Gadjah Mada International Journal of Business
Vol. 27, No. 3 (September-December 2025): 297-319

Machine Learning vs. Human Investors:
Analyzing Adaptive Herding Behavior in U.S. Stocks vs.
Shariah-Compliant Stocks in Malaysia and Indonesia
Kok Loang Ooia*, Sevenpri Candrab
a
Universiti Malaya, Malaysia
b
BINUS University, Indonesia

Abstract: This study examined the effectiveness of machine learning models in capturing
adaptive herding behavior in the U.S., Malaysia, and Indonesia. Utilizing data from January 2010 to December 2023, the study incorporates market sentiment (Thomson Reuters MarketPsych indices), news sentiment (Bloomberg sentiment analysis), and investor
happiness measures (Hedonometer). The methodology employs both static and adaptive
herding analyses using the CSAD approach, enhanced by real-time sentiment analysis and
various machine learning models, including single- and multi-layer neural networks. The
results indicate significant differences in herding behavior across the three markets, with
machine learning models demonstrating superior performance in capturing herding behavior and faster normalization after major macroeconomic events than traditional methods. These findings highlight the potential of machine learning models to challenge the
static assumptions of the efficient market hypothesis and provide insights for designing
better trading algorithms by considering the impact of market sentiment, news sentiment,
and investor happiness.
Keywords: machine learning, adaptive herding, market sentiment, financial stability,
stock returns
JEL Classification: G02, G11, G14, G15

*Corresponding author’s e-mail: markooi@um.edu.my
ISSN: PRINT 1411-1128 | ONLINE 2338-7238
https://journal.ugm.ac.id/gamaijb

297

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

Introduction

Herding behavior, a well-documented phenomenon in financial markets (Bikhchandani
et al., 1992), occurs when investors mimic the trades of others rather than rely on their
own information or analysis. This behavior can lead to significant market inefficiencies,
contributing to the formation of bubbles and subsequent crashes. Traditional studies have
primarily focused on static measures of herding that do not account for the dynamic nature of financial markets (Chang et al., 2000). Static herding assumes constant behavior
over time, ignoring the evolving nature of market sentiment and external information. By
contrast, adaptive herding captures the changing nature of investor behavior in response
to market conditions, sentiment, and external information. Despite its importance, adaptive herding is underexplored, primarily because of the complexity and computational intensity required for real-time sentiment analysis and the integration of dynamic variables,
such as market sentiment, news sentiment, and investors' happiness (Choi & Park, 2023;
Hwang & Salmon, 2004).
Machine learning (ML) offers several advantages for studying herding behavior
and stock market performance. ML algorithms can process vast amounts of data and uncover patterns that may not be apparent using traditional statistical methods (Henrique et
al., 2019). Moreover, ML models can enhance the accuracy of stock price predictions by
incorporating a wide range of variables, including technical indicators, sentiment scores,
and macroeconomic factors (Amanda & Pradipta, 2024; Bollen et al., 2011; Gao et al.,
2024; Tetlock, 2007). A fundamental assumption is that machine learning models, being
purely algorithmic, should not exhibit behavioral biases. However, the data used to train
these models often reflect human behaviors and biases, raising questions about whether
ML models can exhibit herding behavior similar to that of human investors (Feng et al.,
2019). If machine learning can eliminate these behavioral biases, it could potentially offer
a more reliable and unbiased approach to trading, outperforming human investors, who
are prone to irrational herding (Zhang et al., 2011).
While ML models have shown promise in stock prediction, their efficacy in down
markets remains uncertain. Downward markets present unique challenges that existing
ML models may not fully capture, necessitating further investigation into their performance under these conditions (Bihari et al., 2025; Henrique et al., 2019; Kumbure et al.,
2022; Soni et al., 2022). Additionally, there is a significant gap in understanding how algorithmic trading systems interact with herding dynamics, particularly adaptive herding,
which is affected by market sentiment and news. While behavioral finance extensively
examines human biases, the potential of machine learning models to replicate or mitigate
these biases remains underexplored (Barberis & Thaler, 2003).
Herding and investor performance are affected by market efficiency, as explained
by the efficient market hypothesis (EMH). According to EMH, markets are efficient if
asset prices fully reflect all available information (Fama, 1970). However, the level of information efficiency can vary significantly between markets, which in turn affects the
presence and nature of herding behavior. Developed markets, such as the United States,
typically exhibit high levels of information efficiency and robust regulatory frameworks,
ensuring that information is quickly and accurately reflected in asset prices. In contrast,
developing markets, such as Malaysia and Indonesia, are characterized by rapid economic
growth, evolving regulatory environments, and diverse investor compositions, which can
lead to information inefficiencies (Bollen et al., 2011; Gao et al., 2024; Oliveira et al., 2016;
298

