Skip to content

CADE-adjacent cobidders concentrate in the FL14 stratum

Intuition (plain-language)

A bid-rigging ring needs losers: someone has to submit losing bids on purpose so the auction looks competitive while the pre-chosen firm wins. Those cover bidders 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 to the top of the list (AUC ≈ 0.92). The signal is strong precisely because losing-on-purpose is the cover bidder's job description.

🟢 In São Paulo's BEC procurement platform (2009–2019), always-loser firms that bid alongside direct CADE defendants in adjudicated cartel environments concentrate disproportionately inside the FL14 stratum identified by persistent zero-win participation. The award-layer ranking achieves AUC = 0.924 [0.921, 0.926] against the cobidder target with the binary FL14 indicator, and AUC = 0.939 [0.932, 0.946] with the continuous log(1 + tenders_count) score (AN-004, AN-011).

Under the gatekeeping deployment, FL14 recovers 131 of 193 adjudicated cobidders (68%) while cutting the bid-microdata pool by 83% — the cost-of-evidence headline of §6 of the manuscript.

The concentration is the loser-side adjacency result that motivates the sequential architecture: the award layer prioritizes where the forensic stage should start, without claiming to identify cartelists.

Caveat. The cobidder set (193 firms) is defined relative to CADE- adjudicated cartel environments in 2009–2019; the loser-side scope is therefore tied to adjudication coverage rather than to a universal cartel population. The reading is 🟢 because the result survives every audit (AN-005, AN-006, AN-013, AN-014) and the disconfirming null against direct defendants (AN-007) is the predicted finding. Promotion to a generalizable claim would require independent replication on a non-BEC procurement panel.

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: \valAUCFLfirm (0.924), \valAUClogtc (0.939), \valCobidders (193), \valFL (2,735), \valCobidShareFL (7.1%).
  • Validation: backing scripts scripts/02_analysis.R, scripts/12_build_item_value.R, scripts/33_auc_direct_cade.R, scripts/34_horse_race_fl_continuous.R.