Joint award + bid scoring is the full-observability upper bound — conditional and case-fragile¶
Intuition (plain-language)
Suppose a regulator had everything — every bid, fully recorded — and ran the classic bid-distribution screens too. Would the cheap loser-side signal still add anything? Sometimes. Under a random cross-validation split the combined model does best, but the two layers perform comparably (award ≈ bid), and once folds are grouped by case the combined model falls below award-only on precision–recall. So the loser-side signal is not a budget substitute for bid microdata, and the complementarity is conditional and case-fragile — not a dominance result. This combined model is a best-case ceiling that assumes data a regulator usually does not have.
🟡 On the 651-cobidder target, a transparent bid-moment random-forest benchmark inspired by Imhof–Wallimann-style screens and the award-layer score perform comparably on the pooled diagnostic, and combining them only conditionally helps (AN-010). On the same-sample gatekeeping pool, the bid random forest reaches ROC 0.665, the award FL-only screen 0.665, and the combined model 0.727 — comparable, not a clean ceiling above both.
On the retained pool A (16,731 firms, 651 positives), the two layers are comparable and the combination is fragile:
- Award continuous (fixed score): ROC 0.760 / PR 0.143.
- Award FL14: ROC 0.688 / PR 0.098.
- Bid random forest (Imhof moments): ROC 0.717 / PR 0.116 under random CV, falling to 0.626 / 0.062 under case-grouped folds.
- Combined RF: ROC 0.756 / PR 0.188 under random CV, but 0.689 / 0.103 under case-grouped folds — below award-only's 0.143 PR.
So the combined model beats award on PR under random CV (+0.045) but falls below award-only under case-grouped folds. The "full observability" reading is a best-case ceiling, not a guarantee: the sequential award-then-bid rule approximates it only at lower informational cost, and the complementarity is conditional on this implemented benchmark and case-fragile (Spearman award–bid 0.544).
Caveat. The result is sample-specific (BEC 2009–2019, CADE-adjudicated cobidder labels, pool of 16,731 firms with both award and bid features available) and the increment depends on the bid-moment feature set chosen as benchmark; the paper uses a transparent bid-moment random forest inspired by \citep{imhof2018screening,imhof2019detecting,wallimann2023machine}. The reading is 🟡 because the comparison is restricted to the available adjudication target and the combined gain does not survive case-grouped folds.
Sources.
- Own analysis: AN-010 (Imhof benchmark + joint), AN-011 (horse race continuous), AN-015 (D1 harmonized same-sample), AN-033 (incremental decomposition — complementarity conditional on the implemented benchmark), AN-034 (sequential envelope demonstrates complementarity at operational level — case-fragile).
- Cross-refs: H:award-bid-complementarity; docs/results.md.
- Macros:
\valImhofFull(0.717 random / 0.626 case-grouped),\valImhofFLBin(0.688),\valImhofComboCont(0.756 random / 0.689 case-grouped),\valAwardCont(0.760),\valImhofPoolN(16,731),\valMainCobidders(651). - Validation: backing scripts
scripts/31_imhof_full_pipeline.R,scripts/49_imhof_incremental_value.R,scripts/34_horse_race_fl_continuous.R,scripts/36_gate_d1_harmonized.R.