Skip to content

Cobidder concentration survives exposure, leakage, and timing audits

Intuition (plain-language)

The natural skeptic's reply is "of course cartel-linked firms score high — they bid a lot, and your score just counts bids." The audits stress-test exactly that: strip out raw bidding volume, remove the firms that built the score, and force the score to predict cobidders it has never seen. The concentration weakens but does not vanish (operational performance is roughly half the in-sample best case). What is left after the cheap mechanical explanations are removed is genuine signal, not an artifact of counting.

🟡 The cobidder concentration result from CADE-adjacent cobidders concentrate in the FL14 stratum survives the three audit disciplines introduced in §4.2 of the manuscript — though the operational deployment metrics are roughly half the in-sample upper bound.

  1. Exposure discipline (AN-005). Participation-volume placebos (sham FL permutation) put the permutation AUC at 0.713–0.783 — substantially below the observed 0.924. Volume-matching alone does not reproduce the concentration.

  2. Leakage discipline (AN-014). Item-level raw AUC of 0.995 drops to 0.891 [0.887, 0.894] under out-of-fold CV at firm level, then to 0.864 [0.858, 0.870] under temporal holdout. Drop ~0.10–0.13 under audit; remaining AUC above 0.85 sustains the operational claim. Direct-defendant AUC stays at ~0.51 (random) in every scenario.

  3. Timing discipline (AN-006, AN-013). Strict ex ante (score formed on 2009–2016 only): firm-level AUC drops from 0.924 (FL14 in-sample) to 0.767 [0.734, 0.800] under ex ante training. Precision@500 drops from in-sample 0.132 to temporal- holdout 0.070 (retention 53%; lift 11.5× → 6.1×).

The audit-survived numbers are what support the operational architecture of §6.

Caveat. Audit retention is partial. The 53% precision retention is honest; the AUC stays above 0.85, the precision drops by half. The manuscript reports both columns. The reading is 🟡 — informative but single-source — because the audit chain has multiple specification choices and the BEC panel is the only labeled environment.

Sources.

  • Own analysis: AN-005 (sham FL placebo, formal p < 1/2,000 / 32 σ above null), AN-006 (timing discipline), AN-013 (precision@k temporal holdout), AN-014 (leakage audit D3), AN-027 (universe scope matrix), AN-028 (within-FL standardized-diff battery, 7 dimensions), AN-029 (three-classifier × cobid_post2019 truly-OOT target), AN-030 (8.7% firm persistence — temporal holdout is fresh population).
  • Cross-refs: H:exposure-discipline; H:timing-discipline; docs/robustness.md.
  • Macros: \valAUCitemCV (0.891), \valAUCitemTemp (0.864), \valAUCStrictFirmFL (0.767), \valPrecInSFivehu (0.132), \valPrecTHFivehu (0.070), \valOpRetentionFiveHund (53%), \valLeakTautLow–\valLeakTautHigh (0.10–0.13 attenuation).
  • Validation: backing scripts scripts/25_sham_fl_permutation.R, scripts/27_strict_prospective_holdout.R, scripts/40_leakage_audit_d3.R, scripts/43_precision_at_k_audit.R.