Predicting abnormal trading behavior from internet rumor propagation: a machine learning approach

Abstract

In 2021, the abnormal short-term price fluctuations of GameStop, which were triggered by internet stock discussions, drew the attention of academics, financial analysts, and stock trading commissions alike, prompting calls to address such events and maintain market stability. However, the impact of stock discussions on volatile trading behavior has received comparatively less attention than traditional fundamentals. Furthermore, data mining methods are less often used to predict stock trading despite their higher accuracy. This study adopts an innovative approach using social media data to obtain stock rumors, and then trains three decision trees to demonstrate the impact of rumor propagation on stock trading behavior. Our findings show that rumor propagation outperforms traditional fundamentals in predicting abnormal trading behavior. The study serves as an impetus for further research using data mining as a method of inquiry.

Introduction

In January 2021, GameStop’s stock price increased from $16 to $347 as a result of the trading behavior of many small investors rather than their underlying financials (Umar et al. 2021). The unusual, or abnormal, short-term stock price variations highlight the significant influence of the volume of Internet stock discussions on corresponding stock trading activities (Anderson et al. 2021; Lyócsa et al. 2021), prompting calls for the Securities and Exchange Commission to address the issue (Anderson et al. 2021). The efficient market hypothesis in traditional finance states that “unemotional” investors are constantly updating their beliefs about market directions as new information about national economies or firm fundamentals becomes available, causing stock prices to fluctuate around their intrinsic values (Wang et al. 2019). The health of stock markets is dependent on the accuracy, timeliness, and transparency of information. However, rumors spread through stock discussions interfere with investors’ decision-making, influencing their sentiments (Wang et al. 2020) and, ultimately, their trading behavior. More importantly, as Internet technologies and electronic commerce tools advance, rumors on digitized platforms make larger interferences with the stock market. Furthermore, rumors evolve, gain visibility, and become more influential as they spread through the media (Shin et al. 2018). These not only undermine information transparency, but also put the stock market at risk.

Today, investors obtain information about stocks online (Wen et al. 2019), and rumors are one of the most widely studied types of misinformation on the internet. They refer to information that has not been verified by official sources and may or may not be correct. Social media platforms are frequently used to quickly share information among users (Bondielli and Marcelloni 2019) and are a breeding ground for rumor propagation (Zubiaga et al. 2018; Wang et al. 2021). Social media platforms are also valuable sources of stock market information. Fake financial news causes temporary fluctuations in stock prices and an increase in trading volume, resulting in abnormal trading behavior (Kogan et al. 2019). As such, sentiments expressed in tweets have been used to forecast stock market reactions (Arif et al. 2016; Bustos and Pomares-Quimbaya 2020). Although they improve user information dissemination, they also have the potential to disrupt markets (Klein 2021), resulting in unexpected and unusually large trading volumes.

However, we identify two gaps in the literature. First, unlike traditional managerial and financial variables, the impact of Internet stock rumors and rumor propagation on trading behaviors has received little attention. There are no satisfactory explanations for how or why rumors influence stock prices (Fong 2021), despite the fact that their impact on stock markets (Prabhala and Bose 2019) is recognized (Majumdar and Bose 2018). Although several recent studies have used text mining for stock market prediction (cf. Nassirtoussi et al. 2015; Oliveira et al. 2017; Zhang et al. 2018; Vanstone et al. 2019; Bustos and Pomares-Quimbaya 2020; Gupta et al. 2020), studies that specifically focus on rumors and their propagation are lacking. Studies have demonstrated the impact of rumors on the GameStop squeeze (Anderson et al. 2021; Lyócsa et al. 2021), but Klein (2021) called for more research in this area that goes beyond this single incident. Second, machine learning algorithms, though increasingly used in stock market analysis today (Zhong and Enke 2019), are still comparably less used to predict trading behaviors despite their higher accuracy (Barboza et al. 2017). To the best of our knowledge, no study has used machine learning to predict abnormal stock trading behavior based on internet rumor propagation.

In our exploratory study, we use machine learning on social media and numerical data to address these two gaps in a novel way. Our goals are to compare the impact of internet rumor propagation and fundamental predictors, and to train a model that can predict abnormal stock trading behavior that is challenging and risky for firms (Anderson et al. 2021). The findings, specifically identifying rumor propagation variables that are missing in the literature, are useful to financial analysts and academics studying stock trading predictions, and they contribute to the growing body of knowledge on stock markets. Global stock market regulators can use a similar approach to forecast based on each country’s definition of abnormal trading behavior, echoing a call for regulation to avoid market inefficiency (Anderson et al. 2021; Umar et al. 2021). Bankers and academics also advocate a need for financial regulation that can reduce systemic risks (Kou et al. 2019).

