1 Introduction

In 2023, it became evident that banks were once again in turmoil after the 2008–09 financial crisis. For instance, in March 2023, three US-based banks, namely Silicon Valley Bank (SVB), Silvergate Bank, and Signature Bank failed due to bank runs. The US banking crisis also transmitted quickly to Europe, with Swiss authorities organising a related takeover of Credit Suisse by UBS on 12 June 2023. On March 9, 2023, a day before the collapse of SVB, the combined market value of four major banks, namely Bank of America, Citigroup, JPMorgan Chase, and Wells Fargo, experienced a collective plunge of USD 52 billion (Choi 2023). Notably, according to Shaw (2023), the potential bankruptcy of SVB could have far-reaching consequences for the landscape of technology startups in both the US and internationally, particularly in the UK. Additionally, SVB was the 16th largest bank in the US with assets of approximately USD 200 billion, and its collapse stands as the second most significant bank failure in US history since the 2008 Global Financial Crisis, according to The Economist (Economist 2023). In Guardian, Yerushalmy (2023) also reported that the implosion of SVB was the second most substantial bank collapse in the history of the US.

Bloomberg Law, Weinberger (2023) documented the recent SVB collapse as the first “Twitter bank run”, prompting banks and policymakers to urgently devise strategies to prevent potential future bank runs. Speculations about a bank failure can span several months or even years before culminating in a collapse. In the specific case of SVB, however, it happened in a matter of hours due to social media sentiment. Furthermore, the most significant event, the 2008 Global Financial Crisis, unfolded over 8 months, whereas the collapse of SVB occurred in only 2 days based on social media tweets. Over the past decade, academic literature has indicated that the role of social media-driven speculation is pivotal in triggering the collapse of the banking system and stock markets (Chen et al. 2022; Ge et al. 2020; Guan et al. 2021; Hamraoui & Boubaker 2022; Lee et al. 2022; Liang et al. 2020; Onsare 2017; Seok et al. 2021; Shah & Albaity 2022; Shah et al. 2023; Wang et al. 2022), among others.

In response to this recent US banking crisis, a bourgeoning literature has explored the contagion effect of the SVB collapse. In this stream of literature, research scholars have linked SVB collapse with global financial markets (Aharon et al. 2023; Pandey et al. 2023; Yadav et al. 2023; Yousaf et al. 2023), financial markets, banks, and non-financial firms of G7 countries as well as Brazil, China, India, and South Africa (Akhtaruzzaman et al. 2023), US equity market sectors (Yousaf & Goodell 2023), Euro area banking sector (Perdichizzi & Reghezza 2023), social media-based bank runs and global market contagion (Dosumu et al. 2023; Khan & Anupam 2023), major global financial assets (Azmi et al. 2023), and cryptocurrency market (Capalbo & Galati 2023).

Intriguingly, the main reason for the collapse of SVB was social media-based speculation; however, earlier studies have used classical models (mostly the event study method) to explore the contagion effect of the SVB run. One exceptional study by Khan and Anupam (2023) conducted a social media-based sentiment analysis of SVB implosion through the application of Natural Language Processing (NLP). Another recent study by Bales and Burghof (2024), noted that a higher level of public attention and Twitter sentiment led to a significant decrease in SVB returns, which, in turn, resulted in SVB implosion. In this paper, we build upon the work of Khan and Anupam (2023) and Bales and Burghof (2024) to explore the impact of social media-based investor sentiment on the SVB collapse. Furthermore, we take a step further to examine the spillover effect of the social media-based investor sentiment and SVB run on the bank stock indices from twelve countries where Global Systemically Important Banks (G-SIBs) operate, using advanced machine and deep learning algorithms. The selection of G-SIBs bank stock indices is motivated by their significance in relation to potential systemic shocks that these G-SIBs may generate and subsequently transmit to the global financial market.

The main findings of this paper, based on advanced machine and deep learning algorithms, reveal that negative social media-based investors’ sentiment regarding SVB started in early March, which, in turn, played an important role in the implosion of SVB on the 10th of March 2023. The results also show a significant plummet in US bank stock index returns in early March, along with the persistence of negative sentiment and its contagion shock transmitted to the G-SIBs bank stock indices.

The contribution of this study is twofold. First, this paper explores the role that social media-driven speculation plays in triggering the collapse of SVB and its spillover effect on the G-SIBs bank stock indices. The spillover effect of any bank run on G-SIBs is important because G-SIBs have the potential to be too-big-to-fail and can cause systemic risk due to their large degree of interdependence within the global financial system. In this regard, Elliott et al. (2014) and Egger et al. (2023) documented that adverse shocks from and on G-SIBs can have devastating aggregate economic effects worldwide.

