Avra @ Bloomberg Línea Summit · Nov 23See the agenda

Cases · Lend

How a leading bank cut SME credit card defaults by 70% while keeping 95% approval

By adding Avra's predictions to its credit policy, the bank flagged 70% of future defaulters before approval and lifted its in-house model by 4 KS points, including for companies it had never served before.

Client
A leading bank
Stage
Lend
Measured on
The bank's own backtest
Status
Under evaluation for new approvals

The challenge

A leading Brazilian bank wanted to grow its SME credit card portfolio without letting defaults rise. It also suspected its current policy was turning away good companies, especially those with little or no history at the bank.

The in-house rating worked well for known customers. For new companies, the bank had little to go on.

What Avra built

Avra built a default risk model for the bank's SME card, trained on the bank's own credit history. For each company, the model uses only its CNPJ and signals from Avra's foundation model, trained on the graph of Brazilian companies and their relationships. This lets it predict for companies the bank has never seen before.

Avra's prediction was designed to work alongside the bank's in-house rating. Its correlation with that rating was only 0.11, so it brings information the bank did not have. The bank's own risk team then ran an independent backtest combining both.

Results

With Avra's prediction in the policy, the bank can approve almost every applicant and still cut defaults sharply.

−70%
default rate among approved companies
95%
approval rate kept at the same cutoff
70%
of future defaulters flagged before approval
+4 KS
points when combined with the in-house rating, in the bank's own backtest

Next

The bank is now evaluating Avra for new approvals and for recovering applicants rejected by the current policy.

Grow your credit portfolio with less risk.

Avra builds risk models on your own credit history and plugs them into the policy you already run. Talk to our team to see how much signal your current model is missing.

Talk to sales