Alternative Data and NLP in Quant Strategies › Sentiment Analysis for Financial Text

Sentiment analysis (determining whether text expresses positive, negative, or neutral views) is the most common NLP application in quantitative finance. The intuition is straightforward: positive news should correlate with positive price movements, negative news with declines. Implementation, however, requires attention to domain-specific language and the unique characteristics of financial communication.

The simplest sentiment analysis methods use predefined word lists: positive words add to sentiment, negative words subtract. The Loughran-McDonald financial sentiment dictionary, specifically developed for financial text, has become a standard resource.

Generic sentiment dictionaries perform poorly on financial text because many words have domain-specific meanings. In everyday English, liability is negative. In finance, it is a neutral accounting term. Volatile might be negative in most contexts, but in finance it describes a factual characteristic of returns. The Loughran-McDonald dictionary addresses these issues by classifying words based on their financial context.

Converting word counts into a standardized sentiment metric requires capturing net sentiment intensity while controlling for document length. This normalization matters because we need to compare sentiment across documents of varying lengths. Without it, a 1000-word document with 10 positive words appears more positive than a 100-word document with 5 positive words, even though the shorter document has five times the sentiment density. By dividing by total words, we obtain a measure of sentiment intensity that is comparable across documents regardless of their length.

The sentiment score is computed as the difference between positive and negative word counts, normalized by document length. The score ranges from approximately -1 (all words negative) to +1 (all words positive), with 0 indicating neutral or balanced sentiment.

$$\text{Sentiment} = \frac{N_{\text{positive}} - N_{\text{negative}}}{N_{\text{total}}}$$

where:

• $N_{\text{positive}}$: count of words in the document that match the positive sentiment dictionary
• $N_{\text{negative}}$: count of words in the document that match the negative sentiment dictionary
• $N_{\text{total}}$: total word count in the document (all words, not just sentiment words)
• The score ranges from approximately -1 (all words negative) to +1 (all words positive), with 0 indicating neutral or balanced sentiment

This provides a standardized measure between -1 (maximally negative) and +1 (maximally positive). The numerator captures the net sentiment (positive minus negative words), while the denominator normalizes by document length. Without normalization, longer documents would mechanically have higher absolute scores even if their sentiment intensity was the same. The theoretical bounds of -1 and +1 are achieved only in extreme cases where every word matches the sentiment dictionaries. In practice, most documents contain many neutral words, so sentiment scores typically fall in a much narrower range around zero.