Decoding Sentiment in Quarterly Earnings Reports with LLMs

Research evidence excerpt

Decoding Sentiment in Quarterly Earnings Reports with LLMs

Idea

June 23, 2026 07:00 AM GMT

Morgan Stanley MUFG Securities Co., Ltd.+MJapan Quantitative Strategy | Japan Kanji Ohi, CFA

Quantitative Strategist

Decoding Sentiment in Kanji.Ohi@morganstanleymufg.comSho Nakazawa +81 3 6836-8924

Equity Strategist

Sho.Nakazawa@morganstanleymufg.com +81 3 6836-8926

Quarterly Earnings Reports with Ukyo Haraguchi, CFA

Ukyo.Haraguchi@morganstanleymufg.com +81 3 6836-8925

LLMs Masahiro Tran

Masahiro.Tran@morganstanleymufg.com +81 3 6836-8878

We used an LLM to measure sentiment in results comments and Kazuya Hayashi

forecasts from quarterly earnings reports. The analysis extracted EquityKazuya.Hayashi@morganstanleymufg.comStrategist +81 3 6836-8923

stock-selection signals hat are not fully captured by

conventional style factors. Guidance sentiment remained

effective after risk neutralization even among large caps.M

Key Takeaways

We extracted text from Operating Results in quarterly earnings releases and used

an LLM to quantify management’s view of current performance and outlook.

We defined seven categories, including Demand, Margins, Earnings Quality, and

Guidance, and assessed sentiment separately for each.

The Guidance category focuses on future information, not past earnings, and

weakly correlates with other categories, suggesting it captures unique

information.

Sentiment is positively correlated with earnings factors such as ROE and profit

growth, and this information appears to be partly reflected in share prices.

Risk-neutralized Guidance sentiment showed a distinct stock-selection effect,

even among large caps (TOPIX 500), suggesting it can complement existing

factors.

Exhibit 1: Quintile Spread Returns Using Sentiment (After Risk Neutralization)