Can data predict who pays back a loan?

We trained a model on 150,000 real credit records. Now you can see exactly what it learned — and why it thinks what it thinks.

670out of 670

From raw data to a risk score

No black boxes. Every step is transparent, inspectable, and based on established statistical methods used in real credit scoring.

01

Clean the data

Missing incomes get median-imputed. Outliers are capped. Duplicates removed. The boring stuff matters.

02

Bin the variables

OptimalBinning splits each feature into risk segments that make statistical sense — not arbitrary buckets.

03

Convert to WoE

Weight of Evidence turns each bin into a number that tells you: is this group more likely to default?

04

Score it

Logistic regression converts the pattern into a probability, then scales it to a 100–670 credit score.

What you actually see

Most ML demos spit out a number and call it a day. This one shows you the full story behind every prediction.

A credit score from 100 to 670

Higher is safer. The scale maps directly to the model's confidence.

Default probability

The raw percentage chance the model assigns to someone going 90+ days late.

Per-factor breakdown

See which variables helped the score, which hurt it, and by how much.

670Very Low Risk

Default probability

4.84%

Base score

385

Payment history+16 pts
Credit utilization-13 pts
Age-12 pts

Try it with your own numbers

Plug in a hypothetical applicant's details and see what the model predicts. No data gets stored — it's all computed in your browser session.

Built as a data science demo using the Give Me Some Credit dataset.

Not for actual lending decisions.