Rocket Science Retailing and The Research Opportunities It Creates For Us

Rocket Science Retailing and the
Research Opportunities
it Creates for Us
Marshall Fisher
The Wharton School
"The Role of Optimization in Supply Chain Management"

Institute for Mathematics and its Applications (IMA)
September 23-27, 2002
Harvard/Wharton Project: How can recent advances in Information Technology
improve the way retailers forecast demand and plan supplies?
Apparel, Consumer Books, CDs, Other

Footwear Electronics Jewelry, Toys Product
and PCs Theme Stores Categories and
Multiple Product
Categories
David’s Bridal CompUSA Borders Group Ahold
Footstar Office Depot Bulgari Christmas Tree Shops
Gap Inc. Radio Shack The Disney Store CVS
GH Bass Staples Tiffany & Co. Federated
Maurices the good guys! TransWorld Enter HE Butt
Nine West Tweeter etc Warner Bros. Iceland Frozen Foods
The Limited Zany Brainy JC Penney
World Marks & Spencer
Zara QVC
Sears
Marshall Fisher, Ananth Raman & Anna McClelland, “Rocket-Science Retailing is Almost Here – Are
You Ready?” Harvard Business Review July – August 2000
© 2002 Marshall L. Fisher – The Wharton School
Agenda
• Overview of retail operations – what do retailers do?
• State of retail supply chain management today – findings

from the Harvard/Wharton Rocket Science Retailing Project
• ‘Rocket science’ retailing – injecting high grade analytics

into retailing
• Some steps toward the vision
• Suggested research opportunities

Research collaborators
Gerard Cachon Rocco Mosconi

Nicole DeHoratius Kumar Rajaram
Karen Donohue Walter Salmon
Vishal Gaur Zeynep Ton
Gurhan Kok Noel Watson
Anna McClelland Giulio Zotteri

Retail operations
Goal = Max Profit = revenue – variable cost (cost of goods, Stores Consumers
inventory holding and obsolescence)
Raw
New Product
Materials/ Plants DC
Development/
Components
Selection
Sourcing
Buyer/Planner
Buyer & Designer Store Manager
•Forecast Demand •Execution e.g.
•Assortment planning
• Set Inventory levels product display,
customer service
•Pricing & Promotion
•Lead time •POS price
changes
management e.g.
when to use air
•Accuracy of
freight
inventory data
State of retail supply chain management today
31%
Department 26%
Too many of
store the wrong
markdowns 21%
as a
products
percentage of
sales 16%
11%
6%
1970 1980 1990 1995
“One third of customers entering Too few of

a store leave without buying the right
because they can’t find what products
they came for”

Two knit tops appeared well bought at the chain style color level …

But at the style/color/size/store level, the blue hooded top was seriously under-
bought, especially in the largest size
Week Week

And SKU-door level analysis reveals huge lost margin

What’s causing these problems?
Long lead times Up to 11 months for apparel

products sourced from Asia, but
some – World and Zara –
achieved 2 week lead times
Inaccurate forecasts 50% -100% errors for a season

forecast at chain level is typical

Preseason forecasts are highly inaccurate
(Each Blue Dot Corresponds to a Particular Style/Color of
Women’s Dress for a Major Cataloger)
4000
3000
Actual
Demand
2000
1000
0
0 1000 2000 3000 4000
Forecast Demand
Average Error is 55%

What’s causing these problems?
Long lead times Up to 11 months for apparel

products sourced from Asia, but
some – World and Zara –
achieved 2 week lead times
Inaccurate forecasts 50% -100% errors for a season

forecast at chain level is typical
Store level execution Stock outs caused by inventory

cannot be assumed record errors or product in the
backroom

Albert Heijn example – store sales vs customer entrances reveals problems
1 unit left over - looks like 100%

service, but possible shrink
results in lower service
Incorrect
Refill from
backroom
Retailers have data that can help, but don’t use it
Scanners and data

warehousing has
created huge “We are awash in data and starved
databases, but for information”
retailers lack ability
to analyze this data
“How can history History -> seasonality, price

possibly be useful elasticity, store/sku patterns,
in a fashion size mix, reaction to
business?” promotion, etc.

