BIS Working Paper: Extracting Causal Response Logic from Earnings Calls with Large Language Models

Extracting causal response logic from earnings calls using large language models is a more refined text-analysis approach than simple keyword counting. BIS Working Paper 1386, published in October 2026 by the Bank for International Settlements (headquartered in Basel, Switzerland), proposes such a pipeline: the model determines whether a document states that a firm is undertaking a particular action in response to a geoeconomic pressure, and summarises the rationale for it. The authors—Chris Clayton, Antonio Coppola, Vanessa Hu, Matteo Maggiori, Jesse Schreger and Serdil Tinda—claim that the critical advantage of LLMs over classical natural language processing is their ability to extract subtle conditional relationships in prose, going beyond a simple keyword count.

Concretely, the model must identify conditional structures of the form "because of pressure X, the firm did Y in order to achieve Z." For example, when a Chinese semiconductor company states in an earnings call that "due to US export controls, we increased domestic R&D spending," the pipeline must simultaneously capture the source of pressure (US export controls), the firm's action (increased domestic R&D) and the implied logic (building substitute technology), rather than merely counting occurrences of words like "controls" or "R&D." The authors regard this ability as the core advantage distinguishing large language models from classical natural language processing.