OPTIMISATION IN CREDIT - Credit Research Centre · Summary • Many business problems encountered...
Transcript of OPTIMISATION IN CREDIT - Credit Research Centre · Summary • Many business problems encountered...
OPTIMISATION IN CREDIT WHERE CAN OPTIMISATION HELP YOU MAKE BETTER DECISIONS
AND BOOST PROFITABILITY
CSCC XIII
Martin Benson
Jaywing
Many business problems that arise in credit
management can be tackled through
constrained optimisation
“How can I assign loan prices to applications in a way that maximises
profitability, but subject to meeting targets for lending and losses and
being compliant with advertising regulation?”
“How can I assign customers to collections strategies in a way that
maximises cash collected but subject to operational constraints?”
Strategy
Optimisation
Scenario
Decision Units Offers / Actions
Constraints Objective
Assignment
Optimisation
Algorithm
General Decision Optimisation
Strategy
Optimisation
Scenario
Loan
Applications Loan APR
Hit Lending Targets
Don’t Breach Risk Appetite
>50% Loans at Typical APR
Objective
Assignment
Strategy
Design
Loan Pricing
Strategy
Optimisation
Scenario
Credit Card
Accounts
Limit Increase
Amount
Control Incremental Risk
Manage Attrition Levels
Maximise
Profitability
Assignment
Strategy
Design
Credit Card Limit Strategy
Strategy
Optimisation
Scenario
Customers Due
To Roll Off
Discount Period
Retention Offer
Hit Retention Target
Control Risk
Manage Margins
Maximise
Profitability
Assignment
Strategy
Design
Pro-active Retention (Mortgages & CCs)
Strategy
Optimisation
Scenario
Collections
Entrants
Contact
Strategy
Minimum/Maximum Daily
Call Volumes
Hit Target Cure Rates
Maximise Cash
Collected
Assignment
Strategy
Design
Collections Entry Streaming
Benefits of optimisation
Optimisation is proven
to return higher portfolio
profitability than more traditional
techniques, across a range of
industries and use cases. It is
the current state-of-the art in
strategy design. Traditional Approach
Optimised Strategy: Up to 20% uplift
SOLVING OPTIMISATION
PROBLEMS
The Maths of Optimisation
max𝑜∈𝑂𝑓 𝑜
Subject to
𝑐𝑖 𝑜 ≥ 𝑣𝑖 𝑖 = 1, …
Maximise
some function
over all possible parameter values
subject to a set of constraints
Solving is difficult in general
max𝑜∈𝑂𝑓 𝑜
Subject to
𝑐𝑘 𝑜 ≥ 𝑣𝑘 (𝑘 = 1,… )
• General solution techniques exist:
Gradient methods
Nelder-Mead algorithm
Genetic algorithms, etc.
• But, they converge to local maxima, not global
maximum (if they converge at all).
Global Maximum
Local Maxima
𝑓 𝑥
𝑥
Tractable formulation of decision problems
max𝑜𝑖,𝑗 ∈ 0,1
𝐼×𝐽 𝑜𝑖,𝑗𝑓𝑖,𝑗𝑖∈𝐼𝑗∈𝐽
Subject to
𝑜𝑖,𝑗𝑐𝑖,𝑗,𝑘𝑖∈𝐼𝑗∈𝐽
(𝑘 = 1,…)
For each decision unit (enumerated
by I), should you make a particular
offer (enumerated by J): yes=1 no=0
The optimisation goal is a sum of
contribution values, each
corresponding to assignment of an
offer to a decision unit
Also, assume that all constraints are
linear in 𝑜𝑖,𝑗
Tractable but not trivial!
max𝑜𝑖,𝑗 ∈ 0,1
𝐼×𝐽 𝑜𝑖,𝑗𝑓𝑖,𝑗𝑖∈𝐼𝑗∈𝐽
Subject to
𝑜𝑖,𝑗𝑐𝑖,𝑗,𝑘𝑖∈𝐼𝑗∈𝐽
(𝑘 = 1,…)
• Binary Integer Program - convergence to
global optimum is
guaranteed...eventually
• Still difficult - BIPs are NP-complete
• Most statistical software packages can
solve BIPs using generic algorithms, but
may take a very long time, even for
small problems
• However, use of proprietary heuristic
algorithms can enable even large
problems to be solved quickly
PREDICTIVE MODELS
Objective
Modelling requirement
• A key requirement for optimisation is
the ability to predict objective values for
every decision unit, for every offer
Offers Decision
Units
• Could attempt to model the
relationship directly…
• …but it’s not a good idea! Too many
moving parts in the middle. Instead…
Model?