Ooi & Candra

Tetlock, 2007).
Although herding behavior has been extensively investigated across developed and
emerging markets (Bikhchandani et al., 1992; Chang et al., 2000; Chiang and Zheng, 2010;
Mahadwartha et al., 2023), there remains a paucity of empirical research on how ethical
investment principles, particularly those underpinning Shariah-compliant equities, shape
such behavior within different market structures. Shariah-compliant stocks, which constitute a significant portion of publicly listed firms in Malaysia and Indonesia, are governed
by Islamic financial principles that prohibit excessive leverage, interest-based transactions
(riba), and speculative activities (gharar), thereby imposing a unique layer of ethical and
financial discipline (Medhioub and Chaffai, 2019; Aziz et al., 2022; Nugroho & Pratiwi,
2023). Prior studies suggest that these constraints may mitigate irrational investor behavior, reducing the propensity for speculative herding (Loang and Ahmad, 2022; Maulida
and Sari, 2023). However, the dual governance model of conventional regulation alongside Shariah oversight may introduce distinct behavioral dynamics that are not observable
in conventional equity markets. Despite the growing prominence of Islamic finance, limited attention has been paid to whether these structural and normative differences influence
investor herding, particularly during periods of heightened market uncertainty.
These differences in market efficiency necessitate a detailed examination of how
herding behavior manifests across various economic contexts. Moreover, understanding
the role of adaptive herding requires analyzing dynamic variables, such as market sentiment, news sentiment, and investor happiness (Gopal & Loang, 2024). Existing literature
has largely focused on static herding measures, leaving a gap in comprehensive studies
that incorporate real-time sentiment analysis (Chang et al., 2000; Chiang & Zheng, 2010).
Additionally, there is a need to explore whether machine learning models can effectively
capture and predict these adaptive behaviors, particularly during market downturns when
traditional models may falter (Henrique et al., 2019; Kliegr et al., 2021; Kumbure et al.,
2022; Soni et al., 2022). The potential of machine learning to eliminate behavioral biases
and outperform human investors presents another critical gap, as current studies have yet
to fully explore this capability (Athota et al., 2023; Feng et al., 2019; Soni et al., 2022).
This study investigates adaptive herding and machine learning models' stock return predictions during down markets. Three markets—the U.S., a mature market; Malaysia, a developing market with unique regulatory and market characteristics; and Indonesia, a rapidly increasing emerging market—are studied. This study measures herding
behavior in these markets and the impact of adaptive herding on market dynamics using
the Thomson Reuters MarketPsych indices for market sentiment, the Bloomberg sentiment analysis for news sentiment, and the Hedonometer for Twitter feed data on investor
happiness. This study could reveal algorithmic trading system capabilities and limitations.

Literature Review

The EMH, introduced by Fama (1970), posits that asset prices reflect all available information, making it impossible to consistently outperform the market through stock selection
or timing. While the EMH has been extensively validated in developed markets such as
the United States, emerging markets such as Malaysia and Indonesia often display deviations due to less stringent regulatory frameworks and slower information dissemination.
Recent developments have challenged the static nature of the EMH. Lo (2004) proposes
299

