Not logged in.
Quick Search - Contribution
Contribution Details
| Type | Master's Thesis |
| Scope | Discipline-based scholarship |
| Title | Trump Tweets and the Stock Market |
| Organization Unit | |
| Authors |
|
| Supervisors |
|
| Language |
|
| Institution | University of Zurich |
| Faculty | Faculty of Business, Economics and Informatics |
| Number of Pages | 61 |
| Date | 2021 |
| Zusammenfassung | In this thesis, we analyse whether Donald Trump’s tweets have predictive power on U.S. stock indices. While several studies show that general tweet sentiment can forecast stock markets moves, showing there is significantly less literature about the effect of Trump’s tweets. With on average eight tweets per day, Trump was by the far the most active Twitter user in the White House. Not only has he polarized the political establishment, numerous of his tweets have shaken the interest rates market (Salem, Younger and St John (2019)). Previous research on the effect of his tweets on stock indices is rare and the reported results are very dissimilar. Compared to previous research, we thus conduct a relatively comprehensive study with multiple indices analysed over a time spanning from Trump’s inauguration to the end of 2020. For our study we downloaded all of Trump’s tweets, second-by-second levels of Center for Research in Security Prices (CRSP) US capitalisation-based and sector indices, and daily levels of the Chicago Board Options Exchange Volatility Index (VIX) in our defined timeframe. We then utilize two distinct approaches: Using two sentiment tools we conduct sentiment analysis on Trump’s tweets. We utilize the well-known Loughran and McDonald library, which is based on financial reports, and often used by researchers in finance. Additionally, we use Valence Aware Dictionary for Sentiment Reasoning (VADER), which employs more advanced sentiment scoring and is tailored towards social media texts. Daily sentiment scores are then aggregated using different formulas to form sentiment time series. As in Mao, Counts and Bollen (2011) and Mao, Counts and Bollen (2015) we use an Autoregressive Distributed Lag (ARDL) model. We regress daily returns of the US capitalisation-based and sector indices on our sentiment time series and the VIX. For each independent variable we include five daily lags, but no contemporary lag. Using Trump’s tweets, we train Robustly optimized BERT pre-training approach (RoBERTa), a transformer-based machine learning model, on daily, 1 hour, 20 minute and 5 minute returns of the CRSP US Total Market and E-Mini S&P 500 futures. To the best of our knowledge, this is the first time this model has been used in the context of Trump’s twitter messages and financial markets. Our main findings are: Using our ARDL model to regress daily index returns on our sentiment, we find no or weak statistically significant results for multiple sectors. In the case of the Oil and Gas and Materials sectors, however, we find highly statistically significant sentiment coefficients. The coefficients are negative and lagged by three to four days. These results imply that a positive present-day sentiment, has a negative effect on returns in three to four days. A possible explanation for this is what De Long, Shleifer, Summers and Waldmann (1990) describe: High sentiment, that is irrationality of noise traders, may cause a stock to deviate from its fundamental value for a short period, after which it reverts to its fundamental value. This assumption is also supported by the negative coefficients of our lagged index returns. In regard to both sentiment analysis tools, our results suggest that VADER outperforms the Loughran and McDonald library. Using daily data, our RoBERTa based model suffers from a too small sample size and our results exhibit very high variance. Using intra-day data yields somewhat better accuracy than a random prediction, however our precision and recall results are distributed very unevenly. We conclude that our machine-learning setup is not adequate for our problem. |
| Export |
BibTeX
|