Retailers have data that can help, but don’t use it
Inadequate tools - Reality is many products like

Planning paradigms shoes, home fabrics, books &
are replenishment of music, toys, etc have features
staples and ‘one and of staples and fashion –
done’ category limited life but can replenish &
planning for fashion items matter
products
•Need to manage at But can’t do this without analytic

the SKU store level tools – too much data for a
manual approach

Too many decisions to actively manage
Data for median retailer in our survey
• $1.1 billion annual sales
• 2.3 million store/sku stock points e.g.

230 stores x 10,000 sku’s
• 45% gross margin
---> $4.14 gross margin per week per stock point
Only way to actively manage this is by computer

The payoff from fixing this can be huge
Better analytics can Data + analytic tools

easily double a
= better in-stock + fewer mark-
retailers profit downs
=profit increase of 5 – 10% of

revenue

10% higher in stock leading to 10% more sales doubles profit
Gross Margin Profit before tax
Jewelry Stores 52.4% 2.7%
Consumer Electronics 33.1% 2.3%

and Computer Stores
Apparel 36.1% 4.0%

and Accessory Stores
Department Stores 32.2% 2.5%

Retail operations
Goal = Max Profit = revenue – variable cost (cost of goods, Stores Consumers
inventory holding and obsolescence)
Raw
New Product
Development/
Components
Selection
Sourcing
Buyer/Planner
Buyer & Designer Store Manager
•Forecast Demand •Execution e.g.
•Assortment planning
• Set Inventory levels product display,
customer service
•Pricing & Promotion
•Lead time •POS price
changes
management e.g.
when to use air
•Accuracy of
freight
inventory data
Scientific retailing
Stores Consumers
Raw
New Product
Development/
Components
Selection
Sourcing
Determine Track Data Accuracy

Optimal Inventory,
Estimate
Price and
Demand Density •Sales & Inventory by
Promotion over
Time for Season SKU/Store
•Loyalty Card Information
Data Warehouse for model calibration
-seasonality, price elasticity,
size mix, store differences
Product Life Cycle Planning
Phase Down / End of
Merchandise Test Read & React Replenishment / Model Life / Markdown
Period Stock Setting
Pregnancy Youth Middle age Retirement

Forecast initial Read early sales, Traditional Manage down
demand
update forecast long life cycle replenishment
•Survey experts and reorder as replenishment flow
needed – set target
•Mall intercepts Drop product in
stock levels
Expedite selected stores
•Prior products for each sku-
•Merchandise test
resupply? store Markdown price
or transfer to
Initial buy Rollout to more outlet stores
stores or drop
Preposition
stores?
materials
Merchandise Depth Test
Typical calendar for Fall season
Early
Test Read Primary sales season Closeouts
May Aug Sept Dec Jan Feb
Goal is to predict total chain primary season sales

from test sales
“Accurate Retail Testing of Fashion Merchandise: Methodology and

Application”, Marshall Fisher and Kumar Rajaram, Marketing Science,
Summer, 2000
Test design issues
Test Stores

Testing practice
• Written survey
– 25 out of 27 conduct tests
– Median effectiveness = 6 on 10 point scale -> most retailers test but
think they don’t test well
• Interviews
– Retailer Y says “Retailer X has found 2 stores that perfectly predict
next season.”
– Retailer X says “Our tests are inaccurate because of wrong test
stores, poor execution, weather …. “

Typical testing process
• Choose 25 average stores e.g. average sales rate
• Place product in those stores for 4 weeks
• Forecast chain season sales as
(stores in chain/25) (weeks in season/4) * Total test sales

Predicted vs actual using typical process with 25 test stores at apparel retailer
Actual
Sales
Predicted Sales
New testing process uses store clustering to find representative test stores
• Cluster stores based on similarity of sales mix. Use K-median

clustering for this.
• Choose most representative store from each cluster to be a test store.
• Fit formula that most accurately predicts chain season sales from test
store sales. Test stores aren’t weighted equally.
• Can also predict sales of each cluster, and hence store sales

Test store selection problem
Season sales Difference in sales

mix between store i
of store i
& store j
= 1 if store j
=1 if store i chosen as a test
assigned to store
cluster
represented by
test store j