The remainder of the paper is structured as follows. The “Literature review” Section examines the literature on internet stock rumors and how they affect stock trading behavior, as well as other traditional fundamentals. The “Method” Section discusses the data and the machine learning method. The “Results” Section summarizes the findings. Finally, the “Conclusionanddiscussion” Section concludes the paper.

Literature review

Rumor propagation

Information exchange is a critical factor in shaping public opinion and decisions. Access to electronic word of mouth has transformed the way Internet users communicate (Ma et al. 2019), allowing for closer social connections and interactions (Li et al. 2022). This exchange is propagated by new media sources, such as discussion boards and social media, among others (Wang et al. 2021), where peer opinions influence investors and their stock trading behavior (Li et al. 2018). Text mining methods on social media data have been used previously for stock market prediction because they can be used to analyze investor attitudes and opinions (Kou et al. 2019). Table 1 summarizes recent related articles that demonstrate the impact of web-based communication–news and social media–on stock markets. However, the majority of these are based in the United States, where abnormal stock trading behavior is not officially defined and is instead based on sample distributions (cf. Joseph et al. 2011). We focus on rumor propagation over time in the days preceding specific stock-day trades. This is not explicitly and thoroughly addressed in the literature. We quantify rumor propagation by investigating rumors in the 30-, 60-, and 90-day periods preceding each stock-day trade.

Table 1 Recent articles using text mining for stock market prediction

To define rumor propagation, we first examined how misinformation–false or unverified information–spread via the Internet, strongly influences stock prices (Fong 2021) and attracts significantly more attention to stock price reaction than legitimate information (Clarke et al. 2020). There is a scarcity of research on the propagation of misinformation (Jang et al. 2018) and how it influences human behavior (Bastick 2021), or in this case, stock trading behavior.

Uninformed traders who react to misinformation prefer to trade in equity markets rather than options markets (Brigida and Pratt 2017). Furthermore, heavy social media users gaining information from these platforms are four times more likely to blindly follow the actions of other traders. Misinformation has been found to increase abnormal stock trading activity and stock price volatility by more than 50% and 40%, respectively, in the United States (Kogan et al. 2019). These can also be used to promote specific stocks, resulting in stock trading speculative campaigns (Tardelli et al. 2020). Opinion leaders can influence the behavior of other users (Chen et al. 2021); in crowdsourced systems, they can have a greater influence in stock price manipulations through their posts (Wang et al. 2017), similar to how expert reviews via electronic word of mouth influence purchase decisions (Naujoks and Benkenstein 2020). Taken together, misinformation can lead to market instability and should be detected as soon as possible (Bondielli and Marcelloni 2019).

Official sources, opinion leaders, and crowdsourcing can all be used to validate information spread via the Internet (Chen et al. 2021). However, this takes time to happen. An efficient financial market should reflect accurate stock prices (Anderson et al. 2021), but misinformation comes at a high cost (Bondielli and Marcelloni 2019) that reduces market efficiency and stability because it undergoes more content modifications in the propagation process than accurate information (Jang et al. 2018). Therefore, it is worthwhile to explore the impact of information propagation prior to verification. As shown in the recent GameStop phenomenon, the sheer volume of verified and unverified information can heavily influence stock trading behavior (Anderson et al. 2021; Lyócsa et al. 2021).

Thus, rather than examining the impact of misinformation, we examine rumor propagation. Rumors are a type of web-based communication that has the potential to influence stock markets. We define rumor propagation as the spread and growth of rumors about a particular topic over time. This goes beyond diffusion, which is defined by spreading. Stock market traders are more concerned with receiving information quickly than with verifying its accuracy (Oberlechner and Hocking 2004), emphasizing the importance of studying rumors. Rumors are circulating relevant information that has yet to be verified, which is consistent with longstanding literature (DiFonzo and Bordia 2007; Donovan 2007). Rumors become legitimate information or misinformation after being verified. The spread of rumors occurs in four stages: detection, tracking, stance, and veracity stages. The detection stage detects information and determines whether it is a rumor based on its veracity. In the tracking stage, the identified rumors are disseminated through various media platforms. Meanwhile, rumors are classified as supporting, denying, querying, or commenting on a specific topic in the stance stage. Finally, they are classified as true, false, or unverified at the veracity stage (Zubiaga et al. 2018). Our research focuses on rumors in the tracking stage because our goal is to develop an early-stage detection of abnormal trading behavior. At this point, rumor classifications like supporting or denying are irrelevant.