Second, this paper contributes to the methodological front. Specifically, it employs advanced machine learning algorithms to examine the influence of social media-based investor sentiment on the SVB collapse and its spillover effects on the bank stock indices of twelve countries. Machine learning algorithms possess robust estimation features as compared to the classical models. Machine learning and deep learning models can capture the contextual meaning underlying a given sentence or social media post. This is usually done by accounting for preceding and succeeding words in each sentence. On the other hand, lexicon-based models are often unable to capture such contextual nuances. The usage of a language is very dynamic and evolves with time. Machine learning models exhibit a greater adaptability in learning new ways of showing one’s emotions on social media. Training of machine learning sentiment analysis models can be performed using a variety of datasets not limited to only social media data. Hence, the model performance is often very robust (Kolchyna et al. 2015; Nguyen et al. 2018).

The rest of the paper is organised as follows. Section 2 describes the data and sample of the study. Section 3 presents the machine learning models. Section 4 discusses the results, and finally, Sect. 5 provides the concluding remarks.

2 Data description and sample of the study

This study combined two types of datasets for empirical investigation. The first dataset was created from social media data using the following procedure. In the first step, data related to social media was extracted from Twitter. To scrap the tweets data, different hashtags were applied namely #SiliconValleyBank, #SVB, #SVBCollapse, and #SiliconValleyBankCollapse. The scrapping from Twitter was conducted in April 2023 using Python, snscrape application, with a final count of 279,779 tweets from 1st March 2023 to 4th April 2023.

In the second step, three real-world datasets compiled by Kotzias et al. (2015) were used for training of the models including Amazon, IMDB, and Yelp. The composition of reviews from Amazon, Yelp, and IMBD was 36%, 36% and 28%, respectively. The total number of observations in this merged dataset was 2,748. Notably, the dataset was meticulously curated to ensure an equal representation of sentiments, being evenly split between negative and positive reviews. This careful distribution mitigates any potential biases or skewness, thereby enhancing the reliability and fairness of the model trained using this dataset.

In the third step, testing data was prepared. To this end, manual tagging was performed for Twitter data (gathered in the first step) to separate data regarding financial tweets regarding SVB. To ensure a comprehensive understanding of sentiment distribution, this dataset consisted of an equal number of tweets across three categories: 60 positive, 60 negative, and 60 neutral. Each tweet was carefully annotated and assigned a sentiment label positive, negative, or neutral by one of the authors, having expertise in the financial markets. This provided robustness in the accuracy and relevancy of the sentiment classification. This well-balanced and expertly annotated dataset served as the foundation for evaluating the effectiveness of our developed models.

The second dataset was based on daily prices of bank stock indices from twelve countries where thirty of the world’s largest Global Systemically Important Banks (G-SIBs) are located. The bank stock indices’ raw data in the form of daily closing prices were taken from the websites of Refinitiv Datastream, Financial Times and Investing.com and then converted to return series using the following formula.

$$R_{t} = ln\;P_{t} - ln\,P_{t - 1}$$
(1)

where is \(R_{t}\) return series, \(ln\;P_{t}\) is the natural log of the current price of each series, whereas \(ln\;P_{t - 1}\) is the natural logarithms of the lagged price of each series.

3 Methodology

This study is broadly divided into two parts, first building the model to correctly assign the sentiment score to the given tweet, and second making a pipeline to concatenate the sentiments of daily tweets to the daily stock prices of the selected countries to study the relation between the tweets and stock price fluctuations worldwide. This section will describe the methodology adopted in this study along with a justification for the chosen approach.

3.1 Sentiment analysis models

To identify the robust sentiment analyser suitable for this study, a brief literature review was conducted (Hartmann et al. 2023; Kearney & Liu 2014; Umer et al. 2021), and the latest text sentiment classification models were reviewed which include Valence Aware Dictionary and Sentiment Reasoner (VADER), Gaussian Naïve Bayes (GNB), Support Vector Machine (SVM), Decision Tree (DT), and hybrid LSTM-CNN.

