Please rotate your device to landscape mode to view the charts.

The Colour of Finance Words

Journal: Journal of Financial Economics

Date: 20230301

Author: Garcia, Diego; Hu, Xiaowen; Rohrer, Maximilian

Abstract:
Our paper relies on stock price reactions to colour words, in order to provide new dictionaries of positive and negative words in a finance context. We extend the machine learning algorithm of Taddy (2013), adding a cross-validation layer to avoid over-fitting. In head-to-head comparisons, our dictionaries outperform the standard bag-of-words approach (Loughran and McDonald, 2011) when predicting stock price movements out-of-sample. By comparing their composition, word-by-word, our method refines and expands the sentiment dictionaries in the literature. The breadth of our dictionaries and their ability to disambiguate words using bigrams both help to colour finance discourse better.

Link: Google Scholar

Key Findings

Machine Learning Outperforms Traditional Dictionaries

ML dictionaries demonstrate significantly stronger predictive power for stock price movements compared to standard Loughran-McDonald (LM) dictionaries, with R² values more than double (4.6% vs 2.1%) in out-of-sample tests.

Enhanced Dictionary Coverage

ML dictionaries achieve greater coverage of financial text despite using fewer words - ML positive words cover 8.4% of earnings calls text with just 57 words, compared to 1.9% coverage from 329 LM positive words.

Robust External Validity

ML dictionaries constructed from earnings calls maintain strong predictive power when applied to 10-K filings and WSJ articles, demonstrating broader applicability across different types of financial text.

Comparative Performance in Earnings Calls Analysis

  • ML dictionaries achieve more than double the explanatory power of LM dictionaries
  • The overlap between LM & ML dictionaries shows the strongest performance
  • ML bigrams demonstrate comparable performance to ML unigrams

Dictionary Coverage Comparison

  • ML dictionaries achieve superior coverage with significantly fewer words
  • ML positive dictionary achieves 4.4x better coverage than LM with only 17% of the words
  • ML negative dictionary achieves 3.2x better coverage than LM with only 3% of the words

External Validity Across Different Text Types

  • ML dictionaries maintain superior performance across different document types
  • Performance difference is most pronounced in earnings calls analysis
  • ML dictionaries show consistent advantage over LM dictionaries in all contexts

Contribution and Implications

  • Demonstrates the potential of machine learning to improve financial text analysis beyond traditional human-coded dictionaries
  • Provides new, more efficient dictionaries that achieve better coverage and predictive power with fewer words
  • Offers a methodology for creating context-specific sentiment dictionaries that can be applied to different types of financial documents

Data Sources

  • Performance comparison chart constructed using data from Table 2, showing R² values for different dictionary approaches
  • Dictionary coverage comparison constructed using data from Table 5, comparing word counts and coverage percentages
  • External validity chart constructed using data from Tables 2, 3, and 4, showing R² values across different document types