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

the adaptive market hypothesis, suggesting that market efficiency evolves with changing
market environments, supported by empirical evidence from Noda, (2016) and Urquhart
& McGroarty (2016), indicating fluctuating market efficiency over time and conditions.
Machine learning has emerged as a transformative tool in financial market analysis, capable of processing vast amounts of data and identifying patterns that traditional statistical
methods may overlook. Fischer and Krauss (2018) demonstrate that ML models, including
neural networks and support vector machines, outperform conventional models in predicting stock prices and managing financial risks. Gu et al. (2020) apply ML techniques to
financial econometrics, showcasing improved predictive accuracy in asset pricing, while
Krauss et al. (2017) use ML to detect financial misstatements, highlighting ML's potential
to enhance understanding of financial anomalies and market behaviors. However, the application of ML in assessing market efficiency and predicting market downturns remains
underexplored, particularly in emerging markets.
The sentiment analysis further enhanced the predictive power of the ML models.
By incorporating textual data from news articles, social media, and financial reports, sentiment analysis provides a more comprehensive understanding of market sentiment and
its impact on asset prices (Loang, 2025). Oliveira et al. (2016) illustrate that integrating
social media sentiment into ML models improves market predictions. Similarly, previous studies (Hu et al., 2021; Schmeling, 2009; Tetlock, 2007) demonstrate that combining
news sentiment with traditional financial indicators significantly enhances the accuracy
of stock market forecasts. Additionally, Zhang et al. (2011) use sentiment analysis of financial news to predict market volatility, further highlighting the importance of real-time
data in financial predictions. Nevertheless, most existing research focuses on static models, lacking comprehensive studies that leverage real-time data to dynamically understand
market efficiency. This gap is critical, as dynamic factors, such as market sentiment and
news sentiment, significantly impact market movements.
Although ML models have advanced the accuracy of financial forecasting, their
performance during market downturns remains uncertain, particularly in contexts where
behavioral biases such as herding are more pronounced (Dantas and Cyrino Oliveira,
2018; Kumbure et al., 2022; Tsai and Hsiao, 2010). While behavioral finance has extensively explored human-driven herding, it is less clear whether ML systems trained on
human data replicate or correct these patterns (Kliegr et al., 2021). Furthermore, most
ML applications have been developed in developed markets like the U.S., with limited
testing in emerging or ethically constrained environments (Zakamulin and Giner, 2020).
In markets such as Malaysia and Indonesia, where Shariah-compliant stocks must adhere
to both conventional regulations and Islamic ethical criteria, investor behavior may differ significantly due to stricter screening, long-term orientation, and reduced speculative
trading (Medhioub and Chaffai, 2019; Aziz et al., 2022; Loang and Ahmad, 2022). Given
these institutional and behavioral differences, it is plausible that herding dynamics in Shariah-compliant stocks diverge from those in mature markets.
H1: There is a significant difference in herding behavior between U.S. stocks and
Shariah-compliant stocks in Malaysia and Indonesia, with herding being
more pronounced in Shariah-compliant stocks due to market inefficiencies
and lower regulatory stringency.
H2: Machine learning models exhibit superior performance in predicting herding behavior compared to traditional statistical models, particularly during
market downturns.
300