The first model considered in this study is VADER (Elbagir & Yang 2019), which belongs to the Natural Language Toolkit (NLTK) (Bird et al. 2009). In addition to VADER, NLTK also supports other types of tools in Python to work efficiently with human language. VADER can analyse text data and assign a score representing the underlying sentiment associated with the text. VADER starts with the polarity score which can vary between –1 and + 1 where –1 represents the extreme end of negativity and + 1 represents the extreme end of positivity. A neutral sentiment is represented by 0. Unlike TextBlob (Loria 2018), VADER also returns compound score which is representative of the overall sentiment of the given text. These days, emotions are not only expressed as mere text but are often combined with emojis and capitalisation of text. VADER can process such types of emotions reasonably well. Likewise, intensifiers and negations are also taken care of in a VANDER algorithm.

The second model applied in this study is GNB which is a variant of parent classifier Naïve Bayes which assumes that the features within the dataset follow a Gaussian distribution. The widely used Bayes probability theory serves as the foundation for the Naive Bayes classifier family of classifiers. Based on the premise that the features in a dataset are mutually independent, the probabilistic model of Naive Bayes classifiers is derived from Bayes' theorem. Even though the independence condition is frequently broken in real life, Naive Bayes classifiers nonetheless frequently outperform other models in this imaginary scenario. Naive Bayes classifiers are employed in many different disciplines due to their relative robustness, ease of implementation, speed, and accuracy. On the other hand, really subpar results may result from severe breaches of the independence assumptions and nonlinear classification issues (Raschka 2014).

The third model employed in this study is SVM which is a supervised machine learning classification technique. SVM locates the dividing hyperplane that separates vector space into sub-sets of vectors. Each separated sub-set is allocated to a single class. The margin between two sub-sets must be maximised by the dividing hyperplane. Let us assume that we have a set of p-dimensional vectors, each falling into one of two categories. Such vectors can be classified by many p–1-dimensional hyperplanes, but only one hyperplane maximises the margin between two classes. This is known as the maximum-margin hyperplane, and it is regarded as an SVM classifier (Nguyen 2017).

The fourth model applied in this study is DT. The DTs may display the structure of decisions made throughout the classification process and give the findings in an easy-to-understand style for specialists. This is why in data mining DTs are among the most often used classification methods. In DT, the fundamental idea is to build a tree with inner nodes that reflect descriptive qualities and leaves that are labelled with specific values for the class attribute. The offspring of an inner node N corresponds to several possible values of the corresponding descriptive characteristic. After a decision tree is constructed, a new instance’s class value is found by following a path from the root to a leaf based on the descriptive attribute values of the instance. The leaf’s label will serve as the class value (Stiglic et al. 2012).

The final model applied in this study is a hybrid of CNN and LSTM. CNN networks come under the family of deep learning and often exhibit excellent performance. CNN has been extensively used in computer vision, speech recognition, object detection, video analysis, picture classification, and bioinformatics. CNN is deeper than other neural networks since it has several layers, unlike regular neural networks. Additionally, CNN uses a nonlinear activation to use weights, biases, and outputs. The three main layers of a standard CNN architecture are convolutional, pooling, and fully connected. A type of recurrent neural network (RNN) called a Long Short-Term Memory (LSTM) network makes use of memory blocks to help it function well and learn more quickly than a conventional RNN. In the LSTM network, in addition to RNNs, a cell state is employed to store long-term states such as input, forget, and output gates. As a result, the network can recall past information and relate it to current data. A CNN-LSTM model is a combination of CNN layers that extract the feature from input data and LSTM layers to provide sequence prediction. They are mainly developed for the application of visual time series prediction problems and generating textual annotations from image sequences (Tasdelen & Sen 2021).

3.2 Integration of daily average sentiments with bank stock indices returns

The first part of the study involved building the model to correctly assign the sentiment score to the given tweet. This part is very much reliant on machine learning models which require training data. To train the competing machine learning models, review data from sources like IMDB Amazon, IMDB, and Yelp were collected. Altogether five models namely, VADER, DT, SVM, LSTM-CNN, and GNB were chosen based on the findings from the literature base. Amongst these chosen candidate models, VADER does not fall under the realm of the machine learning family and belongs to a more traditional sentiment analyser paradigm known as the lexicon-based model. A total of 180 (60 positive, 60 neutral, and 60 negative) tweets were manually labelled by a co-author who is a finance expert. The manual labelling of 60 tweets helped in the preparation of the testing dataset. All the five candidate models were then tested using the testing data and based on the accuracy measures only three of them namely, VADER, GNB, and LSTM-CNN were selected for further analysis. The performance of all three models was subsequently quantified using Confusion Metrics and F1 score. To merge the results of this step with the next step involving stock data, the average sentiment of tweets on a daily basis was calculated.

