1

Markdown Optimization for an Indian

Apparel Retailer

2
Apparel Retail Industry

Seasonal Short shelf life Unpredictable

Excess Inventory
3
Apparel Retail Industry

Seasonal Short shelf life Unpredictable

Excess Inventory
4
http://www.dailymail.co.uk/femail/article-4170226/Australian-Open-2017-tennis-WAGS-
opt-casual-dress.html 5
 Problem Statement

CURRENT STATE GAP / TRIGGER FUTURE STATE

There is an opportunity to use
 High Inventory at hand  Client has a scientific way
advanced analytics to
of clustering similar stores
establish a scientific method
of optimizing markdown
 Retail industry standard of Client has a model to
decisions
overall markdown is 17-18% predict demand for the
identified segment
 Overall markdown given by
Client is around 23-25%  Client has the model which
predicts the optimal
 Maximize the revenue by KEY QUESTION markdown for the segment
optimizing markdown • Which bricks or styles to
markdown in the segment?
• At what level to
markdown?
• What is the optimal
markdown percentage ?

6
Scope : End of Season for ‘Women’s Ethnic Wear’

7
Understanding Current EOSS Markdown Process

8
 Solution Approach / Methodology
Data Aggregation
Perform Store Build Demand Models, and
In Cluster-Brand-
Clustering Brick
Optimization Model, solve for each Validation
group
Groups
Cluster 1.Brand1. Brick1
Store to
Cluster 1
Cluster1. Brand5. Brick4 Compare

Demand Models for each of 252

Optimization Model 1
Maximized
Cluster2.Brand1. Brick 1 Optimization Model 2 Revenue with
Store
Cluster 2
to Actual

combinations
Optimization Model 3
Cluster2.Brand5. Brick 4 Revenue
“ “
earned in the
Store
Cluster3.Brand1. Brick 1
to
“ “
most recent
Cluster 3
Cluster3. Brand5. Brick 4
Optimization Model 252 season

Cluster4. Brand1. Brick 1

Store to
Cluster 4 Cluster4. Brand5. Brick 4

9
Clustering- An Analytical approach

10
Background
 Present scenario- Single markdown strategy for a Brand-Brick-Style
combination across all stores. Is this right?
 At what level should Clustering be done? Brand? Brick? Style?
 Present distribution of average markdown:

 Metrics used:

11
Explanatory Data Analysis
 Do stores perform differently across various brands?

 Two Factor Anova Results:

12
Methodology
 No missing values. Outliers capped at 97th Percentile to avoid isolated
clusters.
 Distance Matrix: Gower’s Coefficient, Zone lower weightage (Semi-
supervised clustering)
 Number of clusters:

 Clustering Algorithms:
• Hierarchical
• PAM- Robust version of K-Means

13
 Gower’s Similarity
• Gower’s similarity Dij
n
 Dijk Wijk
Dij  k 1
n
 Wijk
k 1

• Dijk for continuous variable

X ik  X jk
Dijk  1 
Rk

Xik and Xjk are the values of observations i and j for variable k. Rk is the range of
variable k.

• Dijk for Binary variable

1, X ik  X jk
Dijk  
0, otherwise

14
Model Selection

 Validation and
Model
Selection:

 Overall Cluster
separation
Graph:

15
 Profiling
 Brand-wise Separation of clusters

 Cluster Profiling:
0.80 Mean Centre Markdown Sensitivity 10.00 Mean Centre NPSS
0.60 8.00

0.40 6.00
Brand-1
4.00
0.20 Brand-1
Brand-2
2.00 Brand-2
Brand-3
- Brand-3
Brand-6 -
Brand-6
(0.20) Brand-8 (2.00)
Brand-8

(0.40) (4.00)

