Performance Cues in News Coverage Predict Presidential Approval and Election Outcomes
Abstract
Does media coverage reflect candidate quality? A fundamental measurement problem has stymied this question: existing sentiment tools assign tone to documents, not to the candidates those documents discuss, making it impossible to isolate what coverage says about individual political actors. We introduce a measure of candidate-specific coverage using stance detection from natural language inference (NLI), which evaluates whether news headlines imply that a given candidate is performing well or poorly. Applying this to nearly 900,000 newspaper headlines covering U.S. presidential candidates and sitting presidents from 1948 to 2024, we show that these performance cues predict month-to-month shifts in presidential approval, with short-run effects more than twice those of standard sentiment classifiers, and that incorporating them into election forecasts reduces prediction errors by 28–36% over an economic fundamentals-only model. These findings establish the press as a more powerful driver of democratic accountability than existing measurement approaches have revealed.
presidential approvalnews coveragemedia effectspolitical communicationpublic opinion