In the second part of the study, a pipeline to concatenate the sentiments of daily tweets to the daily stock prices of the selected countries was made to study the relationship between the tweets and stock price fluctuations worldwide. By following Adams et al. (2019) and Wang et al. (2020), we normalised the bank stock indices returns of each country using a logarithmic transformation to mitigate the impact of outliers. Following this, a concatenation of stocks with average daily sentiment of tweet was carried out. A standard set of Exploratory Data Analysis (EDA) and visualisations were performed to reveal a correlation between stock price and tweet sentiment.

4 Results

4.1 Descriptive statistics

Table 1 provides the descriptive statistics of the sample. Panel A of Table 1 reports the results for polarity and subjectivity corresponding to the SVB collapse. The polarity shows that half of the tweets were negative, suggesting a significantly negative investor sentiment towards the collapse of SVB. Likewise, in the case of subjectivity, 53% of comments were objective.

Table 1 Descriptive statistics
Full size table

Specifically, both the polarity and subjectivity results reveal that the news related to SVB since the 1st of March was not positive for the bank and various stakeholders due to strong and crisp messaging from investors on Twitter. Panel B of Table 1 presents the returns behaviour of selected G-SIBs bank stock indices. Overall, all markets show variation in their returns during the period. Interestingly, the returns of five markets, Italy, Spain, Sweden, Switzerland, and the USA were negative from March to April 2023 suggesting that these markets observed the shock of SVB implosion.

Figure 1 presents the results regarding the word cloud. The word cloud corresponding to SVB implosion reveals critical and negative words. Intriguingly, the word cloud regarding negative tweets shows that the implosion of SVB is linked to Credit Suisse. This finding confirms that the contagion shock from the SVB implosion was transmitted from the US market to the European market. As a result, one of the largest European Global Systemically Important banks, Credit Suisse, declared bankruptcy and was taken over by UBS.

Fig. 1
figure 1

Word Cloud based on keywords regarding SVB collapse

Full size image

4.2 Performance evaluation parameters of sentiment analysis models

In our empirical strategy for estimating the results, the first phase was to build a robust model for sentiment analysis. After training and comparing in-sample and out-of-sample testing scores, the best three performing models out of five were selected, namely Gaussian Naïve Base, Vader NLTK (Pre-Built NLTK Library model) and the Hybrid LSTM- CNN (Convolutional Neural Network, Long Short-Term Memory) Model. Table 2 presents the results regarding performance evaluation parameters corresponding to the top three selected models. In addition, the confusion matrix corresponding to each model is presented in Figure A1 in Appendix A.

Table 2 Model summaries based on manual tagged tweets data
Full size table

4.3 The spillover effect of investors’ sentiment generated from SVB collapse on the G-SIBs bank stock indices

To test the effect of social media-based SVB collapse on bank stock indices from twelve countries, Hybrid LSTM-CNN architecture was selected to analyse the investors’ sentiment associated with their tweets from our scraped dataset. Moreover, to quantify the daily sentiment tres, we used the following formula:

$${\text{Average daily sentiment }} = { }\frac{{\left( {{\text{No of positive tweets }}{-}{\text{ No of negative tweets}}} \right)}}{{\left( {{\text{No of positive tweets }} + {\text{No of negative tweets}}} \right)}}$$

This approach enabled us to capture the overall sentiment for each day. Concurrently, we calculated the log returns for bank stock indices from twelve countries. Subsequently, we merged these two datasets, aligning them by date to correlate the stock market performance with investors’ sentiment regarding SVB collapse. Importantly, we normalized both the average sentiment scores and the log returns using z-score method. This normalisation ensured that sentiment values were on a comparable scale to the log returns, facilitating a more accurate and meaningful analysis of the relationship between investors’ sentiment and stock market fluctuations.

We started our analysis to test the spillover effect of the social media-based investor sentiment and SVB collapse on the US bank stock index where SVB was operating. Figure 2 presents the results. We observed a consistent 1-day-lagged correlation between the average social media-based investor sentiment generated from the SVB implosion and the US bank stock index returns. The pattern shows that a decline in average investors’ sentiment leads to a decrease in the stock returns on the subsequent day. This trend persisted throughout March and into early April, highlighting a clear link between social media investors’ sentiment and stock returns fluctuation.

