2. 2
“Retail is just the identification, intensification,
and presentation of merchandise.
It’s simple really. I didn’t say it was easy.”
William “Mac” McCarthy, Senior Buyer, Macy’s –
San Francisco, CA 1980
3. 3
Agenda
• Creating an atmosphere of innovation for
analytics in a non-data driven landscape
• Team transformation
• Iterative approach to model building
• Modeling customer tenure in the context of fit
solutions and returns
• Application for other marketing analytics
problems
• Q & A
4. 4
Profile
Ramp Up of Nordstrom.com
• I was one of 512 hires for the Ecommerce business in
2012
• 2-dimensional data driven business assimilating into
a 3-dimensional customer service culture
• Analytics: a work in progress
5. 5
Prologue
Transformation of Digital Analytics at Nordstrom
Sought out candidates with these specific skills:
• Customer centric preferably in Omni-channel
• Proficient in SQL to mine the database directly
• Strong statistical “firepower” and acumen
• Comfortable using R; Python an added plus (and other
open source tools)
• Innate curiosity and creativity coupled with a healthy
skepticism around data
6. 6
Does Nordstrom Fit Solutions mitigate
merchandise returns?
Methodology and Analysis
• Is the data available?
• Postulate a general class of models
• Fit model and parameter estimation by way of the
iterative process
• Validate model for reliability of conclusions
8. 8
Survival/Hazard Models were selected
over Generalize Linear Models
• A well established statistical technique that provides a
way to understand time-to-event in the context of:
• customer behavior based on tenure
• skewed distribution of data
• explanatory variables which can be isolated
for the effect on merchandise returns
• handling “censored” data, i.e. merchandise
that is never returned
9. 9
Two Views of the Same Group of Customers:
Calendar and Tenure
Customer 8
Customer 7
Customer 6
Customer 4
Customer 5
Customer 3
Customer 2
Customer 1
Customer 8
Customer 7
Customer 6
Customer 5
Customer 4
Customer 3
Customer 2
Customer 1
Calendar
Tenure
10. 10
Data Extraction and Prep
• Clickstream and demographic data from Customer
Analytics (CA) – datamart in Teradata
• orders were mined by week for a time series
comprising 16 weeks
• a fit solution flag was attached to each order
• SQL code was used to calculate the weeks to
return and identify censored data
• other explanatory variables were extracted
including, age, gender, and division
12. 12
Results and Conclusions
• Fit Solution #1 seems to have no effect on returns
• did not meet statistical muster for reliable
inferencing
13. 13
Results and Conclusions
• Fit Solution #2 has a small but significant effect on
returns
• There is a 8% less hazard or risk of returns
for those that did use Fit Solution #2 versus
those that did not
15. 15
Next Steps
• Revisit method and re-calculate using empirical
hazard estimation (discrete)
• Extend time-series longer than 16 weeks or
choose a different 16 weeks to see if original
findings hold
• Apply Cox Proportional Regression to other
Omni-channel phenomenon