Rumors spread through strong ties among users in a social network (Zubiaga et al. 2018); specifically, social media networks that allow users, including retail and institutional investors, to share information (Umar et al. 2021). Stock rumors are the spread of information about stocks. Because of their influence on market movements, Majumdar and Bose (2018) recognized the importance of detecting financial rumors in stock market regulation (Prabhala and Bose 2019). We focus on three stock rumor propagation variables in this study: the number of “posts,” “replies,” and “likes,” which amount to the extent to which stock rumors influence stock trading behavior (Anderson et al. 2021; Lyócsa et al. 2021). The number of posts represents the extent of stock-related discussions or rumors. This is similar to how Internet search volume reflects investor interest (Wen et al. 2019) define replies as direct responses to posts (O’Dea et al. 2018). The number of responses indicates the level of interest in the posts or rumors. “Likes” denote positive responses or appreciation (O’Dea et al. 2018), implying the sentiments of responses that have been shown to influence trading behavior (Sabherwal et al. 2011).

Management shocks

Unexpected news and events about a company have been said to influence stock market trading behavior (Sindhu et al. 2014; Prasad and Prabhu 2020). Some investors base their stock trading behaviors on stock rumors, whereas others pay attention to relevant company news and events to make informed trading decisions. Management shocks are among the unexpected circumstances that influence investor sentiments, which in turn, translate to stock trading decisions. We include management shocks for comparison to assess the impact of rumor propagation, which has received less attention in this regard.

We anticipate that stocks that have experienced management shocks, such as strikes and unexpected top management changes, will trade abnormally. Strikes have been found to negatively influence stock prices because they cause investors to lose confidence and perceive an increase in risk (Wisniewski et al. 2020). As a result, this loss of stability can trigger increases in trading behavior in the short term. Furthermore, senior executives have a significant impact on a company’s strategic direction. In turn, their decisions affect how a company’s stock trades in the market. Stock prices have been found to be significantly influenced by senior executive characteristics and changes in senior management in a company (Yilmaz and Mazzeo 2014). Thus, the unexpected deaths of these senior executives cause significant stock price reactions (Salas 2010). In the US, stock prices were found to decrease by an average of 0.85% following the death of a director (Nguyen and Nielsen 2010).

Other factors

Additionally, investment decisions may be influenced by industry wide events that trigger trading behavior across firms within the same industry. For example, bankruptcy protection of US airlines influences many of their stock prices (Gong 2007). It follows that the higher the shareholding, the larger the potential to affect stock prices via trading behavior. Thus, in this study, we include both industry type and shareholding variation, akin to the common inclusion of demographics in surveys.

Method

The main steps of our research methodology are depicted in Fig. 1. The following subsections provide explanations for these. In Stage 1, we identified 476 stocks listed on the Taipei Exchange (TPEx) and compiled stock-day records to determine whether any of them had been flagged for abnormal trading. Each stock was broken up into 365 stock-day records for September 2019–August 2020, with each record containing the result of whether it was flagged for abnormal trading. As abnormal trading flags apply to each trading day rather than specific stocks, we computed stock-day trades as records for the final data set in Stage 2. Then, using stock-trending keywords to identify posts related to trading, we compiled the corresponding stock discussions on these chosen stocks in Stage 3. This involves compiling the volume of posts, likes, and replies for 30, 60, and 90 days prior to each stock-day record. This enabled us to compare the extent of influence from rumors in the time period leading up to the normal/abnormal trading day for each stock. In Stage 4, we acquired corresponding management-related data on these same stocks. The result is a flat file with all the machine learning-related variables. In Stage 5, we use descriptive statistics to describe the data. Our initial model, which only includes management variables, is discussed in Stage 6. Stage 7 saw the addition of rumor propagation variables to demonstrate model performance gains. Finally, in Stage 8, we built on our earlier findings by boosting the model with variables for management and rumor propagation to create a model that is useful for predicting abnormal trading behavior.

Fig. 1

Research method illustration

Stock trading in Taiwan

We chose stocks listed on the Taipei Exchange (TPEx) as the study’s focus for two reasons. First, our study requires an established standard to define abnormal trading behavior, the target variable, in order to determine its predictors. Such standards can be defined by trade volume or price differentials relative to existing baselines, and they can vary across contexts. However, defining abnormal trading is difficult. Using arbitrary standards leads to poorly defined labels and, as a result, inaccurate classification models. In Taiwan, there is a legal system in place whereby stocks are flagged by the Taiwan Stock Exchange Corporation (TWSE) for abnormal trading by day as an advisory caution to investors (Law Source Retrieving System of Stock Exchange and Futures Trading 2021), because the same stock can be flagged multiple times in a given period of time. There are several conditions that legally define an abnormal stock-day trade under Taiwanese law, such as having a trading volume of at least 3,000 trading units, an intraday transaction price exceeding 9%, and exceeding the amplitude of change in the TWSE Capitalization-Weighted Stock Index by at least 5% (Law Source Retrieving System of Stock Exchange and Futures Trading 2021). Following that, stock days flagged for abnormal trading are kept in a public database that explains each legal violation for transparency’s sake (Taiwan Stock Exchange Corporation 2022). For clarity, we used the availability of rigorously and legally defined labels of abnormal trading in our study. Second, rather than the offline grapevine, our study requires easy access to online stock trading discussions. Due to the widespread use of the Internet in Taiwan, it is common for investors to discuss and share relevant stock news and information online at CMoney, the go-to platform for these discussions. While Yahoo Finance, Reddit, Twitter, and Facebook are popular stock discussion platforms in the United States, CMoney is Taiwan’s most popular social media platform for stock discussions. Using discussions from this primary source enriches the data, allowing us to capture the majority of stock discussions in Taiwan more comprehensively.

