Media Text Sentiment and Prediction of Stock Returns, a paper co-authored by our school’s Professor Jiang Fuwei, PhD student Meng Lingchao of the School of Economics of Peking University (an undergraduate of our school’s Outstanding Academic Talent Cultivation Program for 2019), and Assistant Professor Tang Guohao (a PhD having graduated from our university) from the College of Finance and Statistics of Hunan University, has been officially accepted by China Economic Quarterly.
The paper touches upon the predictive relations between the sentiment of financial media texts and asset prices in the stock market. It builds a brand-new Chinese financial text sentiment dictionary (which can be downloaded at https://github.com/MengLingchao/Chinese_financial_sentiment_dictionary) by employing the Loughran-MacDonald Dictionary (2011), word2vec algorithm and artificial screening, and then uses it to extract media text sentiments about stock markets. The paper finds that media text sentiments can effectively measure investor sentiment changes in China’s stock market. And this indicator demonstrates a remarkable prediction ability for stock returns both within and outside the samples, which is stronger than the oft-used macroeconomic indicators or historical averages. Moreover, the media text sentiment indicator also shows a marked ability to predict some macroeconomic indicators.
The paper’s key innovation lies in the development of a Chinese financial sentiment dictionary. Text analysis in the field of finance mostly employs sentiment dictionaries, but the existing Chinese sentiment dictionaries are all general-purpose, with limited applicability in the financial context. Hence, a widely-accepted, specialized and open-source financial sentiment dictionary is still not available now. This paper has ultimately established a Chinese financial sentiment dictionary by converting the English LM dictionary into the Chinese version, screening out words suitable for financial contexts from existing Chinese general-purpose sentiment dictionaries, and extracting sentiment words and expressions from text corpora through the word2vec algorithm. Then it compares the dictionary with existing ones through various methods, and proves the new dictionary is more applicable and suitable.
Next it sets up China’s stock market media text sentiment index using the financial sentiment dictionary it develops and text analysis technologies. The text sentiment index and text sentiment analysis technologies have the following four strengths: (1) Compared with publicly available financial market transaction data, text sentiment information has strong complementarity, but is usually ignored in previous asset pricing studies. Therefore, it is probable that studies on the impact of sentiment on asset pricing using text analysis methods may lead to new conclusions. (2) Given the huge scale of text data, gathering sentiment information from such massive data sources helps to reduce previous sentiment measurement errors. (3) Text sentiment falls into the category of direct sentiment measurement. (4) Given the high frequency of text data updates, it is possible to build higher-frequency sentiment indices (on a daily or minute basis).
Then the paper tests the sentiment index’s ability to predict stock markets. And the result shows, the text sentiment indicator can notably make forward predictions about stock market returns within and outside the samples, with a prediction ability no weaker than common economic indicators. Moreover, text sentiment indicator also plays a role in asset allocation. All these findings show that the text sentiment indicator has relatively great application value in investment practices.
Finally, the paper looks into the source and mechanism of text sentiment’s prediction ability. It first finds that media text sentiment can apparently affect investors’ expectations for the macro economy, while investors will adjust their degree of participation in the financial market based on such expectations to prompt corresponding responses from market returns. Then the paper also finds that the prediction ability of text sentiment cannot be properly explained based on the risk compensation theory, which indicates the way this indicator works on stocks conforms better with the irrational communication channels under the noise trader model developed by De Long, Shleifer, Summers and Waldmann (1990).
