Every quarter, the CEOs and CFOs of publicly listed companies host earnings calls — structured discussions with equity analysts reviewing financial results, business developments, and forward guidance. These calls are carefully scripted in their prepared remarks sections, but the question-and-answer segments that follow reveal a great deal about management confidence, business trajectory, and potential concerns that numbers alone cannot convey. Natural language processing is unlocking this information at scale.
The Information Content of Earnings Calls
Academic research has repeatedly demonstrated that textual features of earnings calls contain information beyond what is conveyed by the headline financial metrics. Studies have found that linguistic cues such as hedging frequency, first-person pronoun usage, forward-looking statement density, and negative word counts predict future stock returns, earnings surprises, and analyst forecast revisions — even after controlling for the quantitative content of the results.
The intuition is straightforward: managers who are confident about the business tend to speak more directly, with fewer qualifiers and more specific forward-looking claims. Managers facing undisclosed difficulties may use more vague language, avoid certain questions, or become defensive under analyst questioning. NLP systems can detect these signals systematically across thousands of calls simultaneously — an information advantage impossible to achieve through manual analysis.
Data Collection and Processing
Earnings call transcripts are typically available through financial data providers such as Bloomberg, FactSet, and Refinitiv, or can be scraped directly from company investor relations pages and services like Seeking Alpha. Raw transcripts require preprocessing: separating the prepared remarks from the Q&A section, attributing statements to specific speakers (management vs. analysts), and segmenting the text into analysable units.
Speaker attribution is particularly important. The predictive content of analyst questions differs fundamentally from that of management responses. Questions that analysts choose to ask — and those they avoid — can signal areas of concern or uncertainty. Management responses to difficult questions, detected by matching question-answer pairs, are particularly information-rich segments.
Feature Extraction Approaches
The simplest approach to earnings call analysis uses dictionary-based sentiment scoring with financial-domain lexicons such as Loughran-McDonald. The proportion of positive, negative, and uncertain words in the transcript is computed, and these proportions are used as features in predictive models. This approach is fast, transparent, and surprisingly effective as a baseline.
More sophisticated approaches use topic modelling to identify recurring themes across calls — extracting, for instance, how frequently management discusses supply chain issues, demand trends, pricing power, or specific geographic markets. Latent Dirichlet Allocation (LDA) and neural topic models identify these themes in an unsupervised manner, without pre-specifying the topics of interest.
Transformer-based models achieve the best performance on downstream prediction tasks. FinBERT and similar domain-adapted models capture nuanced contextual sentiment that bag-of-words and dictionary approaches miss. Fine-tuning on labelled pairs of earnings call segments and subsequent return outcomes allows these models to learn the specific language patterns most predictive in a financial context.
Audio Features and Vocal Analysis
Beyond the words themselves, the audio of earnings calls contains additional signals. Paralinguistic features — speaking rate, pitch variation, voice tremor, pauses, and hesitations — have been shown to correlate with management uncertainty and subsequent stock performance. Machine learning models applied to raw audio spectrogram data or to hand-engineered acoustic features extract these signals systematically.
The combination of textual and acoustic features outperforms either modality alone — a finding consistent with the broader principle that multimodal signals contain complementary information. Building robust multimodal earnings call analysis systems remains an active area of research, with practical challenges including audio quality variation and the computational cost of processing large audio datasets.
Regulatory and Ethical Considerations
Using NLP to trade on earnings call information raises important regulatory questions. The use of public information — transcripts of calls that anyone can attend — is generally permissible. However, systems that process audio before transcripts are publicly available, or that extract information not available to other market participants through materials selectively disclosed to favoured investors, can cross into legal grey areas around Regulation FD (Fair Disclosure) in the US and equivalent regulations in the UK and EU. Legal review of data acquisition methods and trading protocols is essential before deployment.