Ooi & Candra

Herding behavior in financial markets, where investors mimic the trades of others
rather than rely on their own analysis, is well documented (Bikhchandani et al., 1992).
Traditional studies often employ static measures of herding, which fail to account for
the dynamic nature of investor behavior and market conditions (Chang et al., 2000). On
the other hand, adaptive herding considers changes in investor behavior in response to
evolving market conditions and external information (Gopal & Loang, 2024). Hwang and
Salmon (2004) highlight the importance of adaptive herding, showing that it provides a
more accurate representation of market movements by incorporating real-time sentiment
analysis and other dynamic variables. Recent studies have further confirmed the significance of adaptive herding, especially during periods of market stress, such as the COVID-19 pandemic, in which herding behaviors intensified in both developed and emerging
markets (Mnif et al., 2020; Omane-Adjepong et al., 2021; Youssef & Waked, 2022).
Market sentiment refers to investors' overall attitudes towards a particular security or financial market. It is often gauged using various indicators, including surveys and
sentiment indices. Baker and Wurgler (2007) demonstrate that high levels of investor sentiment are associated with overpriced stocks and subsequent lower returns. Recent advancements have integrated market sentiments into predictive models to enhance their
accuracy. Da et al. (2015) have developed a novel measure of market sentiment using
Google search volumes and found that it significantly predicted short-term market returns. Similarly, Schmeling (2009) analyzes sentiment across different countries and concludes that sentiment significantly affects stock returns, especially in markets with lower
institutional investor presence. Despite these advancements, there is still a need for more
comprehensive studies that incorporate real-time market sentiment data into adaptive
herding models (Chiang & Zheng, 2010; Goodell, 2020).
News sentiment analysis involves evaluating the tone of news articles in order to
gauge the sentiment of the information being disseminated. This form of analysis has been
shown to significantly affect investor behavior and market outcomes. Tetlock (2007) finds
that high pessimism in news articles predicts downward pressure on market prices. Similarly, Engelberg and Parsons (2011) demonstrate that local news sentiment significantly
impacts local trading behavior and asset prices. Incorporating news sentiment into financial models improves their predictive power. Oliveira et al. (2016) show that integrating
news sentiments with machine learning models enhances market predictions. Nonetheless, the dynamic incorporation of news sentiment into adaptive herding models remains
underexplored, particularly in various market contexts (Chang et al., 2000; Chiang &
Zheng, 2010).
Investor happiness, often measured through social media sentiment and happiness
indices, is an emerging area of interest in behavioral finance. Happy investors are more
likely to make optimistic market decisions that influence their trading volumes and asset
prices. Dodds et al. (2011) use Twitter data to construct a happiness index, finding that it
correlates with stock market movements. Similarly, Zhang et al. (2011) find that positive
sentiment on social media is associated with higher stock returns. Nguyen and Vo (2023)
demonstrate that investor mood, as captured by daily stock message board postings, significantly predicts market performance. Despite these findings, there is a lack of comprehensive studies that dynamically integrate investor happiness into models of adaptive
herding, which could provide deeper insights into the psychological drivers of market
behavior (Bogdan et al., 2022). Therefore, this study proposes the following hypotheses:
301

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

H3: Market sentiment, news sentiment, and investor happiness significantly
impact herding behavior, with positive sentiment associated with reduced
herding and negative sentiment with increased herding.
H4: Machine learning models can better mitigate behavioral biases such as herding, leading to faster normalization of market conditions post-major macroeconomic events compared to human investors.
Shariah-compliant stocks operate under a unique set of ethical and financial
guidelines that distinguish them from conventional equities. These guidelines, which include restrictions on excessive leverage, interest-based transactions (riba), and investment
in unethical industries (such as gambling and alcohol), contribute to a more disciplined
investment environment (Medhioub and Chaffai, 2019). Due to these principles, Shariah-compliant firms are subject to rigorous corporate governance and transparency standards, which may reduce speculative trading and limit irrational herding behavior (Loang
and Ahmad, 2022). Recent studies suggest that Shariah-compliant stocks exhibit lower
levels of herding compared to conventional stocks, particularly during periods of market
volatility, as ethical screening criteria lead to more stable investor behavior (Aziz et al.,
2022). Furthermore, religious and ethical motivations among Islamic investors may foster long-term investment strategies rather than short-term speculative behavior, further
diminishing herding tendencies (Maulida and Sari, 2023). Islamic financial principles can
mitigate irrational investor behavior (Ah Mand et al., 2023). Given these factors, this study
proposes the following hypothesis:
H5: Shariah-compliant stocks exhibit lower levels of herding behavior than conventional stocks due to stricter corporate governance, ethical investment
constraints, and a more disciplined investor base.

