Word Power: A New Approach for Content Analysis (Digest Summary)

The authors develop a new methodology to enhance the quantitative analysis of the descriptive information in corporate financial documents. The primary advantages of their approach are that it does not require a comprehensive lexicon of positive and negative words to maintain effectiveness and it relies on observed market returns to determine the strength of various words in conveying positive and negative tone. By applying this method, the authors examine how the market responds to the tone of 10-Ks during the filing period and how the tone of IPO prospectuses corresponds with subsequent pricing levels.

The authors discuss their approach to quantitatively determine the tone of financial documents. Although there is significant documentation supporting how markets react to quantitative items, such as earnings and analyst upgrades, the authors note that relatively few studies explore how investors interpret descriptive information. Advances in statistical language processing have resulted in an increase in papers that focus on how the market reacts to descriptive information based on tone. The authors present a method that they believe improves on existing techniques by eliminating the need to have a complete lexicon and by objectively determining term weights for positive and negative words based on market reactions.

For 10-K filings, they find that both positive and negative tone are significantly related to returns for up to two weeks, which indicates that the market initially underreacts to tone but that this reaction is corrected within two weeks. From the period analyzed, the top five most positive words in the highest frequency quintile are favorable, strong, gain, efficiency, and opportunity. The top five most negative words are unresolved, unsuccessful, discourage, unauthorized, and insufficient. The authors also analyze IPO filings and find a negative relationship between the tone scores of the prospectuses and IPO underpricing. The IPO results indicate that the term weights determined by observing 10-K market reaction are also applicable when quantifying the tone of IPO prospectuses.

This research could have broad interest to the investment community. The volume of descriptive information contained in regulatory financial filings can be daunting to investors. Any approach that increases the efficiency of reviewing this information will be worthy of consideration by a number of investors. This research may also be of interest to those involved with preparing content for financial documents as an illustration of how the market reacts to tone.

The authors examine all 10-Ks filed from January 1995 through December 2010 from the US SEC’s EDGAR database. They use the firms’ first filing for the year. Market capitalization, book-to-market ratio, and turnover are used as control variables. The stock price had to be at least $3.00 on the filing date. All financial firms are excluded because a number of words that are perceived as negative for nonfinancial firms (such as risk) might not have negative connotations for financial firms. The final sample contains 45,860 filings and 7,606 unique firms. The mean market capitalization was $3.09 billion.

The IPO sample includes all IPOs during the 1995–2010 period for which there was an IPO prospectus on EDGAR as well as an IPO price, first day closing price, and all necessary data to compute control variables in CRSP. There are a total of 1,475 IPOs in the sample.

The authors use two word lists in their analysis. The first list uses the Loughran and MacDonald (LM) list from their 2011 research that identified positive and negative words. The authors then objectively determine term weights based on the market’s reaction following 10-K filings. The second list is a global lexicon of tonal words consisting of the LM list, the Harvard IV-4 Psychosociological Dictionary, and the top and bottom 200 words from the word list developed by Bradley and Lang (University of Florida Technical Report 1999). The authors examine the impact of omitting some of the relevant words by constructing partial lexicons in which 50% of the words are randomly removed. They find that the results after removing words are not statistically different from the complete lexicon results, which indicates that tone can be reliably measured even when presented with an incomplete word list.

This article represents a good addition to the body of research that examines the quantitative analysis of descriptive information contained in financial documents. Improvements in this area may have broad implications for investors. For fundamental analysts, using an algorithm to scan through documents to determine the tone could result in significant efficiencies by quickly identifying documents that warrant closer examination. For quantitative analysts, descriptive analysis may become an increasingly important data input. The authors plan to expand their approach to include such items as news reports and press releases. This area of financial research is relatively new and exciting and worth following.