Vous êtes sur la page 1sur 6

W H I T E PA P E R

Creating a Model of the Retail Business

The use of time-phased planning has been entrenched The Challenge


in manufacturing and distribution in the form of Material
Requirements Planning (MRP), Manufacturing Resource In the 1970s, the planning tenets that defined MRPII were
Planning (MRPII), Enterprise Resource Planning ERP and applied to a distribution operation at Abbott Labs in Montreal,
Distribution Resource Planning (DRP) respectively for decades. resulting in the birth of DRP, which directly connected the
Since the 1970s, these organizations toward the back of the distribution center to the plant and, for the very first time,
retail supply chain have reaped enormous benefits from having enabled two supply chain echelons to be managed with a
an updated, future looking view to their related – both in terms single planning process.
of inventory productivity and higher order fill rates.
In the same fashion, it doesn’t require a huge leap of faith to see
Adoption of these concepts on the retail end of the supply chain that the planning tenets of DRP can also be applied to a retail
have been slow to materialize, but more and more retailers are store. Conceptually, a retail store has more in common with a DC
seeing the potential benefits of having a model of their supply than a DC has in common with a manufacturing plant, so this part
chain from supplier to shelf. According to Supply Chain Digest: of the transition should have been, relatively speaking, easier.

• “There is no way to grow profitably without transforming So why has it taken so long?
supply chains toward new demand-driven strategies and
operations… we need solutions that read POS in near real- First of all, most retailers had a lot of work to do just to get
time and drive supply back across the chain with dramatically the basics in place. When adoption of DRP was at its peak,
improved lead times.” – Gene Tyndall, Tompkins Associates most retailers (even the large, progressive ones) did not have
organized master data, store-level perpetual on hands or a
• Gartner says that a more “predictive” analytics software class
reliable repository of point-of-sale data. This problem has largely
will emerge… that means analytics that aren’t built to help
been solved through widespread adoption of ERP in retail, but it
you understand what happened last week, last month or
was not a quick or easy thing to do.
last year (e.g., scorecards), but rather what is likely to happen
next week or month. While there has always been some of With the main building blocks in place, the next challenge was
that, new approaches to analytic technology and greater (and to a certain extent, still is) processing volume. Seeding a
computing horsepower are taking the capabilities here to DRP planning system with data and maintaining continuously
a whole new level. updated 52-week plans is challenging enough for a large
• “The concepts for “supply chain control tower” coupled with manufacturer with tens of thousands of product/location
more leveraged use of predictive analytics will come to the combinations. The ante is upped significantly, when you consider
forefront. – Bob Ferrari, Supply Chain Matters blog the fact that a moderate to large sized retailer can have product/
location combinations numbering in the tens of millions or even
The culmination of the supply chain evolution that started 30-
the hundreds of millions. The approaches to tackling the volume
odd years ago is a rolling 52-week future view of every product
issue have ranged from inefficient brute force (keep buying
movement in the retail supply chain, from the time the product
hardware until the batch can finish on time) to a “start from
leaves the factory until it’s in the consumer’s hands, updated
scratch” approach that tailors the solution to the problem (if you
every day based on what consumers are actually buying, product
want to win a Formula 1 race, try building an F1 car rather than
by product, and store-by-store – at every point of consumption.
adding more and more horsepower to a minivan). In either case,
This is the “Holy Grail” of retail supply chain planning – a true
advances in computing power over the years have helped both
model of the retail business inside a system.
approaches to the extent that volume is no longer the primary
issue, save for the very largest of retail operations.
Supply Chain

With the basic building blocks in place and the concerns about processing volume more or less overcome, we turn to the last obstacle –
intermittent demand patterns. Retailers on the forefront of adopting integrated time-phased planning (store-level DRP) have come
to realize two things:

1. A very large number of their product/locations are slow sellers, experiencing sales of less than 15 units per year. Depending on
the product mix and market served, intermittent demand streams can account for 30% to 80% of a retailer’s product locations.

2. Most of the forecasting algorithms on the market today were not designed to work in a retail context – they only work when
forecasting fairly continuous demand streams. When you feed a stream of intermittent POS history into one of these models,
you get an unrealistic stream of very small decimal numbers as your forecast.

Regardless, one could argue that: a) it’s statistically impossible to develop an accurate weekly forecast for products that sell intermittently
and b) the quality of the forecast at store level on these items is immaterial, because store replenishment will largely be driven by safety
stock anyway. Both of these arguments are true. But the goal is not merely to replenish the store. It’s to plan all echelons of the supply
chain, from the store right back to the factory, by forecasting the only true demand in the retail supply chain – consumer demand at the
store shelf or point of consumption.

The implication of the shortcomings from using an approach that forecasts small decimals from intermittent history is actually revealed
at the distribution center echelon.