(6.00)
(0.60)
1 2 3 4 (8.00) 16
1 2 3 4
 Insights
 Cluster-1:
 Dominated by North & East Zone
 Brand- 2 & 3 is the least in terms of markdown sensitivity (Highest Markdown given to
generate Sales)
 Brand-1, 3, 6 & 8 is the least in terms of NPSS (Least profit margin earned)
 Cluster-2:
 Dominated by West Zone
 Brand- 1, 6 & 8 is the least in terms of markdown sensitivity (Highest Markdown given to
generate Sales)
 Cluster-3:
 Dominated by South Zone
 Brand- 1, 3, 6 & 8 is the highest in terms of markdown sensitivity (Least Markdown given
to generate Sales)
 Brand-6 & 8 is the highest in terms of NPSS (Highest profit margin earned)
 Cluster-4:
 Dominated by South Zone
 Brand- 2 is the highest in terms of markdown sensitivity (Least Markdown given to
generate Sales)
 Brand-1 & 3 is the highest in terms of NPSS (Highest profit margin earned) 17
 Solution Approach / Methodology
Data Aggregation Build Demand
Perform Store In Cluster-Brand- Models, and
Clustering Brick Optimization Model,
Validation
Groups solve for each group
Cluster 1.Brand1. Brick1
Store to
Cluster 1
Cluster1. Brand5. Brick4 Compare

Demand Models for each of 252

Optimization Model 1
Maximized
Cluster2.Brand1. Brick 1 Optimization Model 2 Revenue with
Store
Cluster 2
to Actual

combinations
Optimization Model 3
Cluster2.Brand5. Brick 4 Revenue
“ “
earned in the
Store
Cluster3.Brand1. Brick 1
to
“ “
most recent
Cluster 3
Cluster3. Brand5. Brick 4
Optimization Model 80 season

Cluster4. Brand1. Brick 1

Store to
Cluster 4 Cluster4. Brand5. Brick 4

18
Accurate Sales Forecast for EOSS : A Strategic Advantage

Sales
Forecast

Price Maximize Promotion

Elasticity Revenue Effect

Ageing
Effect
19
Exploratory Data Analysis: Non Linear Relationships
Sales Vs Discount % Sales Vs Net Price

 Correlation between sales & discount is  Correlation between discount & sales is
positive. negative.

 The relationship is not linear but  The relationship is not linear but non-
exponential. linear.
 If sales generated by giving 20%  If price is increased to higher levels
discount is x then sales generated by sales drop drastically
giving 40% discount is not 2x but 20
much more than that.
Exploratory Data Analysis: Different Bricks behave differently with age

Brick 1 - Sales Vs Age Brick 2 - Sales Vs Age

 Correlation between Sales & Age is  Correlation between Sales & Age is
positive. negative.

 Sales has an increasing trend with Age  Sales has an decreasing trend with
Age
21
Exploratory Data Analysis: Promotions have an effect On Sales

22
 Analytics Approach : Demand Forecasting Model
1 2 3 4
Demand Function & Forecast
ETL Data Cleansing Regression With Lags
Validation

Client
ERP

Customer Behavior Value

Customer Demographics Product Usage Interaction Value
Identif ication Person Services Service Usage Number of Customer value
contacts
Individual Home Price Plan Service Volume Time since last Satisf action
Demographics contact
Address sector Contract Usage pattern Contact history Customer
Geography revenue
Status Micro Market Handset Payment history Campaign history Customer cost

Community Mobility Loyalty programs Customer value

Customer scores scores
Price Calling Circles Top Ups
Account details development

For For For

Profiling Behavior Segmentation Value
only Segmentation

 Two years weekly data  Missing Value  Visualize the original  Identify the demand
is extracted from ERP Treatment time series function
systems and rolled up at  Treat Outliers  Plot Cross  Generate future
cluster, brand, brick level. Correlations forecast
 Regression with lags  Perform Out-Of-Time
model to forecast Validation
demand.
 Transform the time
series & predictors
23
 Analytics Approach: Demand Function
Demand Functional Form

