Off-Exchange Migration and the Limits of AI Attribution: A Data Synthesis of Venue Composition, Execution Quality, and Enforcement in U.S. Equity Markets, 2019 to 2025

Abstract

Background. Off-exchange venues executed a record 51.8% of United States equity share volume in January 2025. A widely repeated interpretation attributes this migration to machine learning embedded in matching and routing infrastructure. This paper tests whether public data can support that attribution, and reports what they support instead.\n\nMethods. Six datasets were compiled from FINRA, SEC, and ESMA publications, from enforcement releases, and from the peer-reviewed and working-paper literature, and are provided in full with a calculation workbook in which every derived statistic is a live formula. A three-tier evidence protocol separates directly measured values from author-derived arithmetic and from interpretive attribution, and ten robustness tests are reported alongside the results they test.\n\nResults. In share terms, non-ATS off-exchange venues account for 86.3% of the 14.6 percentage-point rise in off-exchange share since 2019, a result robust to endpoint and window variation but not to dollar denomination, under which the figure falls to 54.1%. The 15.93 basis-point price-impact differential between retail and institutional execution indicates lower adverse selection in retail flow and is not, as it is often read, an execution penalty borne by retail investors. Sixteen enforcement actions totaling $330.15 million cluster in 2015 to 2018, and none alleges conduct involving a machine-learning system.\n\nConclusion. The public record establishes a compositional shift toward bilateral internalization, a segmented market, and a disclosure regime that has not been updated for model-driven venues. It does not establish that machine learning caused any of these. The governance question stands without the attribution, and a difference-in-differences design using audit-trail microdata is specified that would settle it.