AN-010: Imhof full pipeline benchmark¶
Intuition (plain-language)
The literature's workhorse screen (Imhof–Wallimann) reads seven moments off the full bid distribution — powerful, but it needs expensive bid microdata. The cheap award-layer FL screen uses only who participated and lost. Each alone lands near AUC 0.89–0.90; stacked, they reach 0.955. The economically important word is complementarity, not dominance: the two layers observe collusion at different evidentiary stages, so combining them adds genuine information rather than counting the same signal twice.
Question¶
How does the seven-feature Imhof–Wallimann bid-distribution pipeline perform on the cobidder target, and what is the increment from adding the award-layer score? The joint score is the full-observability upper bound.
Design¶
- Sample: pool of 16,779 firms with both award and bid features available in BEC 2009–2019.
- Specifications:
- Imhof CV-only: within-tender bid coefficient of variation.
- Imhof full pipeline: seven-feature bid-distribution set.
- FL14 alone: binary award-layer indicator.
- log_tc alone: continuous award-layer score.
- Joint (Imhof + FL14): stacked features.
- Joint (Imhof + log_tc): stacked with continuous.
- Outcome: AUC against the cobidder target; DeLong test for pairwise AUC differences \citep{imhof2018screening,imhof2019detecting,wallimann2023machine}.
Results¶
| Specification | AUC | 95% CI |
|---|---|---|
| Imhof CV-only | 0.585 | [0.553, 0.616] |
| Imhof full pipeline | 0.888 | [0.865, 0.911] |
| FL14 alone (binary) | 0.921 | [0.914, 0.928] |
| log_tc alone (continuous) | 0.884 | [0.860, 0.908] |
| Joint Imhof + FL14 | 0.955 | [0.943, 0.967] |
| Joint Imhof + log_tc | 0.962 | [0.954, 0.969] |
Macros: \valAUCImhofCV, \valAUCImhofFull, \valImhofFLBin,
\valImhofFLcont, \valImhofComboBin, \valImhofComboCont,
\valImhofPoolN.
Increment magnitudes
(AN-015, [AN-049 ref] script 49):
\valAUCFLvsImhofDelta = 0.035 (FL14 vs Imhof full, p = 0.014);
\valAUCImhofPlusFLDelta = 0.096 (Imhof + FL14 vs Imhof full).
Figure: AUC across the six specifications on the five-fold CV pool (N = 16,779; 193 cobidders). Imhof CV-only is chance-level (0.585); Imhof full pipeline 0.888; FL14 alone 0.921; joint (Imhof + FL14) 0.955; joint (Imhof + log_tc) 0.962. The award-layer + bid-layer complementarity adds ~0.07 over either alone.
Interpretation¶
The Imhof full pipeline is comparable to FL14 alone (0.888 vs 0.921), not dominated by it. The headline is complementarity: the joint score adds roughly 0.05–0.07 AUC over either layer individually. The two information layers operate at different evidentiary stages — award-layer triages, bid-layer evaluates — and their union is the full-observability upper bound.
The pure bid-only CV reading (0.585) is chance-level alone; what makes the Imhof pipeline informative is the inclusion of participation features. The award-layer signal is therefore not redundant; it is necessary for the bid-distribution pipeline to reach AUC 0.888 in the first place.
This is the cost-of-evidence framing of §6: AN-012 reports the operational gatekeeping numbers; the joint score is the counterfactual that gatekeeping approximates at lower cost.
Follow-ups¶
- Decomposition of the 0.05–0.07 increment by Imhof feature.
- Modal-by-modal increment (AN-016).
- Robustness to alternative bid-distribution feature sets.
