Skip to Ask the data

The line is the approval gap: how much less often a small business is approved than a large one with the same repayment record, in portfolios Beatriz studied. Figures from portfolios Beatriz studied.

measurednot measured

Fair credit, measured.

I'm Beatriz Quintela, an AI risk economist in São Paulo. I audit the machine learning models banks use to decide credit, and I measure how they treat small businesses. My career runs down this page as one line.

Where the line starts

In 1996 my mother asked a bank for a loan to buy a second oven for her padaria in Vila Mariana. She had eleven years of receipts. The answer was no, and nobody could tell her why.

I studied economics at the University of São Paulo: a BSc in 2009 and an MSc in econometrics in 2012. Nobody was measuring the gap between small and large borrowers yet, so on this chart the line is dotted until 2013.

Then I started counting

My first measurement, in 2013: small firms with clean repayment were approved 16.2 points less often than large ones. The models were not cruel. They were blind to anyone with a short credit file.

2013 to 2016
Research economist, Observatório do Crédito Pequeno, São Paulo
2016 to 2020
Model risk lead, Banco Meridiano. First audit of a machine learning credit model, 2017
2020 to now
Independent AI risk economist, auditing models for banks, cooperatives and fintechs

Papers and talks

  1. 2015Thin files: why scorecards miss small firms. Working paper
  2. 2019Reject inference for firms with no history. Working paper
  3. 2022Fair by size: a disparity test for credit models. Conference paper
  4. 2024What a fair model costs a bank. Talk, Credit Risk Forum, São Paulo

What changed when banks measured it

In 2020 the pandemic widened the gap to 12.9 points, because lenders tightened on small firms first. Since then I have audited the models directly. Three of them:

  1. 2021, a digital bank11.8 to 4.1 pp

    We rebuilt the scorecard's features around cash flow. The gap closed in nine months while the default rate moved from 2.9% to 3.0%.

  2. 2023, a cooperative network of 40 branches18,400 more firms

    Invoice data replaced missing credit history. That many more small firms were approved in one year, with losses inside the board's appetite.

  3. 2025, a payments company's lending arm1 model retired

    A gradient-boosted model nobody could explain was replaced by a monotonic one that states the reason for every decline in one sentence.

An audit takes six to ten weeks. I need a set of decisions with outcomes, firm size and sector, and I work inside the bank's own environment, so the data never leaves it.

Ask the data

Scripted answers from Beatriz's notes and the series on this page. Not Beatriz, and not financial advice. For anything else, request an audit and she replies herself.

Enter runs the cell. Shift and Enter adds a line.

Example cells

Request a model audit

Tell me what you lend and to whom. I reply within three working days with a scope and a fixed quote.

Plotted

Your request is on the line. Beatriz replies within three working days.

Nothing is sent from this page; the request would reach her inbox on a live profile.

Not financial advice. The series comes from portfolios Beatriz studied; banks are anonymised and organisations named here are invented.