Methods

This study examines adaptive herding in U.S. (1,742 firms), Malaysia (731 companies),
and Indonesia (693 companies) stock markets using market data and sentiment analysis.
Data from U.S., Malaysia, and Indonesia S&P 500 Top 50 firms was used for machine
learning research. This study chose Malaysia and Indonesia because they are the largest
markets for Shariah-compliant stocks, which cater to Islamic investors. Malaysia is one of
the world's major Islamic financial marketplaces, with 79% of publicly listed businesses,
or 1,148 out of 1,457 on Bursa Malaysia, being Shariah-compliant. Of the 793 firms registered on the Indonesia Stock Exchange, 468 exceed Shariah criteria, representing over
60% of the market.
The U.S. stock market, as a highly developed financial system, is characterized by
stringent regulatory oversight, high transparency, and robust enforcement mechanisms
that minimise information asymmetry and speculative trading. In contrast, Malaysia and
Indonesia represent emerging markets where financial regulation is evolving, and market
inefficiencies can contribute to different patterns of herding behavior. Shariah-compliant
stocks in these markets are subject to dual regulatory frameworks—conventional financial
regulations and additional Islamic ethical constraints. While Shariah governance requires
compliance with ethical investment screening and financial ratio restrictions, it primarily focuses on business activity screening rather than direct market conduct regulation.
302

Ooi & Candra

Therefore, while these stocks face additional ethical requirements, they may not necessarily experience the same level of financial regulatory enforcement as firms in the U.S.
The data covers major market events including the COVID-19 epidemic from
January 2010 to December 2023. The daily stock prices and trade volumes came from
Bloomberg. Table 1 provides the Thomson Reuters MarketPsych market sentiment indices, the Bloomberg sentiment analysis news sentiment ratings, and the Twitter-based
Hedonometer for investor happiness.
Table 1. Variables and Data Sources
Variables
Descriptions
Herding Behavior Calculated using the CSAD
Market Sentiment Indices reflecting investor sentiment
based on financial news and social media
News Sentiment
Sentiment scores derived from analysis of
financial news articles
Investor Happiness Happiness index derived from Twitter
data analysis
Trading Volume
Daily trading volume of the selected
companies
Source: Authors' compilation

Source
DataStream
Thomson Reuters
MarketPsych Indices
Bloomberg Sentiment
Analysis
Hedonometer
DataStream

This study employs the cross-sectional absolute deviation (CSAD) method to
measure static herding behavior in financial markets. The CSAD approach captures the
extent to which individual stock returns deviate from the market average, offering a
straightforward measure of herding behavior. The CSAD at time t is calculated using the
following formula:
(1)
In this equation, CSADt represents the cross-sectional absolute deviation at time
t; NtN_tNt is the number of stocks in the market at time t; Ri,tdenotes the return of stock
i at time t; and Rm,t is the market return at time t. To identify herding behavior, the CSAD
is regressed on the market return and its squared term as specified in the following regression model:
(2)
Here, Rm,t denotes the market return at time t; |Rm,t| represents the absolute value of
the market return at time t; R2m,t is the squared market return; and ao, a1, a2 are the coefficients to be estimated. The error term ϵt captures the unexplained variation in the CSAD.
A negative and significant coefficient a2 indicates the presence of herding behavior, suggesting that individual stock returns converge towards the market return when market
returns are extreme. Adaptive herding, unlike static herding, accounts for the dynamic
nature of investor behavior by considering changes in market conditions and external
information. This study enhances the traditional CSAD model by integrating real-time
sentiment analysis and other dynamic factors, such as market sentiment, news sentiment,
303

