A Longer-Term Evaluation of Information Releases by Influential Market Agents and the Semi-Strong Market Efficiency

Abstract

This paper is an evaluation of long-term cumulative abnormal returns (CAR's) based on Twitter broadcasts by highly influential market agents. We look at the information content of Elon Musk, CEO of SpaceX, Tesla and Twitter Inc. and the former US President Donald Trump. The principal objectives of this research are twofold: 1.) To assess whether markets are semi-strong form efficient and consequentially whether or not returns can be derived from strategies based on such sporadic tweet releases (abstracting from ‘news’). For this purpose, event studies are conducted on multiple companies which were targeted by Musk and Trump tweets. A control group of all Dow Jones companies with earnings releases on Twitter is utilized. We find there appears to be a “pre-post-Twitter-drift” when the release is by exceedingly influential market personalities. The cumulative abnormal returns remain significant over long durations. This indicates that markets are not entirely semi-strong form efficient regarding social media releases and that trading on such tweets may be profitable (even after factoring in varying market phases). 2.) The paper introduces a new theme: long term CAR's of market information events. The paper also notes whether the tweet was during regular market hours or after-market hours.

Keywords:

• After-hours
• CARs
• Control
• Event study
• Extended trading hours
• Investor sentiment
• Market efficiency
• Musk
• Overreaction
• Semi-strong
• Social media
• Trump
• Tweets
• Twitter

JEL Classification:

• G41
• G17
• C10
• C18
• B41

Notes

1 Such as going short in stocks with negatively themed tweets or long in stocks with positively themed tweets.

2 Theoretically, instead of specifying complex models to estimate normal returns, these returns could simply be set equal to a constant e.g. the mean over the estimation window. However, this would yield a very low $R^2$ and thus lead to inconclusive results when performing inferential tests. Refer to section 3.4 for more details.

3 The null-hypothesis being that $u_{it}=0$

4 Since $R^2=\frac{\mathit{SSE}}{\mathit{SST}}=1-\frac{\mathit{SSR}}{\mathit{SST}}=1-\frac{\mathit{var}\left(ϵ\right)}{\mathit{var}\left(y\right)},$ a high $R^2$ implies a low $\mathit{var}\left(ϵ\right)$

5 The Corrado and Zivney test statistic is given as: $\mathrm{z}=\sqrt{{\mathrm{L}}_2}\left(\frac{\overline{{\mathrm{K}}_{{\mathrm{T}}_1+1,{\mathrm{T}}_2}}-0.5}{{\mathrm{S}}_{\overline{\mathrm{K}}}}\right)\hspace{1em}\text{with}\hspace{1em}\overline{{\mathrm{K}}_{{\mathrm{T}}_1+1,{\mathrm{T}}_2}}=\frac{1}{{\mathrm{L}}_2}{\sum }_{\mathrm{t}={\mathrm{T}}_1+1}^{{\mathrm{T}}_2}\overline{{\mathrm{K}}_{\mathrm{t}}}$ For any i where Li,2 denotes the number of non-missing ARi,t during the event window and for any t, we denote the number of non-missing scaled ranks Ki,t by Nt.

6 These companies are included as 1. Fama and French factors are used and 2. The US President only has executive and legislative power over the jurisdiction. Companies listed in the US are assumed to have significant US operations, which is why they are more likely to be affected by Trump-actions than non-US-operating companies.

7 http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html

8 Numerical data output can be supplied upon request, not included due to space considerations.