In Taiwan, stocks that are flagged for abnormal trading face severe penalties, including severely restricted trading. Furthermore, stock investors in Taiwan are not permitted to purchase them using margin accounts and must instead trade with existing cash balances. These discourage similar trading behaviors and imply the TWSE’s commitment to healthy trading in a well-functioning market. This zealous policing also strengthens the definition of abnormal trading.

The results will help the TWSE predict abnormal trading activities before they occur, saving the authorities money in their efforts to curb such behavior. It is important to note, however, that these findings can also be applied to other countries.

Data

From September 2019 to August 2020, we used the 476 over-the-counter stocks listed on TPEx to compute the abnormal trading variable for each stock-day record to reflect whether its corresponding day trading activity was flagged for abnormal trading (1 = “Yes,” and 0 = “No”), as legally defined by TWSE (Law Source Retrieving System of Stock Exchange and Futures Trading 2021). For data on financial rumor propagation, we extracted 487,193 specific discussion posts from CMoney for the same period that were related to these 476 over-the-counter stocks. Following that, we calculated the corresponding variables for each stock-day based on discussion post attributes (volume, likes, and replies). In addition, for each stock-day record, we included the relevant financial and management predictors discussed in the literature review. These were obtained from the Taiwan Economic Journal (TEJ), a large database of financial data from the Greater China region. Table 2 summarizes the definitions of these variables. The computations of these variables are explained in the following subsections.

Table 2 Variable definitions

Stock discussion posts

Text mining can be used to analyze textual data from various media, such as social media, financial news, and reports (Kou et al. 2019). Social media is a breeding ground for rumors (Wang et al. 2021), and meaningful information has been extracted for financial analyses using text mining (Kumar and Ravi 2016; Gupta et al. 2020). We calculated the daily discussion volume of each TPEx-listed stock after performing word segmentation on CMoney discussion posts. The sums of daily posts, daily replies to posts, daily likes, and daily posts with stock-trending keywords comprise the daily discussion volume (Table 3). Because CMoney posts are written in Taiwanese Mandarin, some of the translated keywords are local colloquialisms that may not make sense when back-translated literally. We calculated the number of discussions for each stock-day record in the 30-, 60-, and 90-day periods preceding an abnormal trading day (Fig. 2). These times were chosen based on TWSE regulations for unusual stock-day trades (Law Source Retrieving System of Stock Exchange and Futures Trading 2021). Using official historical records, abnormal stock-day trades are flagged for legal violations (Law Source Retrieving System of Stock Exchange and Futures Trading 2021) within the preceding 30, 60, and 90 days of each stock-day trade. For example, Big Sun Shine Co. Ltd. (formerly known as Sumagh High Tech Corp.) was in violation on July 3, 2020, because its closing price had increased by 250.34% from its corresponding price 60 days earlier. Subsequently, on August 6, 2020, the increase was 192.60% over the previous 90-day price (Taiwan Stock Exchange Corporation 2022). To avoid endogeneity issues, we did not include discussion volume by day. Furthermore, we calculated variation as the difference in discussion volume between the two preceding time periods.

Table 3 Stock-trending keywords on CMoney discussion posts

Fig. 2

Date selection of discussion volume in this dataset

$V_{i,D}$ is the discussion volume for stock $i$ within a day in CMoney on an abnormal trading day $D$. $V_{accumulation}$ is the summed daily discussion volume of a stock up to the preceding $X$ days ($X = 30, 60, 90$) (Eq. 1):

Discussion volume accumulation computation

$$V_{accumulation} = \mathop \sum \limits_{k = 0}^{x} V_{i,D - k}$$

(1)

$V_{variation}$ is the difference in the discussion volume of a stock between the preceding $X_{1}$th and $X_{2}$th day ($X_{1} ,X_{2} = \left( {0,30} \right), \left( {30,60} \right), \left( {60,90} \right)$ (Eq. 2):

Discussion volume variation computation

$$V_{variation} = \left| {V_{{i,D - x_{1} }} - V_{{i,D - x_{2} }} } \right|$$

(2)

Management shocks

We include three types of management data in addition to rumor propagation (Table 1). Managerial shocks assess whether the company had a strike or a change in CEO in the previous year. TWSE defines these shocks in Taiwan, and corresponding data for each company is publicly available. We used a binary variable to indicate whether a company had a managerial shock the previous year.