Strategy
Optimisation
Scenario
Loan
Applications Loan APR
Hit Lending Targets
Don’t Breach Risk Appetite
>50% Loans at Typical APR
Objective
Assignment
Strategy
Design
Loan Pricing example
Expected
Profit
Interest Revenue
Expected Loss
Cost of Capital
Fee Income
Other Costs
Prob. of
Take Up
PD
EAD, LGD
Voluntary Attrition
APR
Offered
Customer
Profile
Fee Amounts
Loan Amount/Term
Cost Information
Profit Per
Loan
Profitability map
Expected
Profit
Interest Revenue
Expected Loss
Cost of Capital
Fee Income
Other Costs
Prob. of
Take Up
PD
Voluntary Attrition
APR
Offered
Customer
Profile
Fee Amounts
Loan Amount/Term
Cost Information
Profit Per
Loan
EAD, LGD
Offers Decision Units Objective
Regression Model Parameter Derived
…create errors in objective estimates …generate sub-optimal decisions
Model errors drive decision errors
0%10%20%30%40%50%60%70%80%90%
100%5
.9%
6.9
%
7.9
%
8.9
%
9.9
%
10.9
%
11.9
%
12.9
%
13.9
%
14.9
%
Pro
bab
ilit
y o
f Take
-up
APR
Actual TU Rate
TU Model 1
TU Model 2
5.9
%
6.9
%
7.9
%
8.9
%
9.9
%
10.9
%
11.9
%
12.9
%
13.9
%
14.9
%
Exp
ecte
d P
rofi
t APR
TU Model 1
TU Model 2
Errors in predictive models…
Historic strategy can create modelling challenges
Historic strategy may confound data…
Ris
k B
an
d 1
Ris
k B
an
d 2
Ris
k B
an
d 3
Ris
k B
an
d 4
Ris
k B
an
d 5
…Masking true offer impacts
0%10%20%30%40%50%60%70%80%90%
100%
5.9
%
6.9
%
7.9
%
8.9
%
9.9
%
10.9
%
11.9
%
12.9
%
13.9
%
14.9
%
Pro
bab
ilit
y o
f Take
-up
APR
0%10%20%30%40%50%60%70%80%90%
100%5
.9%
6.9
%
7.9
%
8.9
%
9.9
%
10.9
%
11.9
%
12.9
%
13.9
%
14.9
%
Pro
bab
ilit
y o
f Take
-up
APR
Extrapolation potentially drives under-performance
0%10%20%30%40%50%60%70%80%90%
100%5
.9%
6.9
%
7.9
%
8.9
%
9.9
%
10.9
%
11.9
%
12.9
%
13.9
%
14.9
%
Pro
bab
ilit
y o
f Take
-up
APR
Extrapolation Over/under prediction of propensities
Overestimation of the value of some offers
Selection of suboptimal offers
Strategy under delivers, models look misaligned
How to prevent problems
Constrain optimisation
only to allow offers that
are ‘similar’ to previous
experience
Introduce small
randomised control
groups to support
future modelling
Updated strategy and
testing generates
richer data
Build models on the
data that are
available
Iterate to unlock the full potential of optimisation in a controlled manner
SUMMARY
Summary
• Many business problems encountered in credit management can be addressed through the
use of optimisation techniques
• Optimisation can deliver significant benefits over traditional approaches to strategy design.
• However, this is not straightforward, and success requires the right expertise and tools
• Careful consideration must be given to what models are required to support optimisation, and
special care taken to ensure that areas of model weaknesses cannot compromise strategy
performance
Interested in finding out more?
Read our White Paper:
www.jaywing.com/creditriskoptimisation/