Context-Based Interpretation of Financial Information

Abstract

To what extent does the narrative context surrounding the numbers in financial statements alter the informativeness of these numbers, i.e., contextualize them? Answering this question empirically presents a methodological challenge. Leveraging recent advances in deep learning, we propose a method to uncover the value of contextual information learned from the (deep) interactions between numeric and narrative disclosures. We show that the contextualization of accounting numbers makes them substantially more informative in shaping beliefs about a firm's future, especially when numeric data is less reliable. In fact, the informational value of interactions dominates the direct informational value of the narrative context. We corroborate this finding by showing that stock markets and financial analysts incorporate the interactions between narrative and numeric information when making forecasts. We also demonstrate the value of our approach by identifying rich firm-year-specific heterogeneity in earnings persistence. We discuss a number of avenues for future research.

Keywords: Earnings, cash flows, LLMs, deep learning, narrative context, earnings persistence, neural networks, heterogeneity, MD&A

JEL Classification: G11, G12, M41