Other data

In terms of shareholding data, we used the Taiwan Depository and Clearing Corporation’s Distribution of Shareholdings Proportion to distinguish between individual and institutional ownership of shares. Due to the continuous variables in the distribution of shareholdings proportion, we divided it into five intervals using two data partitioning methods, including the first multiple partitioning method. Individual investors are those who own up to 20 lots, while major investors own more than 600 lots (each lot is equivalent to 1000 shares).

The Percentage of Centrally Deposited Securities represents the share capital of a company as a percentage of a specific population. If major shareholders continue to hold large amounts of company stock, it indicates that they are bullish on the company’s future. We computed $S_{i,D}$ as the percentage of centrally deposited securities for stock $i$ in the week of an abnormal trading day $D$. $S_{variation}$ denotes the shareholding of the week for the $X$th day ($X = 30, 60, or 90$) (Eq. 3):

Shareholding variation computation

$$S_{variation} = S_{i,D} - S_{i,D - x}$$

(3)

Finally, stock trading activity trends may occur across the industry (Chan et al. 2013). For example, negative news about the retail industry can have a negative simultaneous effect on the stock prices of several retail companies over the same time period. Taiwan has 25 TEJ industry classifications. These are categorized into six industry groups in this study: biotechnology and medicine, cultural and creative, electronics manufacturing, information and communications technology, other manufacturing, and other service industries. In addition, pertinent to the Taiwanese context, stocks are classified according to their incorporation type: General or KY. The former includes stocks from Taiwan-based companies such as Hi-Lai Foods and Mornsun, and the latter includes companies incorporated outside of Taiwan that had their initial public offering in Taiwan.

Decision tree induction

Among data-driven approaches, machine learning algorithms are increasingly used to make accurate predictions on financial risks (Kou et al. 2019), and they can be used to investigate the impact of online misinformation (Choudrie et al. 2021), and are becoming increasingly popular in stock market analysis (Zhong and Enke 2019). In financial analyses, they have been shown to be more accurate than traditional statistical models (Barboza et al. 2017). Furthermore, due to the non-linearity of the data, traditional linear or continuous statistical models may not be appropriate (Antunes 2021).

We trained a decision tree, a supervised learning technique, to predict abnormal trading activities among Taiwanese stock-day trades in this study. Decision trees are a popular supervised machine learning technique (Brieman et al. 1984) and are recommended for analyzing website comments (Tan 2015). They classify outcomes using if–then rules (Osei-Bryson 2004) and can handle non-linear relationships without making assumptions about the data distribution (Pal and Mather 2003). Decision trees have a tree-like structure (cf. Lin and Fan 2019), where the if–then rules partition data repeatedly until the subsets are as homogenous as possible with respect to the outcome variable (Ture et al. 2009; Osei‐Bryson and Ngwenyama 2011). They are simple to interpret and utilize (Murphy and Comiskey 2013), without sacrificing performance (Du et al. 2020), giving them a competitive advantage over other machine learning algorithms that lack transparency and are difficult to understand or apply (Du et al. 2020). Tree-based models, by the way, have been found to be more stable than multilayer artificial neural networks (Addo et al. 2018). Decision trees have been used in medicine (Ture et al. 2009; Murphy and Comiskey 2013; Kobayashi et al. 2013) and marketing (Kim et al. 2011; Amir et al. 2015; Legohérel et al. 2015; Díaz-Pérez and Bethencourt-Cejas 2016). Although used less often in finance, Bondielli and Marcelloni (2019) observed a clear trend in supervised classification approaches to fake news and rumor detection for a variety of projects. Although deep learning models outperform tree-based models in terms of predictive accuracy, they are black boxes (Skrede et al. 2020) and do not explicitly identify the relevance of each predictor (Guo et al. 2016). In contrast, rule-based models are more interpretable (Kou et al. 2021), meeting the need for interpretability when analyzing complex financial data (Li et al. 2021). The goal of this study is to identify specific predictors of abnormal trading behavior in order to compare the impact of rumor propagation versus traditional financial and management predictors. This is analogous to the difficulty in determining priorities in complex issues (Kou et al. 2022). Hence, tree-based models are preferable in this study. Following the identification of these predictors, we discuss the use of deep learning as a future research direction.

The final dataset included 241,332 stock-day records, with only 6110 classified as abnormal trading. Given that data imbalance can have an impact on the performance of classifying algorithms (Budhi et al. 2021), we used an oversampling algorithm to balance the data set, resulting in a 50–50 split of records with normal and abnormal trading. We used these data to perform an 80%–20% training-validation split for supervised learning.

