Source: Loughran & McDonald (2011) — “When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks”
I like finance papers that accidentally become papers about human behaviour. This one does that. It looks like a study about data, markets, or disclosure, but underneath it is really about how people decide what to believe.
Before modern NLP and LLMs, the main tool for measuring sentiment in financial text was word counting: you take a list of positive and negative words, count how many appear in the document, and use the ratio as a sentiment score. The Harvard Psychosociological Dictionary was the standard negative word list, widely used in finance research from the 1990s onward. This paper made a pointed observation about that dictionary: in a large sample of 10-K filings, almost three-quarters of the words it classified as negative were words that are simply not negative in financial contexts. The word “liability” is a classic example — negative in everyday language, neutral or expected in accounting. So does “foreign,” “crude,” and dozens of other words that appear routinely in financial documents without conveying negative tone. Using a negative word list calibrated on general language and applying it to financial text produces systematically distorted sentiment scores. And since those scores were then used to study how sentiment relates to stock returns, analyst reactions, and investor behaviour, the measurement error propagated into the empirical findings. Loughran and McDonald’s contribution was to develop finance-specific word lists — negative, positive, uncertainty, litigious, and others — calibrated on actual 10-K filings.
My takeaway: AI and data do not remove judgment from finance. They move judgment to the part where people decide what the signal means, whether it is trustworthy, and whether they are brave enough to act on it.