Projects / Where to set the approval cutoff for a consumer lender

Where to set the approval cutoff for a consumer lender

A scorecard cutoff approves more applicants at the same bad rate and adds $7.1M of realized NPV.

Self-built case study. Lending Club public loan data. Not client or employer work.

PythonSQLExcel

Board deckFor leadershipTechnical documentMethod, results, notebook and files

A scorecard cutoff approves 85.7% vs. 84.5% at the same bad rate and adds $7.1M

Approval rate

85.7%

vs. 84.5% today

Bad rate

12.5%

held equal

Realized NPV, 2015 loans

+$7.1M

$5.1M or more under every cost assumption

Decision: approve a test on 10% of applications, together with a pricing test.

Source: Lending Club 36-month loans, 2013–2015; analysis. Public data; illustrative results.

The 2013–2014 loans made $71.7M of realized NPV after all costs

Realized NPV of the booked loans, by issue year

$M, after cost of funds, servicing and acquisition cost

$42.9M

2013

$28.8M

2014

$71.7M

realized NPV, 2013–2014 loans

How the lender approves today

Rank applicants by risk sub-grade, A1 to G5
Approve the best sub-grades: 84.5% of applicants at a 12.5% bad rate

Source: Lending Club 36-month loans, 2013–2015; analysis. The approval rule is an assumption: the public data shows only booked loans.

The 2015 loans lost $32.9M as defaults rose to 14.9% and rates fell to 11.3%

Realized NPV of the booked loans, by issue year

$M, after all costs

$42.9M

2013

default 12.3%, rate 13.4%

$28.8M

2014

default 13.7%, rate 12.5%

–$32.9M

2015

default 14.9%, rate 11.3%

32 of 35

sub-grades lost money in 2015

So what: a better cutoff adds approvals, but only the price fixes a book that loses money in every grade.

Rate = mean interest rate. Default = share of loans that defaulted within 36 months. Source: Lending Club 36-month loans, 2013–2015; analysis.

Q: Can a new score approve more applicants at the 12.5% bad rate, and add money?

Yes: test the scorecard cutoff on 10% of applications, together with a pricing test

1. Approve

Swap to a scorecard cutoff

+$7.1M

at the same 12.5% bad rate

2. Price

Test higher rates for grades A–C

1 point

or less to break even

3. Reserve

Plan with lifetime defaults

12.8%

forecast for 2018 loans, vs. 2.7% to date

4. Recover

Rank collections by balance

45.5%

of recovery dollars in the top 20%; a model adds under 1 point

Source: Lending Club 36-month loans, 2013–2018; analysis. Public data; illustrative results.

Near the cutoff the scorecard swaps 24,426 loans for safer ones and adds $8.3M

The 24,426 loans that change sides, 2015 applicants

Actual outcomes of booked loans

Default rate

Loans the scorecard adds

23.6%

Loans the scorecard drops

25.1%

Realized NPV

Loans the scorecard adds

–$3.7M

Loans the scorecard drops

–$12.0M

+$8.3M

from the swap, at the same number of approvals

Why the result holds

Tested on 2015 loans that the model never saw
Stable population: score PSI 0.005, limit 0.1
A complex model adds only 1.5 KS points, under the 2-point rule

Source: Lending Club 36-month loans, 2013–2015; analysis. Public data; illustrative results.

At 2015 prices every grade lost money; grades A–C need 1 point or less

Rate increase for zero NPV on the 2015 loans, by grade

Interest rate points

+0.3

+0.1

+1.0

+2.7

+6.0

+12.9

+19.7

Two tools for two groups

Grades A–C: test a small rate increase
Grades D–G: the increase is too large, so use the cutoff

The estimate keeps volume and defaults fixed. A pricing test must measure both.

Cost of funds 3% a year. Source: Lending Club 36-month loans, 2013–2015; analysis.

The gain stays at $5.1M or more under every cost assumption we tested

Realized NPV of the scorecard cutoff against the current policy

$M, 2015 loans, same 12.5% bad rate

Cost of funds	Acquisition $150	$300	$450
2% a year	+8.4	+7.9	+7.4
3% a year	+7.6	+7.1 base	+6.6
5% a year	+6.1	+5.6	+5.1 lowest

+$7.1M

base case, 2015 loans

The current policy is also an assumption

If it approves 80%: the scorecard adds 1.2 points of approval
If it approves 90%: 0.7 points

Both costs need confirmation from Finance. Source: Lending Club 36-month loans, 2013–2015; analysis.

Run the scorecard cutoff on 10% of applications; stop if 12-month defaults rise 1 point

Actions

Test the scorecard cutoffon a random 10% of applications
Test a higher pricefor grades A to C: 1 point or less
Plan losses with lifetime defaults12.8% forecast, not 2.7% to date
Rank collections by balanceno recovery model

Risks and controls

Declined applicants are riskier than the booked loans	About 1,000 random approvals below the cutoff measure their default rate
Defaults rise in the test	Stop if the 12-month default rate is 1 point above the current policy
2015 prices do not pay for the risk	The pricing test for grades A to C runs with the cutoff test

Source: analysis. Public data; illustrative results.

1 / 8

Realized NPV splits into volume, revenue and cost, with one tested question on each branch

Realized NPV of the loan book2015 loans: –$32.9M
- Loans approvedVolume
  - ApplicantsFixed: the 2015 applicants
  - Approval rateThe lever of this case: the cutoff
    Can a new score approve more applicants at the same bad rate?
    Yes. 85.7% vs. 84.5%.
- Revenue for each loan
  - Loan amountFixed: set by the application
  - Interest rateFell from 13.4% to 11.3%
    Did the 2015 prices pay for the risk?
    No. 32 of 35 sub-grades lost money. Grades A to C need 1 point or less to break even, so test a higher price.
- Cost for each loan
  - Credit loss
    - Default rateThe limit: bad rate 12.5%
      Do the loans that the scorecard adds default less?
      Yes. 23.6% vs. 25.1% for the loans it drops. The swap adds $8.3M.
    - Recoveries after default
      Does a recovery model add dollars?
      No. The top 20% by balance holds 45.5% of recovery dollars. A model reaches 46.2%.
  - Operating costFunds 3% a year, servicing 1% of payments, acquisition $300 a loan
    Does the gain hold if these costs change?
    Yes. +$5.1M to +$8.4M across all cost cases.

Lever what the recommendation changes Limit what must not get worse Fixed outside the decision

Both policies run on the same 2015 applicants, and every number is an actual outcome

Learn: loans from 2013–2014Test: loans from 2015, never seen by the model

Current policy

Scorecard

Approved by both

Dropped by the scorecard24,426 loans
default rate 25.1%

Added by the scorecard24,426 loans
default rate 23.6%

Declined by both

The 2015 applicants, from low risk to high risk. Not to scale. Filled = approved. The swap adds $8.3M of realized NPV.