Sales = 𝑆𝑎𝑙𝑒𝑠 𝐿𝑎𝑔1𝛽1 ∗ 𝐷𝑖𝑠𝑐𝑜𝑢𝑛𝑡𝛽2 ∗ 𝐷𝑖𝑠𝑐𝑜𝑢𝑛𝑡 𝐿𝑎𝑔1 𝛽3 𝑒 𝐼𝑛𝑡𝑒𝑟𝑐𝑒𝑝𝑡+𝛽4∗𝐴𝑔𝑒+𝛽5∗𝑃𝑟𝑜𝑚𝑜𝑡𝑖𝑜𝑛 𝐹𝑙𝑎𝑔

Cross - Correlation Plot

Key Insights

• Price elasticity (β2) – Quantifies how much sales varies when discount is
changed by 1 %
• EOSS Discount Lag Effect (β3) – Captures the lag in sales with EOSS discount
• Ageing effect (β4) - Explains how much sales varies as the merchandise ages
• Promotion Effect (β5) – Quantifies how much sales increases during EOSS
promotion weeks compared weeks without promotion 24
Analytics Approach: Forecast Model Comparison

25
 Clustering , Demand Forecasting to Optimization
Data Aggregation Build Demand Models,
Perform Store In Cluster-Brand- and Optimization
Clustering Brick Model, solve for each
Validation
Groups group
Cluster 1.Brand1. Brick1
Store to
Cluster 1
Cluster1. Brand5. Brick4 Compare

Demand Models for each of 252

Optimization Model 1
Maximized
Cluster2.Brand1. Brick 1 Optimization Model 2 Revenue with
Store
Cluster 2
to Actual

combinations
Optimization Model 3
Cluster2.Brand5. Brick 4 Revenue
“ “
earned in the
Store
Cluster3.Brand1. Brick 1
to
“ “
most recent
Cluster 3
Cluster3. Brand5. Brick 4
Optimization Model 80 season

Cluster4. Brand1. Brick 1

Store to
Cluster 4 Cluster4. Brand5. Brick 4

Optimization model was required to be solved for 252 groups 26

 Formulation for Optimization
Maximize MRP*{ (1-d1)*s1 +(1-d2)*s2 +(1-d3)*s3 +(1-d4)*s4 }
Objective

1 2
Sales <= Demand or Ln(Sales) <= Ln(Demand) Sales <= Starting Inventory

 Ln(si) <= Ln(Di )

 The constraints for four weeks:
 From Demand Model, the functional form for o s1 <= inv1
Demand (D) is : o s2 <= inv2
Ln(D) = Intercept + sales_lag1_coef * o s3 <= inv3
Ln(preweek_sale) + disc_coef * o s4 <= inv4
Ln(current_week_discount%) + disc_lag1_coef
* Ln(prev_week_discount%) +promo_coef *
(promo_current_week_flg) + age1_coef * age 3
Inventory & Sales relationship
 The constraints for each of the four weeks:
 The constraints for four weeks:
o Ln(si) <= Intercept + sales_lag1_coef * Ln(si-1) o Inv2=inv1-s1
+ disc_coef * Ln(di) + disc_lag1_coef * Ln(di-1) o Inv3=inv2-s2
+promo_coef * (promo_current_week_flg) + o Inv4=inv3-s3
age1_coef * agei o eossinv=inv4-s4
27
 Formulation for Optimization
4 5
Discounts should be non-decreasing Age of articles

 The constraints for four weeks can be  The constraints can be written as:
written as:
o age2=age1 + 1;
o d2 >= d1 o age3=age2 + 1;
o d3 >= d2 o age4=age3 + 1;
o d4 >= d3

6 7
All variables are non-negative
Boundary of discounts

 The discounts can be between 10% and  The constraints for four weeks :
60%
o d1<=0.6, d1>=0.1; o d1>=0,d2>=0,d3>=0, d4>=0,
o d2<=0.6, d2>=0.1; o s1>=0,s2>=0, s3>=0,s4>=0
o d3<=0.6, d3>=0.1; o inv1>=0,inv2>=0, inv3>=0,inv4>=0, eossinv>=0
o d4<=0.6, d4>=0.1;
28
 Automation to solve Optimization for all combos
CLUSTER.BRAND.BRICK
Intercept
Demand Model

