Skip to content

Raw ranking concentrates adjudication-anchored cobidders — but the lift is opportunity

Intuition (plain-language)

A bid-rigging ring needs losers: someone has to submit losing bids so the auction looks competitive while the pre-chosen firm wins. Those firms leave a footprint — they show up over and over and never win. So when we rank firms purely by how persistently they lose, the firms that bid next to convicted cartelists rise toward the top. But almost all of that lift is simply opportunity: firms that bid a lot have more chances to sit next to anyone. Once you hold opportunity fixed and ask whether losing intensity adds anything beyond raw exposure, essentially nothing survives — the within-comparison ordering is at chance. The honest finding is that the raw concentration is mechanical, anchor-agnostic co-participation exposure, not a residual cartel signal.

🟡 In São Paulo's BEC procurement platform (2009–2019), the frequent-loser ranking concentrates adjudication-anchored cobidders inside the always-loser stratum. The canonical validation label is the 651 always-loser cobidders — unique always-loser firms that share at least one BEC tender-item with a BEC-active direct CADE defendant; direct defendants are excluded and the frequent-loser flag is never used to construct the label (composition: 341 FL14 / 310 non-FL). "Cobidder" denotes adjudication-anchored exposure, never membership. The raw ranking separates them (ROC 0.761, lift@500 ≈ 5.6×), but that separation is almost entirely opportunity.

The decomposition is the finding. Ranking firms by observed contact reaches ROC ≈ 0.90, but that figure is mechanical label encoding — a cobidder is, by construction, a firm with positive contact — reported only to expose inflation, not a competing model. The honest, genuinely label-blind opportunity benchmark is ROC ≈ 0.553 (AN-004; armor pack). Once we condition on opportunity (a within-stratum comparison that holds participation fixed), the loss-intensity score retains only ROC = 0.471 — indistinguishable from chance (FL14 within-stratum 0.507), and the nested increment of the score over exposure alone is +0.010 (DeLong p = 0.013; AN-011). That marginal positive does not survive the permutation designs: matched-stratum label permutation gives p = 0.127 (ns), and FL-enrichment within matched strata gives p = 0.067 (ns). There is no robust residual signal net of opportunity.

FL14 enrichment is descriptive and real at the unconditional level — P(cobidder | FL14) = 12.5% vs 2.2% for non-FL always-losers (≈5.7×) — but this enrichment is mechanical, anchor-agnostic co-participation exposure and is not significant within matched opportunity strata (p = 0.067). The concentration is what motivates the sequential architecture: the cheap award layer concentrates forensic priority for where to start the costly bid-recovery stage — it does not identify cartelists.

Caveat. The cobidder set (651 firms) is defined relative to CADE-adjudicated cartel environments in 2009–2019; the scope is tied to adjudication coverage rather than to a universal cartel population. The reading is 🟡 because the within-stratum residual is at chance, the permutation and enrichment tests are non-significant, and the result is single-source on BEC. An anchor-agnostic armor battery confirms the verdict is not an artifact: a planted positive control (\(O_i\)) recovers within-stratum AUC 0.953 (so the test is not dead by construction) and the permutation test has power 0.97 at within-AUC 0.55, bounding any residual below ≈ 0.55. Negative controls corroborate the opportunity account (real ≈ placebo, p = 0.46; high-volume-winner null above real, p = 0.91; AN-005, AN-006, AN-013, AN-014). The null against direct defendants (AN-007) is the predicted scope boundary.

Sources.

  • Own analysis: AN-004 (baseline AUC), AN-001 (rank construction), AN-011 (continuous vs binary), AN-012 (operational metrics), AN-025 (cutoff sweep, 19-threshold inverted-U), AN-026 (subsample AUC band 0.89-0.96), AN-027 (universe × class scope matrix), AN-029 (three-classifier × out-of-time targets).
  • Cross-refs: H:cobidder-concentration; docs/results.md.
  • Macros: \valExpOnlyAUC (0.713 exposed / 0.9045 ranking by observed contact — mechanical label encoding, not a competing model), \valArmorLabelBlind (0.553 genuine label-blind opportunity), \valArmorPosControl (0.953 positive control), \valExpWithinAUC (0.471 within-stratum, ≈chance), \valExpIncrement (+0.010 nested increment, p = 0.013), \valMainCobidders (651), \valMainCobFL (341), \valMainCobNonFL (310), \valFL (2,735).
  • Validation: backing scripts scripts/02_analysis.R, scripts/12_build_item_value.R, scripts/12_audit_armor.R, scripts/12b_audit_armor_fixup.R, scripts/33_auc_direct_cade.R, scripts/34_horse_race_fl_continuous.R.