← Back to the work

Risk Analytics

Credit grade predicts default. Income barely moves it.

01 · The finding

Credit grade predicts default. Income barely moves it.

Three-way ANOVA across 32,000 applications decomposed default risk into debt-to-income and prior defaults at roughly 64:36 — a specific re-weighting for the scoring model, and grounds for dropping credit history length from it entirely.

The decision this implies: re-weight the scoring model toward debt-to-income and prior defaults, and stop treating applicant income as a primary signal.

02 · The question

I asked which application fields carry the risk, and which only look like they do.

Scope
32,000+ loan applications, 28 features.
Data
Loan application records.
Method
Three-way ANOVA with Tukey HSD post-hoc comparison.

03 · What the data showed

Credit grade separated default rates more sharply than income did.

Credit grade separates default rates more sharply than income does.

Credit grade dominates applicant income in the variance decomposition. Income is associated with default, but it adds little once grade is in the model.

Debt-to-income and prior defaults split the explained variance roughly 64:36.

Of the variance the model accounts for, debt-to-income takes roughly 64% and prior defaults roughly 36%. That ratio is the re-weighting the scoring model needs.

Loan-to-income above 30% marks a step change in default rate.

The 30% loan-to-income ratio is where the default rate steps rather than drifts, which makes it usable as a threshold in the scoring model.

04 · What I recommend

Two changes follow from the decomposition.

  1. Re-weight the scoring model toward debt-to-income and prior defaults at roughly 64:36.

  2. Drop credit history length from the scoring model.

05 · Method and caveats

ANOVA decomposes variance, and it does not establish a mechanism.

The model is specified as follows.

  • Three-way ANOVA across 32,000+ applications and 28 features.
  • Tukey HSD for post-hoc pairwise comparison between credit grades.
  • SQL for extraction, statsmodels for the analysis, Power BI for presentation.

These limits apply.

  • ANOVA gives association and variance decomposition. Nothing here establishes that credit grade causes default, and the wording throughout is deliberately 'predicts' and 'decomposed' rather than 'drives'.
  • The 64:36 split is of explained variance, not of total risk.