How Earnings Call Sentiment Moves Markets

Quarterly earnings calls contain substantially more information than the numbers they announce. The language executives use, the questions they dodge, and the tone they project are all inputs that systematic analysis can extract and quantify.

The Language of Earnings

NLP models applied to earnings call transcripts measure several dimensions: positive and negative sentiment ratios, lexical certainty — the frequency of hedging language such as "approximately," "around," and "we expect" — management tone shift versus prior quarters, and analyst question sentiment as a proxy for institutional concern. A CFO who shifts from declarative language to conditional language when discussing guidance is transmitting uncertainty that does not appear in the headline EPS figure. These signals are present in the transcript; they require systematic extraction to be actionable.

Building a Sentiment Model

The most effective models use transformer-based architectures fine-tuned on financial language. FinBERT, developed on financial news and filings, outperforms general-purpose sentiment classifiers on earnings transcripts. The model assigns a sentiment score to each sentence, weighted by speaker — CEO and CFO carry higher weight than IR leads — and position in the call, with opening remarks and guidance sections weighted above boilerplate. Backtests on S&P 500 constituents from 2015 to 2023 show a statistically significant relationship between negative sentiment shifts and next-day price declines, particularly when the shift diverges from consensus estimate beats.

Execution Timing

The edge in earnings sentiment is concentrated in the first 15 minutes after call completion — before the written transcript is widely distributed and before most systematic models have processed the audio. Firms running real-time audio-to-text pipelines therefore have a structural latency advantage over those relying on transcript vendors. For participants without that infrastructure, the second-order edge lies in post-call drift: the market tends to underreact to sentiment signals in the initial session and correct over the following two to three days, particularly when the sentiment shift is accompanied by a guidance cut or a reduction in management certainty language.