UC Berkeley researchers tackle bidder collusion in high-stakes auctions

In recent years, economists and regulators have begun examining how artificial intelligence is being used to set prices and bids across a range of markets. In some sectors, including housing, companies rely on automated pricing systems that analyze market conditions and competitors’ data to recommend prices in real time. Supporters argue such tools improve efficiency, while critics warn they can lead firms to adjust prices in parallel, creating outcomes that resemble collusion even when no explicit agreement exists. As these systems become more widespread, lawmakers and legal scholars are debating whether antitrust rules designed for human decision-making are sufficient to address coordination that emerges through algorithms and is therefore harder to detect.

That debate echoes an earlier chapter in the history of collusion. In the early 2000s, U.S. states and cities learned that municipal bond auctions meant to protect public funds had been systematically rigged. Major banks, including JPMorgan Chase and UBS, coordinated bids so preselected firms would win contracts to invest taxpayer money. Federal investigations later found the scheme cost governments hundreds of millions of dollars and led to more than $1.7 billion in penalties.

Such losses, according to new research from UC Berkeley, are not isolated incidents but reflect a structural vulnerability in many auction systems. In “Collusion-proof Auction Design using Side Information,” UC Berkeley industrial engineering and operations research (IEOR) doctoral students Sukanya Kudva, Ed Oak Dowling, and IEOR professor Anil Aswani propose a hybrid auction design that combines classical economic theory with modern collusion-detection tools. The framework preserves incentive compatibility, meaning bidders are still best off telling the truth, while improving welfare and revenue relative to standard approaches when collusion is present.

For decades, economists have relied on the Vickrey-Clarke-Groves, or VCG, auction to allocate resources efficiently. When bidders act independently, the mechanism rewards truthful bidding and maximizes total social value. Kudva, Dowling and Aswani show that these guarantees weaken once bidders coordinate.

In multi-unit auctions, such as those used to sell identical goods or licenses, the study finds that colluding bidders suppress prices by deliberately lowering their bids, even if doing so means winning fewer items. Rather than bidding aggressively to capture more units at higher prices, colluding bidders prefer to forgo some items in order to drive the clearing price down. “They would rather win fewer items at substantially lower prices,” Kudva said, noting that the strategy reduces overall spending and increases profits. The result is lower revenue for the auctioneer, often a public agency, while non-colluding bidders may also benefit from the reduced prices.

“This helps explain why collusion is so difficult to address within standard auction rules,” Kudva said. “The incentives are subtle, and the losses tend to fall on the seller.”

To address this vulnerability, the researchers propose a hybrid auction mechanism that treats colluding and non-colluding bidders differently, based on external or statistical indicators of coordination. Non-colluding bidders participate in a standard VCG auction, while colluding bidders face a posted price, a take-it-or-leave-it offer that is resistant to coordinated manipulation. Crucially, the seller must commit in advance to how many items will be allocated to each group, without adjusting that split once the auction begins. The study shows how to calculate this allocation to maximize expected social value and proves that this ex ante commitment is what preserves truthful bidding. Their hybrid approach can be implemented in different ways to select this split, some of which incorporate realized bids to produce allocations that further increase expected social value.

But incorporating bid data introduces new complications. “Determining the split using realized bids is nontrivial because some bidders, including those who are not colluding, may no longer be truthful,” co-author Ed Oak Dowling said.

In simulations across multiple valuation scenarios, the hybrid approach consistently achieved higher welfare and revenue than VCG restricted to non-colluding bidders. As the number of honest bidders increased, outcomes approached those of an ideal auction in which all participants bid truthfully.

The work sits within a growing field known as learning-augmented, or side-information, mechanism design, which incorporates external data into economic systems. Advances in artificial intelligence have made collusion detection increasingly feasible, even when legal proof remains out of reach.

“Collusion detection does not need to be perfect to be useful,” Kudva said. “If you can reduce the gains from coordination, you can make markets more resilient.”

The study also tested the robustness of the hybrid approach under imperfect collusion detection, finding that it delivered similar expected social value.

The researchers note that future work could extend the framework to auctions involving heterogeneous goods and to settings where collusion detection is incomplete. As automated auctions become more common in digital advertising, procurement and platform markets, the economic stakes of such design choices continue to grow.

The study is available on arXiv and was conducted at UC Berkeley’s Department of Industrial Engineering and Operations Research.