Sales lag1 coefficient Non-Linear

Discount coefficient Optimization
 Lag1 Discount coefficient
Promotion coefficient
Age coefficient

 Optimization
script
Text File
 Running
With inputs for CLUSTER.BRAND.BRICK
conditions
all combos  Discounts for 4 weeks
Data Inputs

o Global Solver
CLUSTER.BRAND.BRICK  Maximized revenue
o Multi-start
 Sales Units for 4 weeks
Opening stock for EOSS
Pre-week discount
Age
 Pre-Week Sales (units)
MRP
29
 Validation

Data Aggregation
Perform Store Build Demand Models, and
In Cluster-Brand-
Clustering Brick
Optimization Model, solve for each Validation
group
Groups
Cluster 1.Brand1. Brick1
Store to
Cluster 1
Cluster1. Brand5. Brick4 Compare

Demand Models for each of 252

Optimization Model 1
Maximized
Cluster2.Brand1. Brick 1 Optimization Model 2 Revenue with
Store
Cluster 2
to Actual

combinations
Optimization Model 3
Cluster2.Brand5. Brick 4 Revenue
“ “
earned in the
Store
Cluster3.Brand1. Brick 1
to
“ “
most recent
Cluster 3
Cluster3. Brand5. Brick 4
Optimization Model 80 season

Cluster4. Brand1. Brick 1

Store to
Cluster 4 Cluster4. Brand5. Brick 4

30
 Validation- Lift in Revenue

Unique Combos Price_Elasticity Actual Discount% Optimal Discount% Actual Revenue Optimized Revenue LIFT
1.BRAND1.KURTAS 0.033 35% 10% ₹ 126,39,186 ₹ 232,75,320 ₹ 106,36,134
4.BRAND1.KURTAS 0.035 34% 10% ₹ 240,84,684 ₹ 323,45,600 ₹ 82,60,916
2.BRAND1.KURTAS 0.293 37% 35% ₹ 131,33,113 ₹ 203,04,250 ₹ 71,71,137
3.BRAND1.KURTAS 0.132 33% 18% ₹ 176,91,380 ₹ 243,58,860 ₹ 66,67,480
2.BRAND1.CHURIDARS 0.194 37% 27% ₹ 7,18,512 ₹ 11,65,663 ₹ 4,47,151
1.BRAND1.CHURIDARS 0.053 30% 10% ₹ 8,62,500 ₹ 12,65,146 ₹ 4,02,646
4.BRAND1.CHURIDARS 0.073 27% 11% ₹ 12,75,567 ₹ 16,54,887 ₹ 3,79,320

Optimized Revenue
for 2014 Festive
Winter EOSS Lift= 18% increase
in revenue
Optimized Revenue for
2014 Festive Winter
EOSS
31
 Pilot Implementation

For the upcoming Spring Summer 2015 EOSS (July – August)

Test Proposed
Stores Analytical
(50%) Model

Bangalore Random Measure

Stores Selection Performance
Using Lift

Control
As Is
Stores
Method
(50%)

32
Recommendations
Current Proposed

• Clustering of stores will help

• Currently markdown strategy &
retailer devise different
percentages are same across all
promotion strategies
the stores in India

• Scientific way of quantifying price

• No mechanism to identify price
elasticity for each of the cluster-
elasticity
brand-brick combinations

• Higher discounts are offered on

• Price inelastic brand-bricks can be
inelastic products assuming that
offered lower discounts e.g.
will increase the sales
Dupattas & Leggings

• Judgement based redistribution

• Analytical approach to
of inventory is carried during
redistribution of inventory even
short EOSS
before the start of EOSS

33
Deployment Architecture
Markdown Engine
Details of
Store, Sales &
Inventory
Clustering Tool

Demand Engine
Demand Engine

Optimization Tool
Discount
Details

34