To avoid bias caused by relying on specific training and validation sets, we used tenfold cross validation in our decision trees. Parsimony is essential in machine learning because highly complex models may be ineffective (Shmueli 2016). By limiting the size of decision trees, researchers can simplify them (Esposito et al. 1997). Hence, when two classification trees have similar accuracy rates, the simpler one with fewer leaves and a smaller size is preferred. We limited the decision tree growth for this study by setting the maximum depth to 8 and the minimum terminal node size to 300. We trained two classification trees to identify abnormal versus normal stock trading behavior in stock-day records. The first tree contains only the financial and managerial variables examined in this study. The second adds variables for rumor propagation. Our use of these two classification trees highlights the differences in impact between rumor propagation and company financial and managerial characteristics.

Finally, we boosted the combined classification tree using a random forest (cf. Brieman et al. 1984), which has been used in finance for things like predicting bank lending (Ozgur et al. 2021). A random forest is made up of several random trees that have been trained using bootstrapped observations from the training data (Shmueli 2016). In general, such ensemble models outperform individual classifiers (Kou et al. 2021). In this study, where the relationships between the rumor propagation variables and abnormal trading behavior are presumably complex, random forests can help learn the decision boundaries for better performance (Bacham and Zhao 2017).

Results

Descriptive analysis

Table 4 summarizes the numerical data in a descriptive manner. The medians of the stock discussion variables differ significantly from their respective means, indicating highly skewed and non-normal distributions. Further evidence is provided by the corresponding skewness and kurtosis statistics. The shareholding variation variables are less skewed, with closer medians and means, but they continue to be non-normally distributed. We performed an Anderson–Darlin (AD) test on all variables, and the AD statistics show that the distributions of all variables are significantly non-normal.

Table 4 Descriptive statistics

Pertaining to post accumulations (PA_30, PA_60, and PA_90), the average number of posts 30, 60, and 90 days after the stock behavioral outcome date increased. In terms of post variations (PV_30, PV_60, and PV_90), the averages increased to roughly comparable levels between the 10th and 30th days, and between the 30th and 60th days from the stock behavioral outcome date. Similar trends can be found in replies (RA_30, RA_60, and RA_90), reply volume variations (RV_30, RV_60, and RV_90), likes (LA_30, LA_60, and LA_90), and like volume variations (LV_30, LV_60, and LV_90). There were more replies and less variation, as well as more likes and less “like” volume variation, further away from the stock behavioral outcome date. These outcomes are expected because there are more accumulated posts and rumor propagation about a stock over time. The large standard deviations and skewness indicate that these rumor variables vary significantly across stock-day trades.

Major and individual stockholders’ shareholdings varied less than their rumor counterparts. Corresponding computations for the 30th, 60th, and 90th days preceding a stock-day trading behavioral outcome show significantly lower standard deviations and skewness. This means that there were no significant changes in stock trading among major or individual stockholders during this time period.

A total of 200 records in the data set that are classified as General Stock. The data set contained 1164 records in biotechnology and medicine, 220 records in the cultural and creative industry, 2198 records in electronics manufacturing, 806 records in information and communications technology, 1178 records in other manufacturing, and 544 records in other manufacturing. In the previous year, there were 1664 records that experienced management shocks.

Predictive analysis

Table 5 displays the outcomes of three decision trees based on five evaluation metrics. Accuracy is frequently used as an aggregate measure to demonstrate how well a classification model correctly categorizes positive and negative outcomes. A model may outperform another in predicting one outcome over another. Sensitivity and specificity indicate how well the model predicts positive (abnormal trading) and negative (normal trading) outcomes. Precision measures the accuracy with which a model predicts a positive outcome. The Receiver Operator Characteristics (ROC) curve plots all possible combinations of the True Positive Rate versus the False Positive Rate for various classification cutoffs. The area under this curve (AUC) represents how well the classification model distinguishes between the two classes (i.e., abnormal vs. normal trading) and is one of the most widely used methods for comparing classification models (Kou et al. 2021). Table 6 shows the definitions and corresponding computations (Novaković et al. 2017). Overall, they predict abnormal trading fairly accurately, with accuracies ranging from 61.40 to 84.40%. Notably, performance improved across the board from Tree 1 (which only included shock and shareholding variables) to Tree 2 (which also included rumor propagation variables) to the boosted random forest with all variables.

Table 5 Performance of decision trees

Table 6 Performance metrics definitions

Financial characteristics tree (Tree 1)

Figure 3 depicts the decision tree for financial characteristics that predict abnormal trading behavior. The names of the predictor variables correspond to the notations in Table 1. The terminal nodes display the number of records in each outcome variable class. The tree predicts abnormal trading behavior reasonably well (accuracy = 0.61). It does a good job of distinguishing between the two classes (AUC = 0.63), but it does a better job of predicting normal trading behavior (sensitivity = 0.61, specificity = 0.67). In addition, its precision is only moderate (0.61), implying that using this tree to detect abnormal trading behavior is not ideal.

