AN-017: Gate D3 — continuous-only thesis¶
Intuition (plain-language)
D3 stress-tests whether the thesis depends on the FL14 cutoff at all. Throw the binary away and use only the continuous score: every specification stays significant (p < 0.001) and the modality asymmetry from D2 survives. So loser-side concentration is a property of the data, not an artifact of one threshold. The eye-popping in-sample item-level AUC (0.995) is held back for the leakage audit (AN-014), which is where the honest 0.86–0.89 number lives.
Question¶
D3 gate diagnostic: does the continuous score preserve the loser-side thesis without relying on the FL14 binary, and what is the item-level raw AUC that the leakage audit then disciplines?
Design¶
- Sample: always-losers and matched items in BEC 2009–2019; FL14 binary deliberately excluded from all specifications.
- Specifications: continuous log(1+tenders_count) only;
- AUC against cobidders;
- price coefficients across robustness specs;
- modal-asymmetry replication.
- Raw item-level: AUC reported on the full item-firm panel (subject to leakage; disciplined in AN-014).
Results¶
- All continuous-only specifications: significant at p < 0.001.
- Modal asymmetry (Pregão > Convite) survives without FL14 binary (AN-016).
- Raw item-level AUC: 0.995 (
\valAUCitemRaw); subject to leakage audit (AN-014). - Continuous firm-level AUC: 0.939 [0.932, 0.946]
(
\valAUClogtc,\valAUClogtcCI).
After leakage discipline:
| Step | Continuous AUC |
|---|---|
| Raw item-level | 0.995 |
| Out-of-fold CV at firm | 0.891 |
| Temporal holdout (firm) | 0.864 |
Interpretation¶
D3 passes. The thesis does not depend on the FL14 cutoff: a continuous-only specification preserves the cobidder concentration, the price relationship, and the modal asymmetry. The FL14 binary is the auditable operational rule, but the underlying signal is the continuous loss intensity.
The raw item-level AUC of 0.995 is reported only after the leakage audit (AN-014) is in place — the operational claim relies on the disciplined band (0.86–0.89). The continuous score is therefore robust to: (i) cutoff choice, (ii) leakage, (iii) modality split, (iv) timing discipline (AN-006).
D3 is one of four gate diagnostics (D1–D4) from 2026-04-30.