Predicted vs actual using new method with 10 test stores at same retailer
Actual
Season
Sales
Predicted Season Sales

Performance of the new method for four retailers
New Test Process Traditional Test Profit increase

Retailer Forecast Error Forecast Error due to reduced
10 Test stores 10 Test stores lost sales and
markdowns
Women’s 12.9% 41.9% 18% of revenue

Apparel
Nine West 17.6% 27.9% 11% of revenue
Meldisco 12.7% 22.7%
Home 17% 43%

Fabrics

What Drives Mix? Correlation Between Sales Mix and Store Descriptors

Product Life Cycle Planning
Phase Down / End of

Forecast initial Read early sales, Traditional Manage down
demand
update forecast long life cycle replenishment
•Survey experts and reorder as replenishment flow
needed – set target
•Mall intercepts Drop product in
stock levels
Expedite selected stores
•Prior products for each sku-
•Merchandise test
resupply? store Markdown price
or transfer to
Initial buy Rollout to more outlet stores
stores or drop
Preposition
stores?
materials
Data from a fashion cataloger – early sales is highly predictive
Expert Forecast by a Committee Forecast Obtained by Extrapolating

of Four Merchandisers the First 2 Weeks (11%) of Orders
4000 4000
3000 3000
Actual Actual
Demand Demand
2000 2000
1000 1000
0 0
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Forecast Demand Forecast Demand
Average Forecast Error is 55% Average Forecast Error is 8%

Reading and reacting to early sales
100.0
90.0 This portion of demand can
80.0
% of Total 70.0
be supplied based on an
Season 60.0 accurate forecast.
Sales 50.0
Achieved 40.0
30.0
20.0
10.0
0.0%
Lead Time Week of season
Initial Shipment to Cover Replenishment

Read/ React Period + Read Market
Supply
Safety Stock Based on Error & Order More of
Margin in the Forecast Hot Sellers
4000 4000
3000 3000
Actual Actual
Demand Demand
2000 2000
1000
1000
0
0
0 1000 2000 3000 4000
0 1000 2000 3000 4000
Forecast Based on Initial
Initial "Gut" Forecast Demand or Test
Forecast Committee Results
Product Anna Laurie Julie Kim Committee Committee Actual Forecast

Average Standard Demand Error
Deviation
Navy 89 86 102 102 95 8 85 -10
Turtleneck
Red 63 76 152 51 86 46 199 113
Cardigan
Blue Vest 30 91 183 76 95 64 29 -66
When the Committee Agrees, They Tend to be Accurate

The Committee Process Is A Powerful Way To Determine What You Can And
What You Cannot Predict
2400
High
Error
2000
Average Average
1600 Error = Error =
Forecast 116 units 645 units
Error
1200
800
400
Low
0
Error
0 100 200 300 400
High Agreement Low Agreement

Standard Deviation of the Individual
Forecasts of the Four Person Committee

The Initial Buy Minimizes the Cost of Stockouts and Closeouts
Initial Buy Replenishment Replenishment End of Life

Order Placed Order Arrives
Initial Demand
Life cycle Demand
Probability
0.04
0.035
0.03
We e k 1 de mand
0.025 S tockout Ris k
0.02
Initial Demand
0.015
0.01
0.005
0
0 .0 5 0 20 40 60 80 100 120
0 .0 4
0 .0 3
Re turnRRis k
Closeout Life cycle
0 .0 2
de mand
0 .0 1
0
- 0 .0 1 0 5Initial
0 Buy 1 00
Quantity
Buy Quantity 1 50 2 00 2
Read/react can be used to evaluate the impact of Lead Time reduction
18%
15%
Incremental $
Gross Margin as a
% of Base Case
Revenue 10%
Benefits of Read
7.5% React
3.5%
0 3 6 9 12
Base Case
Lead Time (Weeks)

Nine West Example
Objective To improve demand forecast using early

sales data and decide mid-season replenishment order
quantities
Jan Feb March April May June
Initial Supply
in Stores
Replenishment Order Arrivals
Orders