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

investor happiness, market volatility, and trading volume. The enhanced CSAD model is
formulated as follows:
(3)
In this model, MSentt represents the aggregated market sentiment score at time
t; Newst denotes the sentiment score derived from news articles at time t; Happt captures
the investor happiness index based on Twitter data at time t; and Volt represents the daily
trading volume at time t.
The enhanced CSAD model is estimated using regression analysis, where the significance of the interaction term β2 indicates the presence of adaptive herding, suggesting
that herding behavior changes with sentiment. This study compares static and adaptive
herding behaviors across the U.S., Malaysia, and Indonesia, analyzing the impact of sentiment on herding separately for each market to understand the differences in herding
dynamics.
A single-layer neural network (SLNN) was employed to understand the impact of
market variables on herding behavior. The SLNN consists of an input layer and an output
layer, which facilitates straightforward data processing while maintaining computational
efficiency. In this study, the input layer encompasses variables, such as market returns,
market sentiment, news sentiment, investor happiness, and trading volume. These features were selected based on their significant impact on herding behavior.
The model calculates the output using the following equation:
(4)
where μi denotes the output of the i-th neuron, wj represents the weights, xj are the
input features, and b is the bias term. The input features, represented as x=(xi1, xi2, xi3,…,
xin)x, are connected to the output layer through weights. The neuron outputs are computed as follows:
(5)
The SLNN underwent training using a learning rate of 0.01 and an array size of
4, with weights initialised via the Glorot uniform method and optimised using the Adam
weight optimiser. Training involves forward propagation to generate outputs, calculating
the loss using the mean squared error (MSE), and performing backpropagation to update weights. This iterative process continues until the model achieves a minimal error. In
the context of herding behavior, the SLNN leverages historical data on market sentiment,
news sentiment, investor happiness, and trading volume to identify periods characterized
by herding or its absence, as indicated by the CSAD.
A multi-layer neural network (MLNN) extends the SLNN architecture by incorporating one or more hidden layers between input and output layers. This enhanced architecture allows the network to capture more complex patterns and interactions within
the data, which is crucial for a nuanced understanding of herding behavior. Each layer in
the MLNN comprises neurones interconnected with every neuron in the preceding and
succeeding layers, thereby enabling the model to comprehend intricate dependencies.
304

Ooi & Candra

The hidden layers utilise the rectified linear unit (ReLU) activation function, defined as:
(6)

The output for each neuron in the hidden layers is computed using:
(7)
where ∅ denotes the ReLU activation function, wi represents the weights, ai is the
input, and b is the bias term. The fully connected architecture ensures that each hidden
neuron is linked to all other nodes or neurones in the succeeding layer with weights assigned to these connections. The neuron's output value ∑wi ai + b reflects the estimated
CSAD. The loss function utilised for training the MLNN is the MSE, expressed as:
(8)
where N represents the total number of instances in the dataset, yi is the actual
CSAD value, and yi is the forecasted CSAD value. This equation ensures that the model
minimises the average squared difference between the actual and predicted CSAD values
during the training.
At the end of all epochs, the final coefficients are saved and can be used for unobserved data (i.e. data used for testing) in the future by using absolute market returns, the
square of market returns, and economic uncertainty news sentiment to calculate the value
of CSAD. This approach ensures that the MLNN model is robust and capable of generalising new, unseen data, thereby providing a comprehensive understanding of herding
behavior across different market conditions. Data supporting this study are available from
the Repository at https://doi.org/10.5281/zenodo.15266086

Results & Discussions
Descriptive Statistics

Table 2 presents the descriptive statistics of the variables from January 2010 to December
2023. The U.S. market exhibited a mean CSAD of 0.015 with a standard deviation of 0.023,
indicating relatively lower herding behavior and variability. In contrast, Malaysia's CSAD
is higher, with a mean of 0.018 and a standard deviation of 0.027, suggesting more pronounced herding behavior and greater market instability. Indonesia's market shows the
highest levels of herding behavior, with a mean CSAD of 0.019 and a standard deviation of
0.031, reflecting significant market volatility and investor conformity. The skewness values
for the U.S., Malaysia, and Indonesia are 1.485, 1.597, and 1.782, respectively, indicating a
rightward skew, with Indonesia's distribution being the most asymmetric. Kurtosis values
further differentiate these markets, with the U.S. at 5.025, Malaysia at 5.215, and Indonesia
at 5.472, all indicating leptokurtic distributions (due to the pandemic) but with varying
degrees of peakedness, most pronounced in Indonesia.
305

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

Table 2. Descriptive Statistics of Variables (January 2010 - December 2023)
Variable

Obs.

