How AI is changing earnings call analysis-and stock picks | Fortune

"LLMs, in contrast, interpret context and language structure, allowing them to recognize that phrases such as "growth slowed less than expected" are positive, despite negative words, according to a study by S&P Global Market Intelligence. The findings demonstrate that LLMs can extract insights from earnings call transcripts and convert them into actionable trading signals. These AI-driven signals closely match those from traditional, rules-based sentiment models, showing both methods measure the same underlying reality."

"For many years, financial sentiment analysis relied on simple word lists. Take for instance, on an earnings call, that would mean counting a CEO or CFO's positive phrases like "strong growth" and negative ones like "unexpected losses"-to assign a sentiment score. This rules-based system was transparent and easy to explain. LLMs, in contrast, interpret context and language structure, allowing them to recognize that phrases such as "growth slowed less than expected" are positive, despite negative words, according to a study by S&P Global Market Intelligence."

Large language models interpret context and language structure, recognizing positive meaning in phrases that contain negative words. LLMs can extract insights from earnings call transcripts and convert them into actionable trading signals. AI-driven signals produced by LLMs closely match signals from traditional rules-based sentiment models, indicating measurement of the same underlying information. Fine-tuned LLM strategies delivered roughly double the excess returns of traditional lexicon approaches as market inefficiencies shrink. A long-short strategy based on LLM signals achieved about 8.4% annual returns compared with about 4.2% for traditional benchmarks. LLMs separate material information from noise, improving signal precision, and flagged high-importance events produced about 6.4% excess annual returns versus 3.2% for medium-importance events.

Read at Fortune