Fig. 2
figure 2

The spillover effect of investors’ sentiment before and after SVB collapse on the US bank stock index

Full size image

These findings clearly suggest that the sharp decline in US bank stock index returns in early March was a result of negative investor sentiment towards the largest US bank, which eventually collapsed on the 10th of March 2023. Importantly, this investor sentiment persisted and led to the collapse of other banks in the US, namely Silvergate Bank and Signature Bank, as well as one European bank, Credit Suisse.

Moreover, Fig. 3 exhibits the results regarding the spillover effect of investors’ sentiment before and after SVB collapse on other G-SIBs bank stock indices. Again, Fig. 3 clearly shows a spillover effect of social media-based investors’ sentiment generated from SVB implosion on selected bank stock indices returns over the sample period. Specifically, in the case of Canada, there were instances where a high sentiment score seemed to precede an increase in log returns, suggesting a potential lagged relationship. Similarly, the bank stock index of France revealed some correlation, with changes in sentiment often followed by similar directional changes in log returns. For Germany and Spain, the findings showed several instances where a peak in sentiment precedes a rise in log returns. The bank stock index of the Netherlands observed some movements in returns that seem to follow changes in sentiment, suggesting a possible lagged effect. Likewise, the bank stock index of Switzerland observed some points where sentiment changes precede stock return movements by a day. Contrary, for China, Italy, Japan, Sweden and the UK, the relationship was less clear, with sentiment peaks and troughs not consistently followed by movements in log returns.

Fig. 3
figure 3

The spillover effect of investors’ sentiment before and after SVB collapse on other G-SIBs bank stock indices

Full size image

4.4 Granger causality analysis

In addition, Granger causality analysis was employed to obtain the unidirectional causality between social media-based investor sentiment triggered by SVB collapse and G-SIBs bank stock indices. In our case, the Granger causality test will help determine whether investors’ sentiment can predict the log returns of these G-SIBs bank stock indices. Table 3 reports the results of the Granger causality analysis. Overall, our estimation results show the rejection of the null hypothesis for no causality for G-SIBs stock markets with social media-based investor sentiment generated from SVB collapse. Hence, the causality exists between returns and investors’ sentiment in unidirectional for G-SIBs stock markets. For example, in the case of the US, p-values at lags 2, 3, 4, and 5 are significant at 10%, 5% and 1% significance levels, respectively suggesting a strong lead and lag relationship between investors’ sentiment and the US bank stock index. Likewise, in the case of Switzerland, p-values at lags 2, 4, and 5 are significant at 5% and 1% confidence intervals, respectively indicating a strong predictive relationship of investors’ sentiment on log returns at these lags.

Table 3 Granger causality test for lead and lag relationship between social media-based investors’ sentiment and G-SIBs bank stock indices from twelve countries
Full size table

Importantly, a delayed response of SVB collapse on G-SIBs was observed. One possible reason for the delayed response of the SVB collapse on G-SIBs markets is the lower degree of connectedness, as SVB was not part of the G-SIBs. Although this response was delayed, the SVB failure triggered turmoil that led to global ripple effects, and in Switzerland, this resulted in the takeover of Credit Suisse by UBS, one of the largest financial institutions worldwide. Importantly, this type of bank run and spillover effect has significant economic and financial costs. For instance, Elliott et al. (2014) and Egger et al. (2023) reported that negative shocks from and on G-SIBs can lead to severe global economic repercussions. Our findings support the existing literature on systemic risk and new financial regulations, such as the Basel III agreement and the Financial Stability Board mandate, which explicitly designate G-SIBs status globally and domestically. These regulations recognise the importance of G-SIBs in mitigating systemic risk in the global and domestic financial industry.

4.5 Robustness tests

Our graphical findings reveal a significant impact of Twitter-based investors’ sentiment on US bank stock index returns, with some evidence also prevalent for other G-SIBs bank indices. To refine our main findings, in this section, we provided cointegration analysis and vector error correction model analysis to explore the long- and short-run relationships between Twitter-based investors’ sentiment and the returns of G-SIBs bank indices. Additionally, we conducted frequency domain Granger causality analysis to provide valuable insights into the temporal dynamics of causality relationships.

4.5.1 Cointegration and vector error correction model (VECM) analysis

Table 4 presents the results of the unit roots test and cointegration analysis.

