Do CEO Voices Move Markets? Vocal Stress and Analyst Tone in Earnings Calls as Predictors of High-Frequency Order Flow

Abstract

This thesis asks whether CEO vocal stress and analyst sentiment during earnings calls predict short-run high-frequency order flow, measured as signed order imbalance shifts in NYSE TAQ millisecond data. Using 40 US firms across eight quarters (Q1 2022–Q4 2023), I synchronize TAQ data with 310 earnings calls, extract 88-dimensional eGeMAPS acoustic features via OpenSMILE, and score analyst Q&A turns with FinBERT. The main regression sample comprises 282 complete cases. I additionally construct a 23-firm European ADR extension (164 complete tone+OI cases) to test cross-market generalizability. Three findings emerge. First, a composite vocal stress index (z-Jitter + z-Shimmer − z-HNR) is null across all specifications (βˆ ≈ −0.001, p > 0.80), with the 95% bootstrap confidence interval ruling out even economically small effects. Because audio was not speaker-diarized, this null applies to the full call rather than exclusively to CEO speech. Second, analyst tone is negatively associated with signed order imbalance during the call window (βˆ = −0.161, p = 0.086), but the effect is regime-dependent: significant in 2022 (βˆ = −0.420, p = 0.040) and absent in 2023 (p = 0.801), with structural break tests confirming instability (interaction p = 0.033). Third, a Benjamini-Hochberg scan of 64 acoustic features identifies one survivor: pause variability (SD of unvoiced segment length, pBH = 0.037). Cross-market validation confirms the analyst tone channel is absent in European ADR markets (βˆ = −0.049, p = 0.801), consistent with a US-specific mechanism, though timing and institutional explanations cannot be separated. Together, these results suggest that at millisecond speed, markets respond more to what analysts say than to how CEOs sound, and that even the linguistic channel activates only when macro uncertainty makes firm-specific call content informative.