Menghan Wang

Menghan Wang

PhD candidate @ HKU

NLP in Finance and Accounting (III): Sentiment Measure

8 minute read

Published:

This is the third post of the series on NLP in finance and accounting. In this post, we will talk about sentiment measures. There are two main approach to construct sentiment measures.

Dictionary-based Approach

Most of the existing studies use dictionary-based approach to identify word sentiment, and the most common measure for the sentiment of a document is the number of words associated with a particular sentiment category scaled by the total number of words in the document. Sometimes, this proportion is adjusted by IDF (inverse document frequency) to account for how unusual the word is. Under dictionary-based approach, the choice (or construction) of dictionary is the main subject decision that researchers need to make. Once it is done, we are back to word counts again. There are several word dictionary commonly-used in literature.

  • Havard IV-4 positive and negative word lists: they are among the first readily available dictionary, though not created using financial text, many papers use them to measure sentiment in newspapers, earnings conference calls, press releases, etc. (Tetlock, 2007; Tetlock et al., 2008; Kothari et al., 2009; Hanley & Hoberg, 2010)
  • Diction optimism and pessimism word lists: the optimism subcategories of praise, satisfaction and inspiration are usually pooled for a positive word list, while the pessimism subcategories of blame, hardship and denial are pooled for a negative word list. Sentiment measures using Diction is more common in accounting than finance (Davis et al., 2012; Davis & Tama-Sweet, 2012).
  • Loughran & McDonald (2011) word lists: they document that almost three-fourths (73.8%) of the negative word counts according to the Harvard list are attributable to words that are typically not negative in a financial context. For instance, tax, cost, capital, board, liability, foreign, vice appear with great frequency in most of 10-Ks without revealing a negative sentiment. To address this issue, they create 6 different word lists list with respect to positive, uncertainty, litigious, strong modal, and weak modal by examining word usage in a large sample of 10-Ks during 1994-2008 (available here). Most of the latter studies are built on LM word lists (Feldman et al., 2010; Dougal et al., 2012; Mayew & Venkatachalam, 2012; Liu & McConnell, 2013; Garcia, 2013; Chen et al., 2014; Allee & DeAngelis, 2015).

It is worth noted that negative sentiment tends to convey more information than positive sentiment or net sentiment scores, as negative words are usually less unambiguous compared to positive words. Namely, it is common for management to use positive words when making a negative statement, but rarely the other way around.

Machine Learning Approach

Alternatively, some studies use supervised machine learning like Naive Bayes to train classification model, when there is no readily available dictionary. Such approach is applicable to classification task in general, not necessarily sentiment. However, it requires a manually-classified training dataset, which could be labor-intensive and suffer from subjectivity.

  • Antweiler & Frank (2004): they classify messages from Yahoo! Finance and Raging Bull to three types, buy, hold, sell, using a Naive Bayes model trained on 1000 messages. They also consider SVM (Support Vector Machine) for robustness check.
  • Li (2010): a Naive Bayes model was trained on 30000 sentences of randomly selected FLS (forward-looking statements) from MD&A section along two dimensions, tone (positive, negative, neutral, uncertain) and content (revenues, cost, profits, operations, liquidity, investing, financing, litigation, employees, regulation, accounting, other).
  • Huang et al. (2014): they classify analysts opinions into three categories, positive, negative and neutral, using Naive Bayes classifier trained on 10000 manually-classified sentences in analyst reports. The overall opinion of a report is measured by the percentage of positive sentence minus the percentage of negative sentence.
  • Buehlmaier & Whited (2018): they model the probability of being financially constrained as a function of word counts in MD&A using Naive Bayes algorithm. Three training samples are used to construct three measures respectively: (1) Factiva sample for general financial constraints, (2) equity sample using keywords from delay list and equity-focused list (Hoberg & Maksimovic, 2015) for investment delays due to issuing equity, (3) debt sample using keywords from delay list and debt-focused list (Hoberg & Maksimovic, 2015) for investment delays due to issuing debt.

The survey paper by Kearney & Liu (2014) reviews studies on textual sentiment published prior to 2013. It is generally agreed that textual sentiment has some predictive power on future earnings and stock return. They provide a detailed summary table of previous literature.

kl-sentiment-1

kl-sentiment-2

kl-sentiment-3


That’s all I have for this post. If you are also interested in applying NLP techniques to finance research, feel free to check out other pieces.

Stay tuned to this series, see you next time!


Reference

You May Also Enjoy

Synthetic DID and its implementation in Stata

9 minute read

Published:

As one can tell from the name, synthetic DID is a combination of synthetic control and DID. One main challenge of DID is the validation of parallel trend assumption, as it not always holds in the data that treated and control groups have identical trend or characteristics prior to the treatment event. Built on the method of synthetic control, synthetic DID loosens the needs for parallel assumption by constructing matched control units.

OCR Tools in Python: Azure Vision and Tesseract

7 minute read

Published:

Optical Character Recognition (OCR) is a technology that enables computers to read text from images. OCR tools have become increasingly popular due to their ability to automate data entry and extract information from scanned documents. In this blog post, I will cover the usage of two OCR tools in Python:

Accelerating Data Processing in Python with Parallelism

4 minute read

Published:

Parallel data processing is a technique used to speed up data processing tasks by dividing the data into smaller chunks and processing them simultaneously on multiple processors. It could be useful as long as do not need to be implemented sequentially. In contrast to threading, which is suitable for tasks that are I/O bounded but not tasks that are CPU bounded, multiprocessing runs on multiple cores and helps when tasks are CPU bounded. Python provides several ways to implement parallel data processing.