A post on Towards Data Science a credit‑scoring grid from a logistic‑regression model. The author starts with a dataset that keeps four variables: loan interest rate, loan‑to‑income share, prior default flag, and housing status. By assigning each category a weight, the model converts coefficients into a 0–1000 score that mirrors the FICO scale, loan_percent_income commanding a 35 % share of risk.

To turn coefficients into points, the author scales each category by the highest coefficient of its variable and multiplies by 1000. With a 10 % loan rate, a borrower scores 181.72, 0 for 25 % income share, 59.52 for no prior default, and 373.94 for home ownership, totaling 615.18. The process repeats, scores are binned into 20 vingtiles that feed risk grid.

After scoring, the author plots default density across train, test, and out‑of‑time sets, confirming defaulters cluster at low scores while non‑defaulters sit higher. The 20 vingtiles are then grouped into six risk classes—Very High, High, Medium‑High, Medium, Low, Very Low—risk, ≥30 % rate gaps, and ≥1 % client share. Stability checks show classes retain order and size over time, validating the grid.