How Model Choice Obscures Forecast Uncertainty: The Rashomon Effect and the 2026 Senate

Abstract

Election forecasts report ostensibly precise predictions that depend on modeling choices.These include the size of the incumbency advantage, the weight on state partisan lean, and how strongly national conditions move individual races.We show that the record of Senate elections between 2010 and 2024 is insufficient to adjudicate such choices.While hundreds of Senate races provide information about racelevel relationships, the role of shared national conditions must be learned primarily from comparisons across only eight election cycles.Indeed, of 120 plausible models varying these weights, 117 predict past elections similarly well.This "Rashomon set" illustrates how different assumptions can produce comparable predictive performance but distinct substantive findings.Our reference model predicts a 50.6% chance of Democratic control of the 120th Senate with 50.5 expected seats (90% interval 45-56).The 117-model Rashomon set, however, puts the chances of Democratic control anywhere between 43.2% and 57.6%.We recommend reporting this range alongside conventional predictive intervals from a single model.