Table 4 Unit roots test and cointegration analysis
Full size table

We started by testing for unit roots in each series at levels and found that all variables are non-stationary at this stage, indicating that the mean of the data is not constant over time. Next, we tested for unit roots at the first difference, using logarithmic transformations of the series. The results of the unit root test revealed that all variables are stationary at the first difference. In addition, for all countries, the trace statistics exceed the critical values at the 99% significance level, indicating strong evidence of cointegration between the Twitter-based investors’ sentiment and the log returns of G-SIBs bank indices. This suggests the presence of long-term relationship between sentiment and G-SIBs stock market performance. The presence of cointegration implies that any deviations from this long-term relationship are temporary and that the sentiment and market returns will converge over time. These results can provide valuable insights for models aiming to predict market behavior based on sentiment or for strategies that consider sentiment as a factor influencing market dynamics.

In the next step, VECM test was employed to explore the short-term relationship among our variables of interest. Table 5 provides the results of VECM analysis.

Table 5 VECM analysis
Full size table

In the case of the USA, for instance, the statistically significant coefficient of the error correction term (ec1_sentiment) indicates a strong adjustment mechanism in sentiment towards long-term equilibrium. The negative sign implies that sentiment adjusts downwards towards equilibrium when it is above its long-term level. In other words, an increase in the sentiment score leads to an immediate decline in stock returns. These findings confirm the spillover results presented in Fig. 2 regarding the sharp decline in US bank stock index returns. Additionally, the coefficient of ec1_stock was not statistically significant at conventional levels, indicating a weak correction towards equilibrium after short-term shocks. The negative and significant short-run impact coefficient (L1.stock) suggests a strong negative short-term impact of sentiment on stock returns. Finally, a long-run relationship was not observed.

4.5.2 Frequency domain Granger causality analysis

Table 6 reports the results of the frequency domain Granger causality analysis.

Table 6 Frequency domain Granger causality analysis
Full size table

Our findings revealed that Canada, China, Switzerland and USA stand out with significant results at higher lags, suggesting a delayed influence of sentiment on stock market returns. This might indicate that sentiment metrics, particularly those capturing longer-term trends or cumulative sentiment, can be predictive of stock returns in those specific markets. Most countries do not show a statistically significant Granger causality from sentiment to stock returns at the tested lags. Qualitatively, the findings of unidirectional causality are confirmed by the frequency domain Granger causality analysis regarding temporal dynamics of causality relationships.

5 Conclusion

The purpose of this study was twofold. First, we aimed to examine the role of social media-based investors’ sentiment in the collapse of SVB. Second, we sought to explore the domino effect of social media-based investor sentiment and SVB’s implosion on major bank stock indices worldwide, particularly those where Global Systemically Important Banks (G-SIBs) operate.

Our empirical strategy was carried out in the following manner. We started the empirical investigation by modelling social media-based investors’ sentiment using tweet data extracted from Twitter, employing various hashtags such as #SiliconValleyBank, #SVB, #SVBCollapse, and #SiliconValleyBankCollapse. In total, 279,779 tweets were extracted from 1st March 2023 to 4th April 2023, including 143,578 positive and 136,201 negative tweets. The extraction process was facilitated through a Python application called snscrape. Subsequently, based on the Twitter data gathered, we utilised advanced machine and deep learning models, namely LSTM-CNN, GNB, and NLTK-VADER, to model the sentiment of social media-based investors regarding the SVB collapse and its potential domino effect on the bank stock indices of G-SIBs.

The findings of our study revealed that negative social media-based investor sentiment was one of the important factors in the collapse of SVB, and it played a key role in the subsequent contagion of risk in G-SIB markets. Interestingly, our study established a temporal lead and lag relationship between investor sentiment and the bank stock indices returns of G-SIBs, with investors’ sentiment preceding market fluctuations. Finally, an examination of the word cloud generated from negative tweets highlights a connection between the SVB collapse and Credit Suisse (Switzerland-based G-SIB). This suggests that the contagion effect of SVB’s failure extended from the US market to the European market.

The findings of this research offer crucial insights for policymakers and highlight improvements in the existing banking prudential framework for maintaining financial stability. Policymakers should place particular emphasis on the impact of investors’ sentiment on social media when crafting policies related to bank runs. By doing so, policymakers can proactively anticipate and address potential external shocks originating from social media, thus safeguarding financial stability and preventing cross-border contagion. Lastly, the insights of this study can serve as a useful springboard for future research. For instance, future studies can use machine learning and deep learning algorithms to examine the impact of social media-based investor sentiment on asset prices in the context of systemic risk and market efficiency from a behavioural finance perspective. Specifically, subsequent studies can conduct deeper research on constructing a bank risk contagion network or exploring the medium- to long-term role of emotions in asset pricing.