is
professor of statistics at
the University of
Wuppertal. He is co-
founder and chief scientist Solving the two central problems of CPFR - sales and
of SAF.
He can be contacted at
arminger@wwst09.wiwi.
order forecasting - requires attention to many issues,
uni-wuppertal.de
using many different measures.
which is typical for the CPFR process. In Integration of forecasts for different
summer 2004
chain management, the demands of the back.
individual customer at the individual Since calendar events (Easter, Fourth of
store should generate the orders by the July, Thanksgiving Day, Christmas),
store from the DC and by the DC from the advertising and promotional events may
vol. 4, no. 1
manufacturer. The customer’s demand have very different effects in the different
therefore constitutes the process of stores depending on the respective
demand chain management by forecasting locations, the forecasting techniques
the future sales to the customers at store should not be based only on the observed
journal
level and building adequate cost-optimal time series of sales but should be able to
orders. Since in this forecasting and order take this information about external
building the pull from the customers influences into account. Otherwise, only
© ecr
demand and the push from the trend and seasonal information can be
manufacturer and the retailer in form of extracted. Forecasts for sales and orders 57
promotions, advertising campaigns and for stores are usually based on short term
special displays have to be integrated, the forecasts, that is daily or between one to
whole process is called integrated demand six weeks depending on the delivery plans
and supply chain management (IDSCM of the DC or the manufacturer.
rather than SCM alone).
generated by the sales forecasts for the between the sales forecasts on the store
stores should influence the forecast for and/or DC level to the CPFR sales forecast
the total shipments from the DC to the is the time horizon. For the CPFR process,
stores and the orders from the DC to the a mid-range forecast of 10 to 13 weeks is
manufacturer. It should be noted that the advised to enable the manufacturer to
aggregated sales forecasts per item from adapt the production process to the
the stores should not be used as the specific predicted demand of the retailer.
forecasted number of SKUs to be shipped However, like the sales forecasts on the
to the stores since the sales forecasts will store level and the shipment forecasts on
not be identical with the cumulated the DC level, the CPFR sales forecast must
orders from the stores. A better way of take into account the basic demand of the
minimizing the total cost of the supply customer (pull) as well as calendar events
chain is to use either the shipments of an and the demand induced by promotional
item from the DC to the stores or the schemes such as different types of
cumulated orders of the stores from the advertisement, displays and price
past to the DC as dependent variable that reductions (push).
summer 2004
events are then automatically taken into forecasting mechanisms for the short-
account. Forecasts for shipments (or term replenishment in the store and the
cumulative store orders) and order DC may have to be replaced by more
journal
building for DCs are also usually based on structural models focusing on the
short term forecasts, that is daily or question how much extra demand will be
between one to 6 weeks depending on the induced by specific promotional schemes.
© ecr
SALES FORECASTS FOR STORES, DC AND The main problem of this approach is
summer 2004
THE CPFR PROCESS obviously that the forecast method never
approaches the store-SKU as unique.
Sales and inventory forecasts for the Therefore the needs of the individual
store level store may be highly over- or
vol. 4, no. 1
Presently, three distinct approaches to underestimated depending on how similar
forecasting the large number of SKUs and the needs of a specific store are to the
store combinations required for store aggregated demand in the store cluster.
ordering are used in semi-automatic or For all practical purposes, one can expect
journal
automatic computer assisted ordering from this forecasting method for the same
(CAO) systems: Aggregate, profile and item many stock-outs in one store and
micro-forecasting. more than necessary overstocking in
© ecr
another store. Also, in this approach it is
Aggregate forecasting practically impossible to take into account 59
Aggregate forecasting manages sparse local effects such as local calendar events,
inventory and reduces processing time by local advertisements and other local
grouping stores into store clusters and promotions into account. In effect,
creating a single forecast for each item computer assisted ordering in the store
across the entire store group. This single will be only semi-automatic rather than
• Create profiles from group aggregates. and varies by the unique store-SKU. The
• Create static weights for the application specific characteristics are:
of multiple profiles. • Each store item is forecasted uniquely.
• Apply various profiles to base forecast to • Each item history and external
produce actual forecast. influences are applied uniquely.
vol. 4, no. 1
• The results can only be as good as the The advantages of such a method are:
quality of the cluster and the clusters • There are no separate profiles or
will always include outlier stores. promotional lifts to maintain or to
© ecr
summer 2004
ordering (CAO) system. and price changes in the past.
