Screen credit-risk variables before you model

Drop a loan-level modeling sample. Your browser runs the pre-model screen - abnormal months, missing rate, IV by organization, month-over-month PSI, a noise screen and a correlation drop - free, before you sign in, and the rows never leave this page. Then a paid run reviews the screen: what to keep, what looks like leakage, which thresholds to change.

The examples are synthetic samples made in your browser. Each has a saved model review, so you can see the whole page for free.

Dataset name

Drop a .csv file (comma, semicolon, tab or pipe; up to 300,000 rows and 600 columns)

One row per application: a date, a 0/1 bad flag, optionally an organization, then the candidate features.

Or paste CSV text

Date column

Label column (1 = bad, 0 = good)

Organization column (optional)

Out-of-sample organizations (comma separated)

Key columns (dedupe; comma separated)

Columns to ignore (comma separated)

Settings - the thresholds of each step

Values treated as missing

Minimum bads per month

Minimum rows per month

Missing-rate threshold

Overall IV threshold

Per-organization IV threshold

Low-IV organizations tolerated

PSI threshold

Unstable-month ratio

Unstable organizations tolerated

Minimum rows per month for PSI

Label permutations (noise screen)

Correlation threshold

Top features protected from the correlation drop

The screen (free, in your browser, no AI)

Steps

Step	Rule	Features flagged

Flags

Features

Feature	Type	Missing	IV	By organization	Max PSI	Null-IV margin	Result

Months and organizations

Month	Rows	Bads	Bad rate	Result

Organization	Sample	Rows	Bads	Bad rate	Months

Free: no account, no run.

About this sample (optional: product, target definition, windows)

Anything to ask (optional)

Load a sample to price the review.

The review is sent the screen's statistics - feature names, IV, PSI, missing rates, counts - and your note. Never the rows.

Raw model reply

The reply did not parse as the structured review, so it is shown as it arrived.

Your recent reviews

What this does, and what it does not

The screen follows the @github/datanalysis-credit-risk recipe: rows without a 0/1 label or a readable date are left out, duplicate keys removed, values -1, -999 and -1111 treated as missing (change them in Settings), out-of-sample organizations set aside, and months with fewer than 10 bads or 500 rows dropped. Then every step runs independently over the same sample: missing rate, IV (overall and by organization, from decision-tree bins with at most five leaves and missing as its own bin), PSI (each month against the previous one, inside each organization), a noise screen and a correlation drop. A feature survives when no step flags it. The IV and PSI arithmetic was checked against scikit-learn trees and the recipe's own PSI code on the three examples.

Two steps differ from that recipe, and the page says so wherever they matter: the noise screen compares each feature's IV with its IV under shuffled labels instead of training LightGBM, and the correlation drop keeps the higher-IV feature instead of the higher LightGBM gain. The review sees only the statistics, not your rows, and it cannot know how a feature was built - it asks you to confirm. Derived from the agent skill @github/datanalysis-credit-risk (github/awesome-copilot, MIT). The example lenders and numbers are fictional.