3. Data Science @ Paychex
3
Data
Science
Team
Marketing
Sales
Information
Technology
FinanceRisk
Human
Resources
Service
4. Strategic Client Retention
2009: developed and deployed first predictive retention model
4
2009 2010 201620152014201320122011
V1 V2 V3 V4
2012: Clients with retention strategy applied we’re 80% less likely to leave
Controllable Losses
Price
Service Issue
Dissatisfaction w/ Product
Uncontrollable Losses Out of Business
Business Acquired
Business Change
Out of Business
Business Acquired
Business Change
Out of Business
Business Acquired
Business Change
𝑃 𝐶𝑙𝑖𝑒𝑛𝑡 𝐿𝑜𝑠𝑠 𝑿𝑖 ) =
1
1 + 𝑒−(β𝑿 𝑖)
5. Retention Tracking System
F
• Business 1: Annual Est. Revenue
• Business 2: Annual Est. Revenue
• …
D
• Business 3: Annual Est. Revenue
• Business 4: Annual Est. Revenue
• …
- Retention specialists proactively
call clients based on retention
score to mitigate potential issues
- Client payroll liability
- Employee counts
- ...
- Business Type
- Geographic
- Credit Info
- …
Model
Scores
Clients
Monthly
Inputs Output
- Probability
of Client
leaving
Paychex
600K clients; 8 full-time equivalent retention specialists
National Retention Team
6. Problem: Concept Drift
6
The underlying assumptions and distributions change over time
with respect to predicting losses
time
Population Stability Index (PSI) scores help us identify shifts in the
distributions of the variables in our models
What happens when the relationship with the target changes?
8. Solution: Dynamic Model Averaging*
8
Dynamic
Adapts to account for concept drift
Allows for the inclusion of new variables as they become available
Model Averaging
Use an ensemble of historical models to estimate probability of client loss
Establishing a weighting heuristic to apply each model
* Not Associated w/ Raftery et al. (2010)
9. Model Averaging
9
Tested four alternatives:
Dynamic Model Averaging
Regression Combiner
Use a weighted average base on ridge regression
,t-12 t-3*
t-12 t-11 t-4 t-3…
Average Model Scores From Each Model
Baseline Dynamic
,t-3*
Use Most Recent Model
One Year Prior Dynamic
,t-12*
Use Model from One Year Prior
The best performance
achieved by the averaged
scores of a trend model (t-3)
and seasonal model (t-12)
10. Dynamic Variables Inclusion
Access to data is continually improving. How can we adapt to this?
Just join new variables to production data set!
What if variables are discontinued?
We can identify which variables were used in our model 12 months ago
but aren’t in our current scoring data set.
Add zero-filled column (not optimal… can retrain if needed)
10
0
0
⋮
11. Model Fitting: Logistic Regression via
LASSO (aka Elastic Net)
Least Absolute Shrinkage and Selection Operator(LASSO)
Regularize logistic regression
Standardize variables (scale them to mean = 0 and sd = 1)
In traditional modeling, we choose to either include or exclude a variable in a
model (domain knowledge, AIC/BIC)
LASSO automatically includes all variables in the model, but shrinks
their impact on the model if it is not important
By shrinking the coefficients in the model, we can drive some
coefficients to zero.
11
12. Example LASSO Plot
12
• A vertical line represents a cross-
sectional view of the coefficients for
a model.
• The red line indicates a chosen
parameter/model. It is a regression
with 3 variables (3 coefficients)
Lcavol: 0.55
Weight: 0.17
Svi : 0.09
• With a very large shrinkage factor,
the sum of usual regression
coefficients are less than the
constraint, so this is exactly the
same as using a regular regression
• With a very small shrinkage factor,
the coefficients must sum to zero, so
there are no coefficients
13. Variables*
13
Continuous Variables
Trend: Mann-Kendall Trend Test Statistic
Standard Deviation
Average
Demand Index
Length of Service transformed using a Spline
Categorical Variables
# of Discount Products > # of Billed Products – {0,1}
Has Penalty in last 6 months & LOS < 2 years – {0,1}
Has Case in last 6 months & LOS < 2 years – {0,1}
Has Return in last 6 months & LOS < 2 years – {0,1}
Start Month - {1,…,12}
401k, Taxpay, Time & Attendance Indicators – {0,1}
* Final variables(w/ non Zero Coefficients) will change month by month