Mean

Max

Min

Std Dev Skewness

Kurtosis

CSAD_US
3,528
CSAD_MY
3,528
CSAD_IN
3,528
Market Sentiment
3,528
News Sentiment
3,528
Investor Happiness
3,528
Trading Volume (' 000)
3,528
Source: Authors' compilation

0.015
0.018
0.019
0.502
0.405
0.547
1,203

0.123
0.132
0.147
0.978
0.902
0.953
5,012

0.004
0.006
0.007
0.103
0.052
0.201
202

0.023
0.027
0.031
0.198
0.248
0.183
798

5.025
5.215
5.472
2.751
3.015
2.845
4.195

1.485
1.597
1.782
0.052
0.215
-0.298
1.247

Market sentiment averages 0.502 with a standard deviation of 0.198, whereas news
sentiment has a mean of 0.405 and a standard deviation of 0.248, indicating moderate
variability in both metrics. Investor happiness averages 0.547 with a standard deviation of
0.183, reflecting relatively stable investor sentiment. Trading volume, expressed in thousands, has a mean of 1,203, a maximum of 5,012, a minimum of 202, and a standard deviation of 798, showing substantial trading activity variability.

Correlation Analysis

The Pearson correlation analysis presented in Table 3 reveals significant relationships between the variables under study. The U.S. (CSAD_US) exhibits a strong positive correlation with the CSAD for Malaysia at 0.752 and Indonesia at 0.689, indicating that herding
behaviors are closely related across these markets. Similarly, CSAD_MY and CSAD_IN
have an even higher correlation of 0.814, suggesting that herding behavior is particularly
interconnected between these two emerging markets.
Table 3. Correlation Analysis
Variable

Obs.

Mean

Max

Min

Std Dev Skewness

Kurtosis

CSAD_US
CSAD_MY
CSAD_IN
Market Sentiment
News Sentiment
Investor Happiness
Trading Volume (' 000)

3,528
3,528
3,528
3,528
3,528
3,528
3,528

0.015
0.018
0.019
0.502
0.405
0.547
1,203

0.123
0.132
0.147
0.978
0.902
0.953
5,012

0.004
0.006
0.007
0.103
0.052
0.201
202

0.023
0.027
0.031
0.198
0.248
0.183
798

5.025
5.215
5.472
2.751
3.015
2.845
4.195

1.485
1.597
1.782
0.052
0.215
-0.298
1.247

Note: Market Sentiment (MS), News Sentiment (NS), Investor Happiness (IH), Trading Volume (TV)

Source: Authors' compilation

Market Sentiment (MS) is negatively correlated with CSAD across all three markets, with values of -0.325, -0.298, and -0.354 for the U.S., Malaysia, and Indonesia, respectively, indicating that higher market sentiment is associated with reduced herding.
News Sentiment (NS) also shows a negative correlation with CSAD, with correlations of
-0.401 for the U.S., -0.422 for Malaysia, and -0.389 for Indonesia, suggesting that positive
news sentiment helps mitigate herding behavior. Investor Happiness (IH) follows a similar
trend, negatively correlating with CSAD values at -0.276, -0.312, and -0.337 for the U.S.,
306

Ooi & Candra

Malaysia, and Indonesia, respectively, indicating that higher levels of investor happiness
are linked to lower herding behavior. Conversely, Trading Volume (TV) is positively correlated with CSAD, with correlations of 0.512 for the U.S., 0.468 for Malaysia, and 0.495
for Indonesia, suggesting that increased trading activity is associated with higher herding
behavior in these markets.

