[Paper Review] Market Trend Prediction using Sentiment Analysis: Lessons Learned and Paths Forward

Andrius Mudinas, Dell Zhang+1 more|arXiv (Cornell University)|Mar 13, 2019

Stock Market Forecasting MethodsDecision Sciences33 references39 citations

TL;DR

The paper tests whether sentiment attitudes and emotions from financial news and social media Granger-cause stock price changes and whether adding sentiment features improves market trend prediction; attitudes generally do not cause price changes, emotions occasionally do for certain stocks, and sentiment features yield mixed, data-source–dependent improvements.

ABSTRACT

Financial market forecasting is one of the most attractive practical applications of sentiment analysis. In this paper, we investigate the potential of using sentiment \emph{attitudes} (positive vs negative) and also sentiment \emph{emotions} (joy, sadness, etc.) extracted from financial news or tweets to help predict stock price movements. Our extensive experiments using the \emph{Granger-causality} test have revealed that (i) in general sentiment attitudes do not seem to Granger-cause stock price changes; and (ii) while on some specific occasions sentiment emotions do seem to Granger-cause stock price changes, the exhibited pattern is not universal and must be looked at on a case by case basis. Furthermore, it has been observed that at least for certain stocks, integrating sentiment emotions as additional features into the machine learning based market trend prediction model could improve its accuracy.

Motivation & Objective

• Clarify what constitutes sentiment in finance by distinguishing sentiment attitudes and sentiment emotions.
• Test Granger-causality between sentiment signals and stock price changes across multiple data sources.
• Evaluate whether incorporating sentiment attitudes and emotions improves machine learning market trend prediction over a technical-indicator baseline.
• Assess how data source and time granularity affect the relationship between sentiment and price.
• Provide guidelines and directions for future research in sentiment-based market forecasting.

Proposed method

• Collect and preprocess three data sources: Financial Times articles, Reddit WorldNews Channel headlines, and stock-tweet data with cashtags.
• Define sentiment as attitudes (positive/negative) and emotions (eight Plutchik dimensions) and use domain-specific lexicons for extraction.
• Apply Granger-causality tests with lag of one or two days to determine directionality between sentiment and price changes.
• Develop a baseline market trend model using fifteen technical indicators and compare SVM and LSTM performance.
• Augment the baseline with sentiment attitude and emotion features and evaluate impact on predictive accuracy across stocks and currencies.

Experimental results

Research questions

• RQ1 Do market sentiments (attitudes and emotions) Granger-cause stock price changes?
• RQ2 Do stock price changes Granger-cause market sentiments?
• RQ3 Can sentiment attitudes and/or sentiments emotions improve market trend prediction beyond a technical-indicator baseline?
• RQ4 How do data sources (FT, RWNC, Twitter) and temporal granularity influence sentiment usefulness for prediction?

Key findings

• Sentiment attitudes rarely Granger-cause stock price changes across datasets.
• Stock price changes more often Granger-cause sentiment attitudes, especially with temporal sentiment modeling.
• Sentiment emotions show predictive power for some individual stocks, but results vary widely by stock and data period.
• Incorporating sentiment signals from headlines generally does not improve prediction; when improvements occur, they depend on the stock and data source (e.g., some FT-news cases).
• Tweets provide mixed results and, in short periods with a bull run, offer limited predictive value.
• Overall, the usefulness of sentiment signals is case-by-case and data-source dependent; universal predictive power is not established.