Figure 1: Credit Suisse Twitter Sentiment
How has Twitter users’ sentiment towards Credit Suisse evolved over the last few months? In Sunday evening’s official press conference about the merger of UBS and CS, FINMA chair Prof. Dr. Marlene Amstad mentioned the potential role social media might have played in amplifying negative sentiment and mistrust of Credit Suisse.
Using a natural language processing model, I analyzed 188,638 tweets mentioning Credit Suisse since March 2022 to predict their sentiment. The figure attached to this post shows the average sentiment from daily tweets, weighted by the number of tweets that day. For reference, the figure also plots the daily stock price movement of Credit Suisse.
There are several instances where negative (and positive) sentiment on Twitter correlates with daily returns. In particular, figure 1 shows a negative sentiment in the days leading up to the UBS-CS merger announcement, indicating that social media sentiment may have played a role in shaping public perception of Credit Suisse during this time. Of course, it should be noted that the observed correlation between the two variables does not necessarily imply causation.
Figure 1 also shows a negative peak in sentiment on March 15th, the time when the chairman of Saudi National Bank (SNB) announced that the SNB wouldn’t provide any further financial help to Credit Suisse.
The last week highlighted that, despite Credit Suisse meeting all capital and liquidity requirements, banking relies on trust and sentiment. This sentiment, in turn, can change quickly. The surge in negative reporting on social media might have been an amplifier and aggravated the dynamics of outflows of client funds.
Last week’s events also underline the importance for policymakers to steadily monitor social media sentiment and news reporting. In certain cases, early intervention and committed, targeted communication might help to dampen negative dynamics in the first place, thus allowing policymakers some time to address the underlying root causes in the most appropriate manner.
Figure 1 displays the sentiment predicted for 188,638 tweets that contain the keyword “Credit Suisse”, along with the daily stock price change (in %) of Credit Suisse shares. To predict the sentiment of a tweet, I use a Natural Language Processing model trained on financial texts (FinBERT, Araci (2019)).
In the plot, the sentiment measure represents the average sentiment per day of all collected tweets. The sentiment is then weighted by the number of tweets collected that day, divided by the total number of tweets. The measure is normalized (mean subtracted and divided by its standard deviation).1 Higher values correspond to a more positive sentiment, whereas lower values represent a negative sentiment. For the daily change in stock prices, missing values over the weekend are linearly interpolated to improve the readability of the chart. The correlation between the two series is 0.43.
A few notes on the negative spikes in sentiment values in figure 1 and news media reporting around the same time:
On September 22nd, 2022, news emerged that Credit Suisse might consider splitting their investment bank (Reuters (2022a)) and that they were sounding out investors about a capital hike (Reuters (2022c)).
Around October 2nd, 2022, news emerged that Credit Suisse might be looking to raise capital. At the same time, the price of Credit Default Swaps increased strongly (CNBC (2022)).
Around October 27th, 2022, Credit Suisse confirmed raising capital, announced a job cut, and informed about their new strategy (Credit Suisse (2022), Reuters (2022b)).
On November 1st, 2022, S&P downgraded Credit Suisse’s credit rating to BBB- (Wall Street Journal (2022)).
On March 15th, 2023, the chairman of Saudi National Bank announced that the bank wouldn’t boost its share of the bank (Bloomberg (2022)).
For more information, there is a growing literature that investigates the links between social media on financial markets. See, for example, Goutte (2022), Broadstock and Zhang (2019), Antweiler and Frank (2004), Sprenger et al. (2014).
For more projects and research, have a look at my webpage here.
References and Sources
The mean and standard deviation are calculated without observations of March 2023.↩︎