Fig. 3

Financial characteristics tree (Tree 1)

Table 7 summarizes the decision tree rules. Records that meet the rules specified in the Description column are likely to have the corresponding Outcome column result. The variation in stockholding of individual stockholders with less than or equal to 20 lots in the previous 30 days is the best predictor. Those with a value greater than 2.29 and less than or equal to − 1.03 are flagged for abnormal trading behavior (Rules #1 and #2), while those with a value greater than − 1.03 are subject to other predictors, including other shareholding variation predictors. It is worth noting that the electronics manufacturing industry contributes to the classification of abnormal trading behavior (Rules #6, #7, and #8). For example, in Rule #8, records in the electronics manufacturing industry are not flagged for abnormal trading behavior despite meeting the conditions related to shareholding variations.

Table 7 Decision rules for financial characteristics tree (Tree 1)

Combined tree (Tree 2)

Figure 4 depicts the decision tree with rumor propagation variables as predictors as well as all variables from Tree 1. Similarly, the names of the predictor variables correspond to the notations in Table 1. The terminal nodes display the number of records in each outcome variable class.

Fig. 4

Combined tree (Tree 2)

Overall, Tree 2 outperforms Tree 1 (accuracy = 0.70), representing an 8.79% improvement over the preceding tree. Its sensitivity, specificity, and precision have all improved, and it can predict both classes more accurately than Tree 1 (sensitivity = 0.70 and specificity = 0.70). Interestingly, despite being better positioned to distinguish the classes (AUC = 0.74), Tree 2 lacks management shock and shareholding variation variables. In classifying abnormal trading behavior, only accumulated “likes,” post accumulation, and post variations are relevant (Table 7).

Table 8 summarizes the Tree 2 decision tree rules. The number of “likes” in the previous 30 days is the best predictor. Following that, the splits are determined by the accumulation of posts and volume variations in posts over the previous 30 and 60 days, respectively. For example, among stock-day records with more than 61,766 “likes” in the previous 30 days and post variations of more than 23, those with less than 112,374 “likes” are not flagged for abnormal trading behavior (Rule #5), while those with more than 112,374 are flagged for abnormal trading behavior (Rule #6). According to these rules, the favorability, volume, and variations of CMoney posts, which appear in the classification rules, primarily in the preceding 2 months are sufficient to predict whether a stock will exhibit abnormal trading behavior. The rules also show that the relationships between these predictors and abnormal trading behavior are not linear, bolstering our decision tree’s use.

Table 8 Decision rules for combined tree (Tree 2)

Random forest (Tree 3)

The addition of rumor propagation variables improved the tree’s predictive ability. To improve prediction, we trained a random forest using the same variables as Tree 2. It shows a significant improvement in all aspects of prediction, as expected (Table 2). At 0.84 and 0.84, respectively, the accuracy and precision are higher than the preceding trees. Furthermore, the random forest can almost equally well classify both classes—abnormal and normal trading (sensitivity = 0.84 and specificity = 0.85). With an AUC ROC of 0.91, the random forest clearly distinguishes between the two classes. These findings imply that the random forest can be used effectively to distinguish abnormal stock trading behavior from financial rumors, and that ensemble models outperform single models (Kou et al. 2021).

Robustness checks

We performed robustness checks on our results using other machine learning techniques (cf. Kou et al. 2021), including k-Nearest Neighbors (kNN), logistic regression (Logistic) and support vector machines (SVM). Their performance metrics are compared to those of decision trees (Tree 1, financial characteristics variables, and Tree 2, combined) and random forest (RF, financial and RF combined). A summary of the comparison is shown in Fig. 5: The top chart compares all models that include only the financial and managerial characteristics variables (similar to Tree 1), while the bottom chart compares all models that include all variables (comparable to Trees 2 and 3). The comparisons are based on the models’ five performance metrics: accuracy, sensitivity, specificity, precision, and AUC ROC. Two findings emerged from the robustness tests. First, in all five models, the combined versions outperform their corresponding financial and managerial characteristics variables-only counterparts in terms of performance metrics. For example, the RF has the highest accuracy among the combined models, at 0.84, which is 11.05% higher than its financial characteristics-only counterpart, which has an accuracy of 0.76. The AUC ROC of the logistic regression comprising all variables (0.72) is 35.85% higher than its financial and managerial characteristics-only counterpart (0.53). Thus, financial rumors evidently have a greater influence on abnormal trading behavior than the financial and managerial characteristics variables included. Second, in both sets of comparisons, RFs outperform the other models. These are the strengths of using decision trees in financial predictions (cf. Barboza et al. 2017; Addo et al. 2018), which support our methodological rationale.