In Figure 1 below, we have a slow selling item being planned in 2 different stores. The stock level at each of these stores is equal to safety
stock (not uncommon for slow sellers) and the forecast has been created using a method that calculates small decimals for intermittent
demand streams:

Figure 1: Plans for a slow selling item in two stores rolling up to a DC

In Store 1, the sales forecast of 0.1 units on Monday is projected to push the on hand below the safety stock, triggering an immediate
replenishment request for 1 unit (which is the minimum order quantity for the store.) The same phenomenon is happening in Store 2.
When these store replenishment requests are aggregated to the DC shipping plan, the DC is now projected to fall below its safety stock,
so a replenishment order of 6 units (minimum order quantity for the DC) is immediately sent to the supplier. This near term “bunching”
problem becomes extremely serious when you think of a full scale retail supply chain with hundreds or thousands of stores and can be
Supply Chain

even further exacerbated if stores must order in case packs, rather than single units (e.g. a forecasted requirement of 0.1 units per
store leads to each store ordering a case pack of 4 units – a 40-fold inflation of the forecasted demand happening at all stores at
the same time.)

The problem is that the replenishment calculation is trying to model a classic smooth saw tooth curve where one doesn’t really exist.
The culprit is a forecast that is not an appropriate1 representation of what will realistically occur at the store shelf (i.e., When’s the last
time you bought 1/10th of a laptop?).

Although this problem is well known and understood among those who have been working on bringing time-phased planning
(store-level DRP) to retail, it has really only been dealt with in a makeshift way.

The “Ignore It” Approach


One way to avoid having a forecast of small decimal numbers is to not forecast at all (in other words, an implied forecast of zero for
every day). In addition to avoiding the bunching problem, this approach also saves considerable processing time, as no forecasts need
to be generated for slow moving items. The major problem with this approach that prevents it from being a viable solution is that it is
only applicable to the store replenishment argument – the safety stock will keep the store in stock anyway, so no forecast is necessary.
However, with no forecast of store-level demand, the link between the store and DC is broken, meaning that the shipment projection
for the DC will be incomplete and a model of the business is not possible. Further complicating matters, any given item can be slow
selling in some stores and fast selling in others. A planner would never know how many stores are actually reflected in their DC shipment
projection and thus would never be able to trust it.

The general idea is to take advantage of the Law of Large Numbers and only apply forecasting logic where it is “worth it” (i.e. at a level
where demand patterns are continuous rather than intermittent), then use relatively simple logic to push the smoothed result down to
store level where it is needed to kick-start the supply chain wide replenishment logic.

The “Bottom-Up, Top-Down” Approach


This approach aggregates history to a regional or national level in order to discern sales patterns that can’t be seen at store level,
generates a forecast using the aggregated history, then prorates this forecast back down to store level based on each store’s contribution,
as per Figure 2 below:

Figure 2: The “Bottom-Up, Top-Down” approach for developing store level sales forecasts

1. The word “appropriate” was chosen deliberately and does not equate to “accurate”. It’s already agreed and established that pure forecast accuracy in the traditional sense is not an attainable
goal for these types of products. We use the term here to mean “a realistic representation of what will happen”.
Supply Chain

There are two significant flaws with this approach:

1. Trying to properly discern a demand pattern across multiple stores can be fraught with problems. This is because many factors other
than geography can affect selling patterns at store level.

“In a Flowcasting simulation for beer sold in Quebec convenience stores, it was noticed that out of 8 stores in the Montreal market, 7 of
them had a spring/summer selling season, but one store had a fall/winter selling season. After an investigation, it was determined that the
8th store was located close to a university. Because the population of students was low in the spring/summer months and high in the fall/
winter months, that store experienced an “upside-down” seasonal pattern.” 2

2. It does not address the root cause of the bunching problem. It is merely a different method of calculating decimal forecasts.

The Solution – Integer Forecasting3


For many retailers, intermittent demand streams are the rule, not the exception.

For the majority of stores in typical retail chain, the current on hand for slow selling items is equal to safety stock. This means that
if a decimal forecasting approach was used for slow selling items at store level, a flood of store replenishment orders would be
triggered overnight.

As the name implies, the integer forecasting approach expresses the forecast as a stream of intermittent integers out into the future.
For slow selling items, we may be able to confidently predict that we’ll sell 4 units in the upcoming quarter, but the exact timing of
when those sales will occur is unknown. What we do know is that on any given day within that quarter, we will either sell 1 unit or none.
Rather than complicating the solution, integer forecasting simplifies the problem. Rather than smoothing out the forecast using small
decimals as a proxy for a probability of a sale occurring on any given day, take the quarterly forecast of 4 discrete sales and sprinkle them
throughout the quarter, as in Figure 4 below:

Figure 4: Actual POS history and forecast output for a slow selling product in a particular store using the integer forecast method

2. André Martin, Mike Doherty and Jeff Harrop, Flowcasting the Retail Supply Chain (2006)
3. The patented integer forecasting approach (U.S. Patent No. 7,552,066) was developed by Darryl Landvater of the RedPrairie Collaborative Flowcasting Group LLC
Supply Chain

Sales history from the prior two years is represented by the green and yellow lines. The dark blue line is the resulting forecast calculated
using the integer forecast method. Notice that the distance between the blue spikes is not the same. This reflects the seasonal selling
pattern for this model. The spikes are further apart during a time of year when this model does not sell as well, and the spikes are closer
together during the time of the year when the product is more likely to sell. For example, during June and July, the spikes are spaced
farther apart than for example during October and November. During December and the first half of January (leading up to Christmas
and up until the Super Bowl) the probability of a sale is so large that one week has a forecast of two. Contrast this with June and July
where the expected sales are only about one a month.

