Does Finance Benefit Society? A Language Embedding Approach

Abstract

We measure popular sentiment toward finance by applying a large language model to millions of books published in eight countries over hundreds of years. We extensively validate this measure both internally and externally. We document persistent differences in finance sentiment across countries despite ample time-series variation. Books written in the languages of more capitalist countries discuss finance in a more positive context. Finance sentiment is correlated with survey-based measures of financial market participation and income inequality. Finance sentiment declines one year before rather than after financial crises. Positive shocks to finance sentiment are followed by higher output and credit growth.

Keywords: sentiment, text analysis, word embedding, BERT, transfer learning, financial crises, large language model

JEL Classification: C82, E44, G52, N40, Q54