In new research, Tomaso Duso, Joseph Harrington, Carl Kreuzberg, and Geza Sapi demonstrate how their screening tool can aid antitrust authorities in identifying potential collusion between firms through public communications.
When most people think of collusion, they imagine secret meetings in hotel rooms, encrypted emails, and calls on burner phones. But in modern markets, collusion does not always take place in dark hidden places. Sometimes, it unfolds out in the open—on earnings calls, at investor presentations, and even in press releases.
The airline industry is a prime example. During earnings calls, executives have conveyed the need for “capacity discipline” in the industry and praised “rational competition.” For instance, during an earnings call in 2021, a Lufthansa executive said, “We just want to avoid the price war out there. I think some capacity discipline on the part of leading carriers like us will help us create a healthy industry.”
While these words are ostensibly directed at shareholders and analysts, concerns have risen about whether their primary target may actually be competitors with the purpose of restraining competition. A recent study of airline markets in the United States found that when all legacy carriers serving a particular route referred to “capacity discipline” (and related terms) in their quarterly earnings calls, the number of seats offered in that market dropped during the next quarter. They fell by an average of 1.45% and as much as 4.2% in smaller markets.
This behavior is not limited to the airline industry. The use of investor communications as a medium for coordination has been documented in a wide range of industries, including broiler chickens, steel, and telecoms. Some competition authorities have also begun to take notice of this practice. The European Commission recently opened an antitrust case on the basis of public announcements containing collusion-facilitating content.
Despite the growing evidence of collusion using public media, enforcement based on public communication remains rare. Although there has been some private litigation, it has largely escaped the attention of competition authorities, leaving a significant enforcement gap. Regulators need to find ways to systematically detect and interpret these messages but, at the same time, large-scale monitoring of public announcements could strain scarce agency resources.
The challenge is clear: How can agencies monitor the flood of public communications without drowning in information?
Natural language processing as a detection tool
It is this challenge that our recent research seeks to solve. We have developed a systematic, workable approach to review corporate communications for collusion-friendly language. The approach uses natural language processing (NLP) to efficiently sift through vast amounts of text in order to flag statements that may facilitate coordination.
To be clear, the goal is not to convict firms on the basis of a few words. Rather, it is to provide regulators with a new method for detecting collusion—a way to spot potentially problematic communications and focus investigative resources where they will be most productive. The practical relevance of this proposed tool has been demonstrated in the field as the results of our early analysis informed the European Commission before it decided to conduct inspections at replacement tire manufacturers in 2024.
How the method works
Let us begin with the raw text upon which the method will operate. Drawing from previous research, these are examples of the type of public announcements that we want to detect:
“We have more price increases planned for this year … and, hopefully, our competitors will follow suit.“ (generic drugs)
“What is needed from the industry is a disciplined approach to bringing on supply and managing capacity.“ (steel)
“[I]f the industry could achieve a 10% reduction in capacity year-over-year by the fall that we’d be in pretty shape.“ (airlines)
Our approach breaks down these messages to their essential components and trains an NLP program to search for firms’ public announcements that contain those components. To enhance accuracy in identifying collusive content, we further require that multiple competing firms employ similar language within the same time window, and that this pattern persists over an extended period.
We begin by constructing a keyword dictionary linked to collusion-facilitating concepts through a three-step process.
First, we compile unigrams (one-word combinations) for strategy-related terms (such as price, capacity, margins) and competition or competitor-related terms (such as peer, industry, competitor). Second, we combine these with unigrams expressing intent, action, or market conditions to create two sets of bigrams (two-word combinations)—one for strategy (combining terms like “price” with “increase” or “rational”) and one for competition (combining terms like “competitor” with “will” or “should”). Third, we manually refine these lists by adding informative bigrams like “win win” and removing potentially misleading ones like “reduce price” that are more likely to be procompetitive. This process results in 108 strategy-related bigrams and 503 competition-related bigrams.
Our screening tool also requires accurate identification of actual competitors. We use detailed data from the S&P Capital IQ database to identify competitive relationships between firms. Under its earnings call functionality, Capital IQ provides a proprietary list of competitors for each company, based on information from public and proprietary sources such as annual reports, analyst reports, and filings.
What the method reveals
Our research demonstrates the practical application of a dictionary-based screening tool for detecting collusion-facilitating communication. We analyzed over 366,000 earnings calls of firms active around the world. We found that less than half contained references to the identified strategy or competition-related bigrams, with about 20% of calls consistently containing both types of sensitive language.
We illustrate the screening tool’s ability to identify concerning statements with messages from two suppliers of mechanical and digital security products, Assa Abloy and Allegion. These two competitors were flagged by our tool because, during 2018–2019, their earnings calls consistently ranked among the top 5%, and occasionally even the top 1%, in terms of strategy- and competitor-related bigrams.
An examination of their earnings calls reveal several elements consistent with coordinated pricing behavior, including price leadership, industry discipline, and reassurances. For example, in 2018, an Abloy representative stated that material cost increases are “a good thing, because, as such, everybody can increase prices,” which is consistent with collusion given that a cost increase is considered a “bad thing” in a competitive market. Allegion reinforced such content, sometimes even on the same day, by stating that the “industry is disciplined” and adding “There’s no reason that my competitors won’t move and should move. It’s important that we protect profitability.” A few days later, Abloy confirmed its position by remarking: “We definitely are not a company that buys market share with reducing the price. That would be clearly the wrong strategy.” These messages convey a common purpose to lessen price competition.
Providing some evidence of empirical validation, our automated screening approach flagged some industries recently subject to cartel investigations, such as paper products and tires. The tool successfully filtered hundreds of thousands of documents down to a manageable subset for manual review by competition authorities, making enforcement efforts more targeted and efficient. Of course, challenges remain with multi-product firms and potential false positives, emphasizing that the screening tool should complement rather than replace traditional investigative methods.
Policy implications for modernizing antitrust tools
The implications for enforcement are profound. Our NLP-based tool allows agencies to scan thousands of transcripts at scale, a task far beyond the capacity of human investigators. This capability enables regulators to prioritize by identifying industries and firms where potentially anticompetitive public dialogue is most prevalent, and to allocate resources accordingly. The tool also complements traditional evidence with a new source of communications between competitors. Just as importantly, systematic monitoring would have a deterrent effect: if firms know their words are being scrutinized, they may think twice before engaging in language that edges into coordination.
Listening more closely
That collusion occurs through firms’ public announcements presents both a challenge and an opportunity. The challenge is that collusion can masquerade as transparency; the opportunity is that public communication can be systematically monitored. Our study demonstrates how to do it. By using a trained NLP screen, competition authorities can move from anecdotal discovery to systematic monitoring, from reactive enforcement to proactive oversight.
If firms are trying to collude in the open, then regulators must learn to listen carefully. In the age of big data, words still matter and they may be the clearest signals of all.
Authors’ Disclosures: The authors report no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket‘s weekly newsletter, Special Interest, to stay up to date on ProMarket‘s coverage of the political economy and other content from the Stigler Center.
