4. In Out
In 63%
of population
12%
of population*
Out 12%
of population*
13%
of population
*Swapped populations are equal in this instance as the
client’s goal was to keep approval rate constant
ML Model
Benchmark
Model
Sources of
Credit Risk Improvement
Better In-In risk-based
pricing leads to higher
booking rates
Swapped populations
reduce credit losses
Bottom line
impact
+$9M
+$8M
+$17M
A swap set analysis highlights the business impact of using the new ML model
Results on an average $100M prime indirect auto portfolio decisioned using a custom scorecard
CASE STUDY: Auto Loans
4
63%
of population
5. In Out
In 63%
of population
12%
of population*
Out 12%
of population*
13%
of population
*Swapped populations are equal in this instance as the
client’s goal was to keep approval rate constant
ML Model
Benchmark
Model
Sources of
Credit Risk Improvement
Better In-In risk-based
pricing leads to higher
booking rates
Swapped populations
reduce credit losses
Bottom line
impact
+$9M
+$8M
+$17M
CASE STUDY: Auto Loans
5
12%
of population
12%
of population
A swap set analysis highlights the business impact of using the new ML model
Lending business: $100M prime indirect auto portfolio
6. But
(there’s no free lunch)
Lenders need to appropriately document,
validate and monitor ML models,
explain model-based decisions to consumers,
and ensure models aren’t biased.
7. 7
Credit is some of the highest-stakes AI in the business world
Credit Models Must
Comply With ECOA,
FCRA, GDPR, OCC and
Federal Reserve MRM
Guidance SR 11-7
9. Traditional Underwriting Machine Learning
Considers only a handful of variables Can consider thousands of variables
Makes simplistic, linear assumptions
Captures complex variable
interactions using real data
Uses relatively simple math Uses complex math
ML’s strengths are its greatest weaknesses when viewed
from a model risk management perspective
9
10. Used in
Guides
Documentation
ZestFinance CONFIDENTIAL
Putting an ML model into production safely involves 3
steps that must be documented
Development Validation Monitoring
• Modeling techniques
• Training and tuning
• Accuracy testing
• Risk analysis
• Adverse Action reason codes
• Model benchmarking
• Economic projection
• Disparate Impact analysis
• Data source reliability
• Input/output distributions
• Multivariate outliers
• Fair lending analysis
Creates
10
(But all this process means nothing if the math is wrong)
12. 12
ML monitors keep you out of trouble
Traditional PSI (blue line) indicates the
model results are stable over time.
ML monitors (yellow line) detected a
sharp rise in the number of
multivariate outliers in mid-2018.
The ML monitors can see the Input
data is no longer as the model
expected; traditional methods can’t.
If your models use advanced math,
so should your monitors.
13. 13
The right explainability helps you avoid denying
protected classes for dubious reasons
Car Mileage & Are you in state 47 vs Default Risk Default
Risk1.5
1.0
0.5
0.0
-0.5
0 50k 100k 150k
0.150
0.145
0.140
0.135
0.130
14. Chapter Three
With sound mathematics and the
right tools, ML can help you make
more profitable, fair and transparent
credit decisions
15. 15
The goal: To accurately explain every model-based decision
Approved applicant’s score: 0.265
This is easy to do with linear models. Not so easy with ML models.
16. 16
There are a lot of approaches out there
There are thousands of academic papers on explainable
machine learning published every year
Source: Kim, Been. “Interpretable Machine Learning: The Fuss, The Concrete, And The Questions.” Harvard University Tutorial. ICML 2017.
18. 18
ZestFinance CONFIDENTIAL
We fit a neural network a sample data set called
“moons”, which has only two variables. The
neural network achieved a high degree of
accuracy, because the modeling problem was so
simple.
Even on a simple model, popular explainability
techniques make lots of errors. ZAML doesn’t.
Correlation
Error Variance Runtime
LOCO 4.1% 0.003953 00.005s
PI 6.8% 0.009088 00.039s
LIME 7.7% 0.010679 48.521s
ZAML 0.1% 0.000253 00.258s
33
But the right mathematics can deliver accurate, consistent and fast explanations
LOCO PI
LIME ZAML
19. 19
ZestFinance CONFIDENTIAL
Game theory and advanced analysis give us the tools to
accurately assess each and every model-based decision
Lloyd Shapley Henri Lebesgue Richard Feynman
20. Once you get explainability right, you can
do amazing things with ML:
Like making the American Dream
possible for more deserving people
21. Machine learning helps lenders turn the knobs down on variables
to increase fairness, while keeping as much accuracy as possible
21
Model Variable
Contribution to
Disparate Impact
Importance in
Original Model
Importance in
More Fair Alternative
Credit Score 28% 32% 22%
Loan To Value 17% 21% 20%
Down Payment % 14% 11% 12%
Monthly Income 12% 8% 9%
Count of Bankruptcies 2% 6% 8%
Removing impact-causing variables from the model altogether,
would make the model fall apart
22. Advanced math can uncover better fairness-accuracy trade-offs
22
ZestFinance CONFIDENTIAL
Instead of omitting impact-causing variables altogether, you can reduce their
influence and increase fairness while still preserving predictive accuracy.
Original ML model
and ZAML Fair
alternatives
Old model and
“drop one”
alternativesA very small
decrease in
accuracy . . .
Original ML Model
23. Advanced math can uncover better fairness-accuracy trade-offs
23
ZestFinance CONFIDENTIAL
Instead of omitting impact-causing variables altogether, you can reduce their
influence and increase fairness while still preserving predictive accuracy.
Original ML model
and ZAML Fair
alternatives
Old model and
“drop one”
alternatives
Original ML Model
. . . can yield a big
increase in fairness.
More Fair
Alternative
24. 24
Original ML Model More Fair Alternative
Accuracy 0.905 0.902
Loss Reduction $50 million $49 million
Fairness 77% AA 84% AA
The result is a model that better balances profits and fairness
100,000 more people of color in homes per year
25. Thanks
j@zestfinance.com
1. Machine learning substantially outperforms traditional credit
risk modeling
2. Robust explainability is the key to safely applying ML
3. With sound mathematics and the right tools, ML can help you
make more profitable, fair and transparent credit decisions