Each bar = a 2% probability band. The spike near zero is not a flaw — it reflects reality. Bankruptcy is rare. When the signal appears it is concentrated and unmistakable.
Four action tiers from the model output. Each requires a different response.
Drag the slider to change the probability threshold. Numbers update live — connecting model output directly to the resource allocation decision every risk team has to make. Where you set the bar determines how many firms you flag.
Each square = one firm, colored by risk tier, ordered highest to lowest risk left-to-right, top-to-bottom. Hover any square to see that firm's exact probability. Click to open its full risk profile. The rarity of red is the point.
Filter by risk tier, search by Firm ID, sort by probability. Click any row for the full risk profile. Export any filtered view directly to CSV.
| Firm ID | Risk Probability | Tier | Action Signal |
|---|
Seven model families, two stack levels, one weighted ensemble. No single algorithm wins — the combination does.
Lessons from building a top-3 bankruptcy prediction model in a live class competition against 37 teams.
A model that always predicts "safe" is 95.3% accurate and completely useless. AUC is the only honest metric for rare-event classification. This is not an academic point — it is the difference between a tool that works and one that misleads.
Financial ratios have extreme outliers and zeros. Log transform fails on zeros. arcsinh handles the full real line — compressing extremes while preserving sign. One preprocessing choice moved the needle on AUC meaningfully.
Letting AutoGluon run 7 model families for 4 hours and stack the best produced AUC 0.9231 — better than any manually selected algorithm. The best analysts know when to step back and let the machine explore the space.
The same predict-risk → quantify-value → act-selectively logic runs inside every major financial institution today.