Fig. 5

Robustness checks

Conclusion and discussion

Summary

This study adopted an innovative approach using social media data (posts, likes, and replies) and decision tree induction to predict abnormal stock trading behavior in the context of trending data-driven approaches to studying financial contexts (Kou et al. 2019). In addition, we used the TWSE’s official legal definition of abnormal trading. We addressed two gaps in the literature: the lack of studies on the influence of rumors despite their relevance (Arif et al. 2016) and the lack of analyses in this area using machine learning despite its higher accuracy (Zhong and Enke 2019). We included variables related to rumor propagation, managerial shocks, and shareholding characteristics using data from September 2019 to August 2020 in our decision trees. Our results show that rumor propagation outperforms management shocks and other variables in predicting abnormal trading behavior. In particular, the extents of “likes” and post volume are critical. Furthermore, the random forest improves prediction in all aspects and can be readily used by academics, financial analysts, and governing bodies seeking to achieve and maintain market efficiency and stability. We performed robustness checks using logistic regression, k-Nearest Neighbors, and SVM; all of these methods produced similar results, but our decision trees outperformed them.

Rationality and the predictors of abnormal trading activities

New media content influences stock trading behavior (Li et al. 2018). False or unverified information can easily spread through these new media channels (Bondielli & Marcelloni 2019), strongly influencing investor behavior and thus stock prices (Clarke et al. 2020; Fong 2021). These may take the form of rumors (DiFonzo and Bordia 2007; Donovan 2007). Our findings support previous research that stock rumor propagation influences stock trading behavior (Anderson et al. 2021; Lyócsa et al. 2021). Furthermore, we demonstrate that rumor propagation can be used to predict whether a stock will exhibit abnormal trading behavior in the subsequent 2 months.

A multitude of variables have been shown to influence stock trading behavior, including firm strikes (Wisniewski et al. 2020) and changes in senior management (Yilmaz and Mazzeo 2014). Rational decisions based on these shocks reflect investor confidence losses. In contrast to the literature, our findings from management and other fundamental variables (Tree 1) indicate that the extent of shareholding variation is more important in predicting abnormal trading behavior.

Meanwhile, rumors are unverified (DiFonzo and Bordia 2007; Donovan 2007), and our results show that rumor propagation is far more influential in prediction. This demonstrates the irrationality of stock trading behavior, in which investors’ irrational emotions and errors in judgment can cause market volatility (Verma and Verma 2007). Our findings provide empirical explanations for GameStop’s recent short-term volatility, correlating with the heavy influence of stock rumors (Anderson et al. 2021; Lyócsa et al. 2021) on large volumes of irrational trading behavior.

Contributions and applications

Our study makes three contributions to the literature on abnormal trading behavior. First, we present empirical evidence that rumor propagation is far more important in predicting abnormal trading behavior than fundamental management shocks and traditional financial variables. This lends support to the notion of investor irrationality in stock trading and steers attention toward the overlooked rumor propagation in the literature on stock trading prediction. Second, we show that machine learning can be used to study financial outcomes, such as abnormal trading behavior. We chose a machine learning approach because it is more accurate than traditional statistical approaches (Barboza et al. 2017). Furthermore, it advances our understanding of how humans react to rumors via new technologies such as social media (Choudrie et al. 2021). The non-linearity of the resulting relationships offers additional methodological justifications (Antunes 2021). Third, our model is a direct response to calls for the Stock Exchange Commission to address stock market volatility (Anderson et al. 2021). This applies to other stock trading regulatory bodies.

Accurate, timely, and open information is essential for healthy stock market trading. With the advent of social media, stock discussions and rumors proliferate, interfering with investors’ decision-making and sentiments (Wang et al. 2020) and thus triggering irrational and abnormal behavior in volatile markets. Our results show that within 2 months, rumor propagation can be used to predict abnormal trading behavior. Because rumors can be monitored and measured, stock trading regulatory bodies can use our model to flag stocks for abnormal trading in advance, preserving market efficiency and controlling potential volatility. Financial analysts can also use our model to investigate stock discussions and trading behavior.

Future research directions

Although our study has many applications, three limitations should be addressed in future research. First, our data were restricted to the period between September 2019 and August 2020. The global pandemic may have influenced stock trading decisions during this time period. Future research can thus use the same approach to investigate rumor propagation and trading behavior in a different time frame. Second, we conducted the investigation in Taiwan, primarily because TWSE flags abnormal trading behavior and provides an authoritative means to define trading behavior. Future research could investigate stock trading in other countries and compare the impact of rumor propagation. Third, although decision trees are more stable than artificial neural networks (Addo et al. 2018), future research can explore into the differences between the two methods to determine the extent to which either one is superior. After identifying the predictors of abnormal trading behavior, future studies can then better design and train deep learning models based on these predictors for improved prediction performance. Overall, with additional research, the novel application of machine learning to analyze stock rumors and trading behavior has the potential to become a more established research program.