Prof. David Simchi Levi, Engineering Systems Professor at MIT and Chairman of OPS Rules spoke at Bristlecone Pulse 2017 about delivering customer value through digitization, analytics and automation.
4. Online Fashion Sample Sales
Industry
• Offers extremely limited-time discounts (“flash sales”) on
designer apparel & accessories
• Emerged in mid-2000s; ~$4B industry with ~17% annual
growth in last 5 years (US)
Page 4
8. Flash Sales Operations
Merchants
purchase items
from designers
Designers
ship items to
warehouse*
Merchants decide
when to sell items
(create “event”)
During event,
customers
purchase items
Sell out
of item?
End
*Sometimes designer will hold inventory
Yes
No
Focus on first time
this happens
9. 0%
10%
20%
30%
40%
50%
60%
70%
0%‐25% 25%‐50% 50%‐75% 75%‐100% SOLD OUT
(100%)
% of Items
% Inventory Sold (Sell‐Through)
Department 1
Department 2
Department 3
Department 4
Department 5
Sell-Through Distribution
of New Products
Page 9
suggests price
may be too low
suggests price
may be too high
10. Page 10
Approach
Goal: Maximize expected revenue from new products
Demand Forecasting
Challenges:
Predicting demand for
items that have never
been sold before
Estimating lost sales
Techniques:
Clustering
Machine learning models
for regression
Price Optimization
Challenges:
Structure of demand forecast
Demand of each style is
dependent on price of
competing styles
exponential # variables
Techniques:
Novel reformulation of price
optimization problem
Creation of efficient algorithm
to solve daily
12. Regression Trees:
Illustration and Intuition
Page 12
If condition is true, move left;
otherwise, move right
Demand
prediction
Why regression trees?
– Use features to partition styles sold in past, and
only use relevant styles to predict demand
– Allow for non-monotonic price/demand relationship
13. Page 13
Approach
Goal: Maximize expected revenue from new products
Demand Forecasting
Challenges:
Predicting demand for
items that have never
been sold before
Estimating lost sales
Techniques:
Clustering
Machine learning models
for regression
Price Optimization
Challenges:
Structure of demand forecast
Demand of each style is
dependent on price of
competing styles
exponential # variables
Techniques:
Novel reformulation of price
optimization problem
Creation of efficient algorithm
to solve daily
14. Pricing Decision Support Tool
ETL
Process
Optimization
Input
Optimal Price
Recommenda‐
tions
Rue La La
Database
Reports and
Visualization
Ad hoc Reports
Query / Drill
Down Visualizer
Standard Reports
Optimizer
Database
Optimizer
Database
Rue La La Enterprise Resource Planning System
Products
Trans‐
actions
Event
Planning
Inventory‐
Constrained
Demand
Prediction
Regression
Tree
Prediction
(Rscript)
Impending
Event Data
R
Predictions
Inventory
Information
Retail Price Optimizer
LP Bound
Algorithm
LP_Solve API‐based Optimizer
Statistics
Tool ‐ R
15. Field Experiment
• Goals
– Would implementing the tool’s recommended price increases
cause a decrease in demand?
– What impact would the price increases have on revenue?
• Identified ~6,000 styles where tool recommended price increase
– Divided styles into 5 categories based on price point; for each
category…
• Raised prices on ~half of styles (treatment group)
• Did not raise prices on other ~half of styles (control group)
Page 15
18. B2W Overview
• Largest online retailer in Latin America
• Its purpose is to connect people, business,
products and services in a digital platform
• Launched in 1999 (Americanas.com) as an
extension to the physical store business
• Fastest growing Marketplace operation in the
world (+153% 2016 vs 2015)
• Competitors include Amazon, Walmart, Ponto
Frio, Extra
Page 18
19. Dynamic Pricing Approach
Page 19
Demand Curve Generation
Extract sensitivity of quantity sold
with change in price
Challenge:
Historical data has multiple
factors that influence quantity
sold
Method:
Random Forest (Bagging)
Learning via Product Clustering
productsIdentify clusters of products
having similar characteristics
Challenge:
Wide range of characteristics in
which products differ
Method:
K-Means
Optimization
Maximize revenue (margin) of a
cluster subject to a min margin
(revenue) constraint
Challenge:
Solving efficiently
Method:
Mixed Integer Linear
Programming
21. Dynamic Pricing--Operations
• Algorithm now runs twice a day;
• Typically, the first few hours of the day generate the traffic
prediction, which is an input to the Random Forest;
• No manual intervention. All prices are pushed directly to the
web-site.
Page 21
26. Competitive Benefits: Dynamic
Pricing is Also a Powerful Tool
for Market Positioning
Page 26
Unique SKUs Sold Daily in Control vs. Treatment GroupsUnique SKUs Sold Daily in Control vs. Treatment Groups
Model sold 40% more unique SKUs everyday (chart below). Will help in positioning
B2W as ‘better price every time, for everything’
Model sold 40% more unique SKUs everyday (chart below). Will help in positioning
B2W as ‘better price every time, for everything’
28. Extension to New Product Introduction
• Large footwear manufacturer
• Seasonal products: short lifecycle and low demand accuracy
• Before: Retailers ordered and returned unsold products or
require expedite shipments
• After: production and allocation is done by ML & Optimization
– Historical sales data available
– Goal: Predict demand for new products twelve months
prior to product in market
– Scope: about 2000 different SKUs
Page 28
35. Summary
• New Challenges from emerging manufacturing and
retail business
High demand uncertainty, short product life cycle and
Multiple Channels
• Combine Machine Learning and Optimization
techniques to impact bottom line
• Our proposed methods are supported by theory,
simulation results and Practice
35