Summary
The Phase-18 test fold was scored with two kinds of rank functions: the locked Phase-13 candidate geometries (composite scores) and a set of single-primitive baselines (each oriented risk-up by negating its z-score). The two strongest single primitives — operating cash flow / assets and interest coverage — slightly outperform the strongest A/G composite on AUROC.
Strict-label leaderboard
| Score | Kind | AUROC | AP | Lift @ 10% | Lift @ 1% |
|---|---|---|---|---|---|
| Operating cash flow / assets (risk-up) | baseline | 0.736 | 0.024 | 2.34× | 0.00× |
| Interest coverage (risk-up) | baseline | 0.726 | 0.031 | 1.93× | 3.35× |
| Altman-like proxy (Phase-19 weighted) | baseline (proxy) | 0.714 | 0.023 | 1.94× | 0.00× |
| Return on assets (risk-up) | baseline | 0.713 | 0.022 | 1.82× | 0.87× |
| Structural distress (composite) | candidate | 0.711 | 0.023 | 2.21× | 0.43× |
| G-aggregate mean (= structural by construction) | baseline | 0.711 | 0.023 | 2.21× | 0.43× |
| G+A aggregate mean | baseline | 0.688 | 0.021 | 2.08× | 0.00× |
| Joint-collapse severity (composite) | candidate | 0.677 | 0.021 | 1.90× | 0.00× |
Key reads
- Two single primitives lead the table. Operating cash flow / assets (0.736) and interest coverage (0.726) slightly outperform the composite (0.711) by 1.5 – 2.5 pp.
- The composite equals its G-aggregate by construction. "Structural distress" and "G-aggregate mean" are the same thing — both AUROC 0.711.
- Adding the A side dilutes rank discrimination. The seven-primitive G+A aggregate mean (0.688) is worse than the G-only mean (0.711). On a wide non-distress population, A introduces sectoral noise; A’s value is in joint-collapse severity, not in pooled rank.
- Altman-like proxy ties the composite. A transparent weighted primitive composite (NOT the actual Altman Z-score; we don’t have market cap, retained earnings, EBIT, or sales/assets reliably from the SEC companyfacts feed for the full universe) gives AUROC 0.714 — essentially indistinguishable from structural distress.
Implication for the corporate story
The corporate A/G composite is not the strongest raw discriminator on this dataset. Its defensible value is twofold:
- Robustness. Single primitives can be unstable when one input is missing or degenerate (e.g., interest coverage for cash-rich firms; OCF/assets for asset-light firms). The composite degrades more gracefully.
- Regime separation. The composite framework distinguishes broad structural distress (G-only), joint-collapse severity (G + A), and divergence / masking (A − G) — interpretive structure that single primitives don’t supply on their own.
Wording the corporate story can use
- "The corporate G aggregate is informationally consistent with the strongest single G primitives."
- "The composite is more robust than any single primitive to missing-data and sectoral noise."
- "Operating cash flow / assets and interest coverage are the strongest single-primitive screens for Chapter 11 / Item 1.03 distress in this dataset."
- "Adding the A layer to the G aggregate dilutes rank discrimination on a wide population; the A side carries joint-collapse severity (
co_collapse) rather than pooled discrimination."
Wording the corporate story cannot use
- "The A/G architecture beats simple financial primitives." (It does not, here.)
- "The composite is more than its parts." (Not in raw discrimination.)
- "Only multi-channel scoring captures distress." (Single primitives capture most of it.)
Source basis: this is a public summary of
CORPORATE_PHASE19_BASELINE_COMPARISON.md in
the project repository. Numbers come from
processed/phase19_baseline_comparison.csv.
Back to Validation · Evidence · Documents index.