Skip to content

Superseded numbers — canonical-target re-estimation (June 4, 2026)

This analysis note documents a historical run under the earlier validation label. On June 4, 2026 the paper adopted a reproducible, non-circular target (651 always-loser cobidders; frequent-loser flag never used in the label) and re-estimated every result. Where this page conflicts with the paper or the changelog, the paper wins.

AN-033: Imhof incremental — formal DeLong tests for complementarity

Intuition (plain-language)

How much does the cheap award layer add on top of the expensive bid-distribution screen? A formal DeLong test answers: +0.096 AUC, p ≈ 10⁻²⁶ — the two are statistically distinct signals, not the same information measured twice. On this same sample the award layer alone is at least comparable to Imhof alone (+0.035, p = 0.014) — read as complementarity and a division of labor, not as the cheap screen "beating" the expensive one. The economic implication is architectural: spend on bid microdata after a near-free award screen has already ordered forensic priority.

Question

How significant is the incremental value of the award-layer score added to the Imhof bid-distribution pipeline, by formal DeLong AUC-difference tests? The headline complementarity result from AN-010 (Imhof 0.888 vs joint 0.955) deserves a formal statistical test rather than a visual gap-reading.

Design

  • Sample: pool of firms with both award and bid features available in BEC 2009–2019: N = 11,676; N+ = 193 cobidders.
  • Baseline: Imhof full pipeline (7 features: cv_mean, cv_sd, skew_mean, kurt_mean, spread_mean, minmax_mean, second_low_mean).
  • Comparators:
  • fl_only: binary FL14 indicator alone.
  • tenders_only: continuous tenders_count alone.
  • imhof_plus_fl: Imhof full + binary FL.
  • imhof_plus_tenders: Imhof full + continuous tenders.
  • Statistic: AUC, 95% CI, delta vs Imhof baseline, DeLong paired AUC-difference p-value (same-sample test).

Two samples, one reading

This page reports the same-sample DeLong pool (N = 11,676), where Imhof full = 0.846, FL = 0.881, joint = 0.942. The manuscript headline (§6) uses the larger labeled pool (N = 16,779), where the canonical numbers are Imhof 0.888 / FL 0.921 / combined 0.962. The two samples differ in which firms have both feature sets; the qualitative conclusion is identical and the cross-sample gap is reconciled in a table footnote. The result is a complementarity / division-of-labor finding — the cheap award layer and the bid layer carry partly distinct information — not a claim that FL dominates Imhof. The joint number is a full-observability upper bound and is leakage-sensitive: it rests on in-sample firm history and falls under temporal holdout (see AN-014, AN-035).

Results

Model Features AUC 95% CI Δ vs Imhof DeLong p
imhof_full (baseline) 7 Imhof features 0.846 [0.819, 0.873] — —
fl_only is_fl 0.881 [0.871, 0.892] +0.035 0.014
tenders_only log(1+tenders) 0.877 [0.857, 0.898] +0.031 0.077
imhof_plus_fl 7 + is_fl 0.942 [0.927, 0.957] +0.096 1.2 × 10⁻²⁶
imhof_plus_tenders 7 + log(1+tenders) 0.944 [0.929, 0.958] +0.098 1.3 × 10⁻²⁵

Source: output/imhof_incremental/imhof_incremental.csv.

AN-033 Imhof incremental DeLong tests

Figure: incremental AUC gains over the Imhof full baseline. FL14 alone +0.035 (p=0.014, significant); tenders_only +0.031 (p=0.077, marginal); Imhof + FL14 +0.096 (p = 10⁻²⁶); Imhof + tenders +0.098 (p = 10⁻²⁵). The joint specifications are the cleanest within-data statistical evidence for complementarity in the paper.

The auxiliary output/auc_decomposition.csv Shapley-like decomposition yields a complementary reading on within-model contributions:

Model Features AUC Marginal contribution to full A
A (full) is_fl + imhof_cv + imhof_spread + tenders + n_bids 0.939 —
B (no FL) imhof_cv + imhof_spread + tenders + n_bids 0.936 +0.003 (FL alone, after controls)
C (FL only) is_fl 0.887 +0.052 (vs FL only baseline)
D (Imhof base) imhof_cv + imhof_spread 0.785 +0.154 (vs Imhof base)

Interpretation

Four readings:

  1. FL14 alone is at least as discriminating as Imhof full at the same-sample level (0.881 vs 0.846; delta +0.035, p = 0.014). The cheap award-layer signal matches the seven-feature bid-distribution pipeline on the same firms at far lower information cost. Following the paper, the increment is an information-cost / complementarity diagnostic, not an outperformance claim.

  2. Joint scoring is more informative than either layer alone in this same-sample comparison (Imhof + FL: 0.942 vs Imhof: 0.846, delta +0.096, p = 1.2 × 10⁻²⁶). This is the formal statistical test of H:award-bid-complementarity. At the same labeled sample, adding FL to Imhof produces an AUC gain that has effectively zero probability under the null of no information.

  3. Continuous tenders dominates binary FL14 marginally (imhof_plus_tenders 0.944 vs imhof_plus_fl 0.942). The continuous contains everything the binary does plus residual information; the joint gain is essentially the same. Consistent with the horse race in AN-011.

  4. The Shapley decomposition reveals where the marginal value concentrates. When tenders_count and n_bids are already in the model, the binary FL14 indicator adds only +0.003 marginal (Model A vs Model B). The continuous participation features carry the load; the binary is the deployable simplification. The marginal value of participation features (tenders + n_bids), starting from the Imhof base, is +0.154 — large. The marginal value of FL14, starting from Imhof + continuous, is +0.003 — essentially zero.

The complementarity claim is therefore about the continuous participation signal, not specifically the FL14 cutoff. The bid distribution carries information that participation alone does not (otherwise Model B would equal Model A, which it doesn't); but the marginal contribution of FL14 binary over continuous participation features is small.

This is exactly the "loser-side concentration is the concept; frequent losers is the operational implementation" framing locked by mr-frequent: the continuous score is the empirical primitive; FL14 is the audit-friendly rule.

Follow-ups

  • Decomposition by Imhof feature: which of the 7 Imhof features carry the marginal value over participation?
  • Same-sample DeLong on temporal-holdout subset (does the complementarity survive timing discipline?).
  • Cross-modality DeLong: Convite-only and Pregão-only DeLong p-values on the same incremental comparisons.
  • Add macros \valImhofIncFLDelong (= 1.2e-26), \valImhofIncTCDelong (= 1.3e-25), \valFLvsImhofDelong (= 0.014), and \valFLMarginalToFull (= +0.003) to the scripts/99_make_paper_values.R pipeline.