The historical sales pattern requires some explanation since there are quantities like 2.5 in some weeks. The sales history for this model
at this store (the green and yellow spikes) shows demand for both this product and the predecessor model (after August on the green
line for this model, and the predecessor model before August). This companie’s products have short lifetimes, typically a year or less,
consequently most models are forecast based on a predecessor model using some assumptions about how well the new model will
sell compared to the predecessor. These assumptions create similar-to sales history in decimals. For example, a new model which is
expected to sell at 125% of the predecessor would have a similar-to sales history of 1.25 if one of the predecessor sold, and 2.5 if two of
the predecessor sold.

Note that the predecessor model sold infrequently during June and July last year (green spikes). Similarly, the forecast for this year has
infrequent spikes during June and July (blue spikes). The placement of these blue spikes is based on the seasonal selling profile, and
also some randomization. This randomization is a recognition of the fact that we can predict that one week may be more likely to have
a sale than another week, but we are not able to predict that a given store will sell a TV in the third week of June, for example. Integer
forecasting solution is not a separate forecasting algorithm that must be separately tuned, maintained and understood by users. The
underlying base of the calculation is a profile based forecasting calculation, similar to what has been used successfully in store-level
forecasting for years. A threshold is used to determine whether or not the forecasts will be expressed as discrete integers or as a weekly
decimal quantity. For example, a model selling 60 units per year will be forecast in weekly decimal quantities such as 1.2 in a particular
week. Another model selling 10 a year will be forecast in integers spaced some distance apart, as shown in Figure 4 during the months
of June and July. While the forecasts are represented as integers, the actual forecast rows in the database are decimal quantities having
a time period of a month or quarter. For example, a model at a store may have a forecast of 3.2 for the month. This quantity will be
represented as a series of integers on the graph and in the DRP spreadsheets. As sales occur, forecast consumption logic is used to
predict the remaing quantity to be forecast during the time period (month or quarter). So, for example, if 2 units were to sell in the first
week of the month, then the remaining forecast for the month would be 1.2. This is the normal forecast consumption logic that has
been in use for decades.

The result is a forecast which represents both the seasonal selling profile as well as the fact that we can predict how many items
may sell in a time period such as a month, but cannot predict with accuracy the week or day when a particular unit will sell. This is a
representation of the real world, and as such, provides a valid model of the business which can be used to calculate dependent demand
at a distribution center or supplier, as well as providing the abillity to do capacity planning, transporation planning, financial planning
and simulations. As demonstrated in Figure 5 below, store level DRP (Flowcasting) gives an aggregation of planned shipments to the
DC echelon without causing the bunching effect described in Figure 1:

Figure 5: Integer forecasting used to


plan 2 supply chain echelons using
only forecasts of consumer demand
Supply Chain

The Bottom line


“Integer Forecasting” combined with store level DRP enables retailers today to finally aspire to create a model of their business inside the
computer. As stated, at the beginning of this paper, the use of MRP, MRPII, ERP and DC level DRP have been entrenched in manufacturing
since the 1970s and these organizations have reaped enormous benefits from having an updated, future looking view to their related –
both in terms of inventory productivity, higher order fill rates and lower operating costs.

Adoption of “Integer Forecasting” combined with store level DRP on the retail end of the supply chain is now enabling retailers to develop,
over a 52-week planning horizon and beyond, very detailed time phased sales forecasts, inventory plans and replenishment/buying
plans for all their products as opposed to only 40% of their products which has been the case until recently.

The VICS New Best Practice Guideline “The Ultimate Retail Supply Chain Machine –Connecting the Consumer to the Factory” does an
excellent job of describing how a vision of a complete retail supply chain is becoming a reality and how it can be fully integrated from
the Consumer to the Factory.

The Twenty First Century Retail Supply Chain

WEB AMERICAS EUROPE ASIA PACIFIC


jda.com
US UK & Northern Europe Singapore Japan South Korea
EMAIL +1 800 479 7382 +44 (0) 1344 354500 +65 6305 4350 +81 3 4461 1000 4-Color
+82 2 3016 0700
info@jda.com
Canada & Latin America France & Southern Europe Australia China India
+1 480 308 3555 +33 (0)1 56 79 27 00 +61 3 9860 1000 +86 21 2327 9400 +91 22 6700 0794

Copyright © 2014, JDA Software Group, Inc. All rights reserved. JDA is a Registered Trademark of JDA Software Group, Inc. All other company and product names may be Trademarks, Registered
Trademarks or Service Marks of the companies with which they are associated. JDA reserves the right at any time and without notice to change these materials or any of the functions, features or
specifications of any of the software described herein. JDA shall have no warranty obligation with respect to these materials or the software described herein, except as approved in JDA’s Software
License Agreement with an authorized licensee.

04.25.14

Vous aimerez peut-être aussi