• Second, the functional requirements for
the effectiveness and the speed of the Learning from each new bit of history
forecasting and optimisation engine in the system constantly improves its ability
vol. 4, no. 1
such a CAO system are much greater to predict future sales, rapidly recognize
than for aggregate or profile forecasting. product trends, and build dynamic
seasons. The retailer not only need not
However, most large retailers have now marry separate profiles and promotion
journal
either data warehouses and/or lifts but the self learning system
transaction systems that provide the continually and automatically analyses
required data base and automatic systems the impact of each sales influence. In this
© ecr
with the advanced features of forecasting way the influences of promotions,
and optimisation have been successfully seasons, and events are considered at one 61
installed by leading retailers in the store time. The compound predictors can
level. To work well, micro-forecasting include price changes, sales promotions,
systems must have two key attributes. event calendar, holiday information, and
They should be ‘self-learning’ and they special circumstances depending on the
should optimize inventory. availability of data provided by the client.
inventory sets the target shelf inventory traditional warehouse method for
level in the store. There is an ideal shelf calculating the ideal stock with a dynamic
set that covers both presentation stock safety stock. The service level is based on a
and the possibility of preventing lost sales symmetrical (bell curve) of inventory
without exceeding the costs of carrying necessary to support the risk of being out
surplus inventory. This optimal inventory of stock. The retailer sets a desired service
in the store sets the lower limit for the level which is the percentage of time that
ordering process. The ordering process they want to be in stock when the stock is
must order at least such a great number of needed based not only on projected
SKUs of one item that the optimal demand but also on risk factors such as
inventory for the first availability day of a variability in forecasted demand as well as
new delivery is equalled or exceeded. lead time variability.
Therefore, not only a forecast of the sales
per SKU in a store is needed but also a Cost-based optimal inventory. The use of cost
forecast of the optimal inventory. functions redefines the traditional
approach to service level and creates a
summer 2004
The retail industry has depended on shelf supply that covers volatility in
three primary methods: minimum stock forecast on a cost-optimum basis. It
with safety stock, minimum stock with minimizes the total cost of lost sales and
service level, and cost-based optimal excess inventory. By recognizing the
inventory. sometimes asymmetrical distribution of
vol. 4, no. 1
methods used by retailers to set their costs associated with lost sales which may
target shelf set (or order up to level). The create a significant loss of a customer’s
safety stock is either a fixed numerical shopping basket on the one hand and
© ecr
summer 2004
stores. The disadvantage of this approach the stores will be an excellent external
is that an item may be not available information or input variable for the
because the manufacturer could not forecast of the shipments. It should be
deliver in time. Therefore, the time series noted that the aggregated sales themselves
vol. 4, no. 1
of shipments can be distorted which should not be used as an input variable for
makes forecasting more difficult. The the shipment forecast since the time lag
second option is to use the cumulated between the aggregated sales from the
store orders as dependent variable. This stores and the shipments from the DC to
journal
option makes only sense if the store orders the stores may already be considerable. In
themselves are based on good SKU-store addition, the aggregate forecasts from the
forecasts and have been optimised to fulfil stores already take the effects of calendar
© ecr
logistic and other restrictions. In practice, and promotion events into account, so
usually only shipment data are available. that there is no need to aggregate holiday 63
However, this will quickly change, so that events and promotion events on the DC
both options will be available in the near level. The same is true for the cumulated
future. store orders as dependent variable.
forecast is short-range and has to be relies on one hand on the data that are
complemented with a forecast of the shared by the retailer with the
optimal inventory to fix the lowest manufacturer, for instance sales data,
possible level of stock that has to be promotion data price data etc.; on the
maintained through the ordering and other hand it may also rely on
replenishment process. Altogether, the information that may not be accessible to
micro-forecasting approach with self the retailer, for instance national
learning and inclusion of external advertisement campaigns, information
information is much more promising than about the whole market of retailers since
aggregate and profile forecasting. the manufacturer will supply many
retailers and not one retailer exclusively,
Sales forecasts for the CPFR process etc.
The sales forecast for the CPFR process
follows in many regards similar Third, the forecast of the retailer and
considerations as the sales forecast for the the forecast of the manufacturer are both
stores and the shipment forecast for the mid-range forecasts. While the short-
summer 2004
DC. The main difference lies in the range forecasts typically rely on the most
following points. recent information to adapt quickly to
new developments, the mid-range forecast
First, the time horizon for the sales concentrate mainly on the effects of
forecast is mid-range that is 10 to 13 weeks calendar and promotion events to improve
vol. 4, no. 1
ahead. This does not mean that the short- production planning for excess demand.
term forecasts are not made. It means that Therefore, the highly adaptive models for
every week a forecast is expected, that forecasting on the short-range should be
journal
covers each of the following 13 weeks. replaced by more structural models for the
From the retailers view, these forecasts mid-range. This implies that the
imply that a short-range forecast may be possibility to include information on
© ecr
used for the first four to six weeks and external influences should be used for
64 that a mid-range forecast will be employed forecasting. In turn, this implies that
for the following weeks. again micro-forecasting should be used
rather than aggregate or profile
Second, the forecast of the retailer forecasting. Again, the aggregate sales
(FoR) is complemented by a forecast of the from the store forecasts could be used as
manufacturer (FoM). It should be noted an input variable for the sales forecast of
summer 2004
many manufacturers and/or retailers are already planned for a latter time point,
managed in this way. Therefore, such a the delivery plan of the DC or the
decision making process should be manufacturer to the store, presentation
supported again by a self-learning micro- stock, minimal order sizes and logistic
vol. 4, no. 1
forecasting system where the FoR and the restrictions such as layers or pallets.
