Earnings Call Analysis

This is the first step toward building AI systems that read financial documents and help people make investment decisions. Mastering earnings call analysis unlocks further study of how language models are fine-tuned for finance (your library mentions FinBERT and BloombergGPT), how to backtest a strategy against historical data without fooling yourself, and how to build an assistant that answers questions about a company's prospects. The skills here—reading tone, spotting evasion, measuring deviation from prepared remarks—are the same ones used in tools that analyse every public company's earnings call automatically.

Every three months, a public company invites analysts and investors to a phone call. The CEO and CFO read a prepared statement, then answer questions. The prepared statement is carefully written, often by lawyers. The questions are not. What the executives say in the Q&A is where the real information lives—they have to respond on the spot, and their words reveal what they believe. An earnings call analyst reads the transcript and looks for: (1) what the company says it will earn in the future (guidance), (2) how confident or worried the executives sound (tone), (3) whether they dodge a question (evasion), and (4) whether they say something different from what they prepared (deviation). This is not about reading minds—it is about finding specific, measurable patterns in the words. A machine can do this at scale, reading every call from every company, and it can notice patterns that a human would miss. But the machine needs to be taught what to look for, and it needs the same discipline a human analyst would have: you can only use information that was available at the time the call happened, not the future.

Imagine you are a detective interviewing a suspect (the CEO). You have a written statement (the prepared remarks) that the suspect prepared in advance, with a lawyer. You ask questions. The suspect answers. You look for: does the answer match the statement? If the suspect says 'we are confident about next year' but then when asked about a specific risk says 'we are monitoring it closely', that is a tone shift. If you ask 'did you lose money on this project?' and the suspect says 'we are focused on the long term', that is evasion. The analogy stops working because a machine cannot read body language or hear a shaky voice—it only has the words. But the words carry signal anyway: the length of an answer, the number of hedging words ('might', 'could', 'maybe'), and the difference between the prepared text and the live answer. Unlike a detective, the machine does not get tired and can analyse every call in the market, not just one interview.

Earnings call analysis is the process of extracting predictive signal from the transcripts of a company's quarterly conference call with investors, focusing on the unscripted question-and-answer section, where guidance, tone, evasion, and deviation from prepared remarks reveal information that the market has not yet priced.

You have notes on 'AI in Finance' as a parent concept—this topic is your first concrete entry into it. You have not yet studied 'FinBERT and BloombergGPT', or 'Backtesting LLM Strategies', or 'Financial Numerical Reasoning', but this topic connects to all of them: the analysis produces signals (tone scores, evasion scores), which need to be tested against historical data (backtesting), and the language models that do the analysis need to handle financial language (FinBERT). You also have notes on 'Balance Sheets' and 'Cash Flow Management'—those are the numbers behind the call; this is the words around the numbers. The point-in-time discipline you will learn here is the same one your library emphasises for backtesting, and it is a core rule in finance. The library says 'Point-in-time discipline is essential' and 'Answers must trace to admissible evidence spans'—both will appear as we work through an example.