Studying Causality Between Investor Sentiment and Stock Returns
Summary
The document outlines ways to study whether investor sentiment influences stock returns or returns influence sentiment. It distinguishes lexicon-based sentiment measures from methods that label text polarity using stock returns. The latter can create circularity: if returns define which words are positive or negative, finding that those words predict returns may be built into the measurement.
For lexicon-based measures, the answer suggests extending dictionary matching with word embeddings while retaining supervised polarity labels. To investigate direction, it proposes comparing lagged returns with later sentiment and lagged sentiment with later returns using a Granger-style econometric approach. Autocorrelation in both series complicates interpretation. It also proposes investor flows as an intermediate link in a sequence from information and sentiment to trading decisions, flows, prices, and observed returns. These are research suggestions rather than reported empirical findings; the document does not specify a complete design for establishing causality.
Key ideas
- Sentiment measures based on return-labeled text can make return prediction findings circular.
- Lexicon-based measures offer a distinct approach to defining sentiment polarity.
- Word embeddings can extend text analysis, but polarity still requires a supervision method.
- A Granger-style comparison can examine whether past returns precede sentiment or past sentiment precedes returns.
- Investor flows may mediate the relationship between information, sentiment, and returns.
Tags
Full text
# Literature recommendations regarding sentiment in the stock market # Literature recommendations regarding sentiment in the stock market I plan to write a paper about the influence of investors sentiment on the stock market. I would like to look specifically at the question of what has an impact on what: does sentiment influence returns, or do returns influence sentiment? However, I have not yet been able to find any literature on this particular question. Can someone help me? ## Answer by lehalle (score 1) https://quant.stackexchange.com/a/68854 They are now a lot of papers about the relationship between "sentiment" and returns using Natural Language Processing, nevertheless it is difficult to recommend any of them because they do not really explain what "sentiment" is. They use two implicit definition of sentiments: - either they use a lexicon of words, and the reference study is When is a Liability not a Liability? Textual Analysis, Dictionaries, and 10-Ks by Loughran and McDonald - either they infer what are "positive" or "negative" words (or local joint distribution of words) using stock returns as a labels. And in this case Predicting Returns with Text Data by Ke, Kelly and Xiu is a good reference. But you see that the second case is not what you want because the definition of "polarity" is spurious now: if polarised means "that moves returns" they by construction polarised words are moving returns... what would you like to study? The first one is more interesting, and you can go further than a lexicon matching, you can use word embedding that is a more modern NLP way to process texts, but you will nevertheless need a lexicon to supervised your polarisations. See Do Word Embeddings Really Understand Loughran-McDonald's Polarities? by Li and L. But your question is deeper than that: you ask of the causality between returns and sentiments? You can of course have a Granger-like approach: take past returns and relate them with sentiments, and on the other hand takes past sentiments and relate them with stock returns. Auto-correlations of both will make you task difficult, but it is a possible way if you invest some time in econometrics. I would add a third suspect in the game: flows. What about a causality link like this - information is available into reports and news - investors take information into account and make investment decisions - the consequences of these decisions are flows - these buying and selling pressures move prices - investor observe returns - these returns influences their decisions (because they became information) This sequence `information/sentiment-->flows-->returns` is probably very interesting and could disentangle some of the causality questions that your study will raise.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.