Weight of Evidence & Information Value Calculator
Bin a credit risk feature and get its weight of evidence per bin, the information value for the whole feature, and a read on whether it is strong enough to keep. Upload a CSV of the feature and a binary target, or paste counts you have already binned.
One numeric value per row.
1 = event (bad), 0 = non-event (good).
How weight of evidence and information value are calculated
WOE describes each bin, IV summarises the whole feature. Both are built from the split of events and non-events across the bins.
The shares are within each class: the share of all non-events in the bin over the share of all events in the bin. Positive WOE means the bin is lower risk than average.
Each bin contributes a non-negative amount, so IV grows with how differently the two classes are spread across the bins.
| IV | Reading |
|---|---|
| < 0.02 | Not predictive |
| 0.02 – 0.1 | Weak |
| 0.1 – 0.3 | Medium |
| 0.3 – 0.5 | Strong |
| > 0.5 | Suspiciously strong — check for leakage |
These bands follow Naeem Siddiqi's scorecard guidance. They are conventions for binary targets, not statistical tests.
A raw split into ten equal-frequency bins is rarely monotonic, because sampling noise nudges neighbouring bins out of order. The sample loan file on this page shows exactly that: a strong feature whose fine bins wobble.
The standard fix is coarse classing — merging adjacent bins until the weight of evidence moves in one direction. That makes the relationship with risk smooth, keeps each bin large enough to be stable, and is far easier to defend in model validation.
Once the WOE pattern is settled, each bin's WOE value replaces the raw feature in a logistic scorecard, which is what makes WOE binning so common in credit risk.
Take the binning back to your notebook
One feature is a quick check. A scorecard means computing WOE and IV across every candidate feature, iterating on the bins, and keeping it reproducible as the data refreshes. Copy the Python above into your own notebook, or let an agent with a live Jupyter kernel run the whole binning pass on data that never leaves your environment.
FAQ
Common questions about weight of evidence, information value, and binning.