Auction Competitive Capacity (ACC)

Replication code for "Auction Competitive Capacity: Characterization and Inference" (Durmus Karatay). The paper axiomatizes a measure of competitive pressure in auctions,

ACC = H(p) · log(1 + ρ/G*),

where H(p) is the Shannon entropy of bid shares, ρ is the top-two bid ratio, and G* is the normalized Gini dispersion of bids. This repository contains the Python package implementing the metrics, the Monte Carlo simulations, the empirical data pipelines, and the figure/table generators behind every number in the paper.

Layout

• src/acc_metrics/ — Python package: ACC components, dispersion measures, equilibrium bid functions, samplers, bootstrap/cluster-robust inference
• scripts/data/ — empirical data preparation (eBay, Operation Car Wash)
• scripts/simulations/ — Monte Carlo simulations (fixed seed 12345)
• scripts/figures/ — figure and table generation
• scripts/verify_*.py — symbolic (sympy) and numerical theorem verification
• tests/ — unit and property-based tests
• data/empirical/ — empirical datasets (see Data below)

Setup

Requires Python ≥ 3.14 and uv. Some scripts use syntax introduced in Python 3.14 (PEP 758), so older interpreters will fail to parse them.

make sync   # install dependencies
make test   # run the test suite

Usage

from acc_metrics import acc, acc_components

bids = [100, 95, 80, 60, 40]
print(f"ACC: {acc(bids):.3f}")
h, rho, gini, acc_val, evidence = acc_components(bids)

Reproducing the paper's results

All outputs are regenerated from scratch; nothing generated is committed.

make simulations  # Monte Carlo runs -> data/simulations/*.parquet
make figures      # figures + LaTeX tables -> figures/generated/
make verify       # symbolic and numerical theorem checks

Simulations use a fixed seed (12345), so regenerated figures and statistics reproduce the paper exactly. The paper's source repository compiles against a committed snapshot of this output; the reproduction check is to run make simulations && make figures here and compare the regenerated figures/generated/ files (in particular generated_stats.tex) against the versions the paper was compiled with. make figures reads the included carwash_auctions.csv directly and does not require the IEEE supplementary file; the xlsx is needed only to rebuild the CSV from scratch via scripts/data/load_carwash_data.py.

Data

• eBay auctions (data/empirical/ebay/): bid-level data for Xbox consoles, Cartier wristwatches, and Palm Pilot M515 PDAs from the dataset accompanying Jank and Shmueli, Modeling Online Auctions (Wiley-Blackwell, 2010), with contributor credits (Shanshan Wang, Sharad Borle) preserved in the directory names. scripts/data/clean_ebay_data.py consolidates the raw CSVs into ebay_auctions.parquet (included).
• Operation Car Wash (data/empirical/carwash/): 101 Brazilian procurement auctions (2002–2013) with confession-based ground-truth labels, from Signor et al., "Collusion Detection in Infrastructure Procurement: A Modified Order Statistic Method for Uncapped Auctions" (IEEE Transactions on Engineering Management). The processed carwash_auctions.csv is included; it is a derived dataset of facts (bids, estimates, computed metrics) extracted from the published supplementary material, redistributed on that basis. The raw IEEE supplementary file (supp2-3049129.xlsx) is not redistributed here; to rerun scripts/data/load_carwash_data.py from scratch, download it from the IEEE article page (DOI: 10.1109/TEM.2021.3049129) and place it in data/empirical/carwash/.

License

MIT. See LICENSE.