Data from a fashion cataloger – early sales is highly predictive
Expert Forecast by a Committee Forecast Obtained by Extrapolating

of Four Merchandisers the First 2 Weeks (11%) of Orders
4000 4000
3000 3000
Actual Actual
Demand Demand
2000 2000
1000 1000
0 0
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Forecast Demand Forecast Demand
Average Forecast Error is 55% Average Forecast Error is 8%

Simple Extrapolation of Early Sales Didn’t Work for

Because price markdowns and stockouts were affecting sales

Needed forecast model that considers impact of price and supply
Forecast of sales of SKU j in week t
Estimate of season sales based on early sales
Seasonality factor estimated from history
Price factors
Inventory
Factor

And then we got accurate forecasts

Weekly forecast vs actual for one sandal

Application of this approach to one style-color
Original Optimized Actual

plan plan based results
on our
work
Unit Receipts 6,210 14,160 12,325
Unit sales 9,882 14,875 14,579
Gross Margin $ $385,497 $580,274 $562,269

Product Lifecycle Planning
Different Challenges require Different Algorithms

New Product Intro Choosing representative stores for entire chain to test product and extrapolate a forecast
Initial Buy Determining the optimal buy quantity, risk adjusted for stockouts and closeouts
Read React Correctly interpret early sales to make an intelligent replenishment purchase
Replenishment Accurately set stock levels at both the stores and the DC's
End of Life Find the optimal price and conditions to liquidate inventory
4R System Clients – Partial List

Retailer/Supplier Collaboration via Common 4R Planning Platform
Sales & inventory Supply chain
data, supply chain
4R Supply Planner parameters
parameters
Consumer forecast, DC and Consumer forecast, DC
store inventory levels inventory levels, Asia buys
Discussions
Buyers Account
planners
POS data & Product

Linens N’ Orders American Flow
Things Pacific
DC Information
DC Flow
Asian
Stores
Stores Suppliers
Stores
On Hand inventory prior to 4R
OH-WOS
35.0
30.0
25.0
20.0 Pilot
4R
Peer
Cont
15.0
Dif.
10.0
5.0
0.0
3/9
4/6
6/1
5/4
2/23
3/23
6/15
6/29
7/13
7/27
4/20
5/18
8/10
Lost sales prior to 4R
Lost Sales
8.0%
7.0%
6.0%
5.0%
4.0%
3.0% Pilot
4R
2.0% Cont
Peer
1.0% Dif.
0.0%
-1.0%
-2.0%

Lost sales for 4R SKUs decreased from 6.6% to 2.2%
Lost Sales
8.0%
7.0%
6.0%
5.0%
4.0%
3.0% Pilot
4R
2.0% Cont
Peer
1.0% Dif.
0.0%
-1.0%
-2.0%
Start of pilot
Weeks of supply relative to Control SKUs decreased 7%
OH-WOS
35.0
30.0
25.0
Pilot
20.0 4R
Peer
15.0 Cont
Dif.
10.0
5.0
0.0
3/9
4/6
6/1
5/4
2/23
3/23
6/15
6/29
7/13
7/27
4/20
5/18
8/10
Start of pilot
Summary – the profit increase from retail analytics is large
Phase Down / End of

Survey experts
Read early sales, Traditional long life Manage down
Initial buy cycle replenishment replenishment flow
update forecast
Merchandise test and reorder as 3% reduction in
needed lost sales with 7%
less inventory
Profit increase = Profit increase = from next
11% - 18% of 3.5% of revenue generation
revenue replenishment
46% gross
algorithm
margin increase
Scientific retailing
Stores Consumers
Raw
New Product
Development/
Components
Selection
Sourcing
Determine Track Data Accuracy

Optimal Inventory,
Estimate
Price and
Demand Density •Sales & Inventory by
Promotion over
Time for Season SKU/Store
•Loyalty Card Information
Data Warehouse for model calibration
-seasonality, price elasticity,
size mix, store differences
Research Opportunities in Retailing
• Enhancing traditional forecasting/inventory management for a real context