FoM for each item are considered as input However, all of this can be taken into
variables for the mid-range forecast of account routinely by a good CAO and can
each item. even be forecasted if only the order for the
journal
first delivery period is used as dependent
Finally it should be noted that some variable. The first delivery period is the
retailers not only expect a joint forecast time between the first day on which the
© ecr
on the aggregate or on the DC level but goods from today’s order are available and
also on the level of the individual store. In the day on which the goods from the next 65
this case, the manufacturer has to possible order day will be available.
undergo a similar forecasting process as It should be seen that this is a rather
for a vendor managed inventory (VMI). restrictive assumption because in practice,
the determination of economic order
quantities tries to optimise the logistic
but for a given series of successive delivery forecast will only be available if only the
periods. Typical examples are the next delivery period is considered as
minimization of logistic costs for one feasible.
supplier by looking at the predicted sales
and optimal inventories for all items of a Order forecast for the CPFR process
supplier for the next K delivery periods to An order forecast for the CPFR process
fulfil restrictions such as the minimum from the retailer’s view is hampered by
order value (for instance $ 1000) across all the same difficulties as an order forecast
items of a supplier or filling up a half or a on the store and on the DC level. Again,
full truck to lower transportation costs. In under the restrictive assumption that
this general case, the building of the order orders are made only for the first delivery
is based on the predicted sales, but it will period, an order forecast can be achieved
be almost impossible to predict how much by looking at the one delivery period
will be actually ordered in detail with orders per item as dependent variable for
great accuracy. the forecast and by using the same
forecasting methods as for the sales data.
summer 2004
Order building and forecast for the DC The forecasted sales (and possibly) the
The same remarks as for the store hold forecasted stock at the beginning of the
true for the DC. In the DC, the logistic delivery period may be used as input
restrictions will play an even greater role variables if a self-learning system with
because the filling of pallets and trucks the possibility to include external
vol. 4, no. 1
much greater influence than in the store. period, the order forecasts will be fairly
Altogether, the forecasts for the shipments difficult to make. To the author’s
to the stores and for the optimal knowledge it has not even been attempted
© ecr
summer 2004
forecasts can be accessed by both partners. This procedure combines some
The retailer also uses a micro-forecasting interesting features. First, it manages the
tool which allows to include external CPFR process not only on the aggregate
influences for sales forecasting and order but also on the store level. Second, it lets
vol. 4, no. 1
building in the store. empirical results decide how much weight
should be given to the FoR and the FoM.
The CPFR pilot process may be described Third, it still includes the element of
briefly by the following steps: personal communication between the
journal
• For each store, sales are predicted by the responsible managers but it supports it
retailer for the all weeks in the short- heavily with IT tools. Fourth, the problem
and mid-range using historical sales as of order forecasting is pragmatically
© ecr
well as information on calendar and solved by using the short-range forecast to
promotion events. These predicted sales adjust for the most recent developments. 67
are aggregated and shown to the
managers in charge at the retailer and at Contrary to other opinions, CPFR does
the manufacturer. work. However, in addition to the right
• Retailer and manufacturer give their structures, processes and symptoms, it
own forecasts on the aggregate level needs a better understanding of what
based on additional data or on their needs to be forecasted, at what level, and
with which statistical and IT support.
Partnership is committed to
furthering the understanding,
improvement and development
of the best-possible models of
economic theory and business
practice, and aims to bring
summer 2004
Robert P. Wilkinson
Great Britain
business thinking to achieve
Co-Chairman
journal
Ulf Kalmbach
Germany
this objective
© ecr
68
ECR Europe Academic Members: ECR Europe Academic Members:
Advisory Panel Council
The ECR Europe Academic Daniel Corsten The ECR Europe Academic Dieter Ahlert
Advisory Panel aims to Switzerland Council represents diverse Germany
guide and support the Georgios I. Doukidis international, national Martin Christopher
Executive Board of ECR Greece and regional academic Great Britain
Europe as a neutral Thomas W. Gruen thinking. It seeks to Raoul Hasselgren
advisory panel, to link USA provoke and inspire best Sweden
between academia and Ludo Van der Heyden business and economic Alan Harrison
practioners, and to enrich France concepts and practice and Great Britain
the teaching of future Arnd Huchzermeier to link academia and Peter Klaus
generations of managers. Germany practitioners. Members Germany
Daniel T. Jones are encouraged to liaise Nirmalya Kumar
Great Britain closely with the national Great Britain
Peter Nolan initiatives of their Peter Lorange
Great Britain country. Switzerland
Luca Pellegrini Alan McKinnon
Italy Great Britain
Werner Reinartz Heribert Meffert
France Germany
Solveig Wikström Peter Schnedlitz
Sweden Austria
Daniel Tixier
France