LendingClub consumer loans · loans · 2007–2015
Which loans should a lender approve?
A credit decisioning engine that decides approve or decline, then measures what that decision does to profit, losses, and fairness. The model learned from 2007–2013, every cutoff was chosen on 2014, and every result below is on 2015 loans it never saw, using what each borrower actually repaid.
more profit than approving everyone, from declining the riskiest (4% cost of funds)
more profit than ranking applicants by LendingClub’s own grade, at the same approval rate
ROC-AUC on the 2015 test set, vs for LendingClub’s grade alone
What the data says
Approval strategy
Rank applicants by
| Policy | Approved | Profit | vs approve all |
|---|
The chosen policy is the approval rate that maximized 2014 profit, applied unchanged to 2015. The hindsight row is a yardstick: it uses 2015 outcomes, so no lender could have picked it.
What this means
Vintage curves
What this means
Applicant explorer
Every decision comes with a probability of default and three plain-English reasons.
Declines are over-sampled for illustration. The chosen policy declines about of applicants. 300 real 2015 applicants: 250 spread evenly across the risk range, plus 50 from above the cutoff. Reasons are the features that raised this applicant’s risk most (SHAP values). LendingClub’s own grade and rate are never given as reasons.
Fairness
Operating point
What this means
Limitation. The data has no protected attributes (race, sex, age), and none were inferred. Income, home ownership and state are not protected classes. The four-fifths screen is a first check for groups that may overlap with protected classes. It is not a finding of discrimination, and it doesn’t replace a proper fair-lending review.
Monitoring
What this means
Method & limitations
Built to avoid the two classic traps in credit data: leakage and survivorship bias.
Data
Validation
Split by time, not at random: train on 2007–2013, choose every setting and cutoff on 2014, report once on 2015. Model inputs are application-time fields only. Automated tests keep outcome columns and geography out of every model, and LendingClub’s price out of the borrower-only model.
Models
Profit
Net profit = total payments received (including recoveries) − collection fees − funded amount, from actual outcomes. Cost of funds = rate × average outstanding balance × years outstanding. The rate is an assumption.
Limitations
- Only loans LendingClub accepted: there are no outcomes for rejected applicants, so the analysis can tighten LendingClub’s policy but not loosen it.
- Profit excludes servicing and acquisition costs and uses a straight-line average balance.
- No protected attributes, so fairness is a first screen only.