– Short life cycle products
– Correcting for lost sales due to stockouts
– Measuring micro consumer response to stockout – substitute, lose sale or
lose customer
– Assortment planning
– Forecasting product A based on sales of B, C, …
– Full optimization of all levers – inventory, price, advertising, store
placement, etc.
– Probability density estimation – forecasting extremal events with sparse
data
• Retail experimentation – merchandise testing
• Store as factory – adapting TQM and TPS to store operations

– Improving data accuracy
– Avoiding phantom stockouts
– Empirical studies to discover best practice like the the MIT auto study

Real demand differs from textbook image

Research Opportunities in Retailing
• Enhancing traditional forecasting/inventory management for a real context

– Short life cycle products
– Assortment planning
– Correcting for lost sales due to stockouts
– Measuring micro consumer response to stockout – substitute, lose sale or
lose customer. Combine marketing’s consumer view with operations
supply chain view
– Full optimization of all levers – inventory, price, advertising, store
placement, etc.
– Probability density estimation – forecasting extremal events with sparse
data
– Data aggregation/disaggregation
• Retail experimentation – merchandise testing
• Store as factory – adapting TQM and TPS to store operations

– Improving data accuracy
– Avoiding phantom stockouts
– Empirical studies to discover best practice like the the MIT auto study

Rocket Science Retailing and The Research Opportunities It Creates For Us

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Rocket Science Retailing and The Research Opportunities It Creates For Us

Hochgeladen von

Copyright:

Verfügbare Formate

Rocket Science Retailing and the

"The Role of Optimization in Supply Chain Management"

Apparel, Consumer Books, CDs, Other

• Overview of retail operations – what do retailers do?

• State of retail supply chain management today – findings

• ‘Rocket science’ retailing – injecting high grade analytics

• Some steps toward the vision

• Suggested research opportunities

© 2002 Marshall L. Fisher – The Wharton School

Gerard Cachon Rocco Mosconi

© 2002 Marshall L. Fisher – The Wharton School

“One third of customers entering Too few of

© 2002 Marshall L. Fisher – The Wharton School

© 2002 Marshall L. Fisher – The Wharton School

© 2002 Marshall L. Fisher – The Wharton School

© 2002 Marshall L. Fisher – The Wharton School

Long lead times Up to 11 months for apparel

Inaccurate forecasts 50% -100% errors for a season

© 2002 Marshall L. Fisher – The Wharton School

Average Error is 55%

Long lead times Up to 11 months for apparel

Inaccurate forecasts 50% -100% errors for a season

Store level execution Stock outs caused by inventory

© 2002 Marshall L. Fisher – The Wharton School

1 unit left over - looks like 100%

Scanners and data

“How can history History -> seasonality, price

© 2002 Marshall L. Fisher – The Wharton School

Inadequate tools - Reality is many products like

•Need to manage at But can’t do this without analytic

© 2002 Marshall L. Fisher – The Wharton School

Data for median retailer in our survey

• $1.1 billion annual sales

• 2.3 million store/sku stock points e.g.

• 45% gross margin

---> $4.14 gross margin per week per stock point

Only way to actively manage this is by computer

© 2002 Marshall L. Fisher – The Wharton School

Better analytics can Data + analytic tools

=profit increase of 5 – 10% of

© 2002 Marshall L. Fisher – The Wharton School

Gross Margin Profit before tax

Jewelry Stores 52.4% 2.7%

Consumer Electronics 33.1% 2.3%

Apparel 36.1% 4.0%

Department Stores 32.2% 2.5%

© 2002 Marshall L. Fisher – The Wharton School

Determine Track Data Accuracy

Pregnancy Youth Middle age Retirement

Typical calendar for Fall season

Goal is to predict total chain primary season sales

“Accurate Retail Testing of Fashion Merchandise: Methodology and

© 2002 Marshall L. Fisher – The Wharton School

© 2002 Marshall L. Fisher – The Wharton School

• Choose 25 average stores e.g. average sales rate

• Place product in those stores for 4 weeks

• Forecast chain season sales as

(stores in chain/25) (weeks in season/4) * Total test sales

© 2002 Marshall L. Fisher – The Wharton School

• Cluster stores based on similarity of sales mix. Use K-median

• Choose most representative store from each cluster to be a test store.

© 2002 Marshall L. Fisher – The Wharton School