Existence of Static Herding in the U.S., Malaysia, and Indonesia

Table 4 estimates Equation (2), the static CSAD regression model, which tests for herding
behavior based on the nonlinear relationship between return dispersion and market returns for the U.S., Malaysia, and Indonesia across the pre-, pandemic, and post-pandemic
regimes. During the pre-pandemic period (January 2010 to March 2020), all three markets
exhibited strong evidence of herding behavior, with significantly negative a2 coefficients
(-0.012, -0.019, and -0.017 for the U.S., Malaysia, and Indonesia, respectively), indicating
a convergence of individual stock returns towards the market return. This suggests that investor behavior is driven by collective market trends rather than by individual stock fundamentals. The pandemic period (March 2020 to December 2022) saw an intensification
of herding behavior, reflected in more significant negative a2 values (-0.021 for the U.S.,
-0.026 for Malaysia, and -0.023 for Indonesia). This heightened herding can be attributed
to the increased uncertainty and market volatility during the COVID-19 crisis, which led
investors to rely more heavily on market signals than on independent analysis.
In the post-pandemic period (January 2023 to December 2023), while herding
behavior persisted, there was a slight attenuation compared to the pandemic regime, with
a2 values of -0.015, -0.020, and -0.019 for the U.S., Malaysia, and Indonesia, respectively.
This reduction in herding intensity may reflect a gradual return to more normal market
conditions and a recovery in investor confidence, although collective behavior remains
significant. Comparing these findings with recent studies, our results align with those of
Bouri et al. (2019), who found that herding behavior intensified globally during the COVID-19 pandemic due to heightened market uncertainties (Bouri et al., 2019). Similarly,
Nguyen and Vo (2023) documented persistent herding during the pandemic, emphasizing
the role of investor sentiment and market liquidity in driving collective behavior.
Figures 1(a) and 1(b) depict the CSAD values obtained through non-ML and machine learning. In the non-machine learning approach (Figure 1a), CSAD values exhibit
higher levels of dispersion, indicating more pronounced herding behavior among investors. For instance, during the pre-pandemic period (January 2010 to March 2020), the
average CSAD values for the U.S., Malaysia, and Indonesia were 0.015, 0.018, and 0.019,
respectively. These values surged during the pandemic (March 2020 to December 2022) to
0.021 in the U.S., 0.026 in Malaysia, and 0.023 in Indonesia.
Conversely, the machine learning approach (Figure 1b) showed lower overall
CSAD values, suggesting that herding behavior is less pronounced when advanced analytical techniques are employed. During the same pre-pandemic period, the machine
learning-based CSAD values were 0.013 for the U.S., 0.016 for Malaysia, and 0.017 for
Indonesia. Notably, the machine learning-based CSAD also spiked during major macroeconomic events such as the COVID-19 pandemic, the 2015 Chinese stock market crash,
and the 2016 Indonesian demonetization, reflecting the immediate impact of these disruptions on investor behavior. However, these surges were quickly mitigated, with CSAD
values returning to normal herding levels faster than those in the non-machine learning
307

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

approach.
The CSAD during the pandemic surged higher than that in non-machine learning at the beginning of the pandemic, as machine learning models initially struggled to
account for unexpected events that disrupted the market. The results of this study indicate that machine learning models, despite their initial challenges in accounting for unexpected events such as the COVID-19 pandemic, ultimately provide a more accurate
and resilient measure of herding behavior compared to traditional methods. This finding
aligns with previous research, highlighting the superiority of machine learning techniques
in financial market analysis. Gu et al. (2020) demonstrated that machine-learning models significantly improve predictive accuracy in asset pricing compared to conventional
econometric models. Similarly, Fischer and Krauss (2018) found that neural networks
and other machine-learning approaches outperform traditional models in forecasting
stock prices and managing financial risk. The rapid normalization of CSAD values in our
machine learning models post-major macroeconomic events emphasized the potential of
algorithmic trading systems to mitigate behavioral biases and enhance market stability.
However, the reason why machine learning shows less herding and better normalization is
unknown, highlighting the need to conduct adaptive herding analyses to further examine
this phenomenon.

Existence of Static Herding in the U.S., Malaysia, and Indonesia

Table 5 highlights the results of the adaptive herding analysis using rolling regression
methods to compare machine learning and non-machine learning models in the U.S.,
Malaysia, and Indonesia with 250-day and 500-day rolling windows. The use of 250-day
and 500-day rolling windows allows for the continuous updating and analysis of herding
behavior over time, providing a more accurate and dynamic understanding of how market conditions and investor behaviors evolve rather than relying on static, single-period
observations.
Table 4. Existence of Static Herding in U.S., Malaysia, and Indonesia
Regime I: Pre-Pandemic
No of Obs
Constant
a_1
a_2
R-squared
Adjusted R-squared
Log-likelihood
Hannan-Quinn
criterion
F-statistic
Prob(F-statistic)
Durbin-Watson stat
Wald F-statistic
Prob(Wald F-statistic)

Regime II: Pandemic

Regime III: Post-Pandemic

U.S.

Malaysia

Indonesia

U.S.

Malaysia

Indonesia

U.S.

Malaysia

Indonesia

2,583
0.010***
(0.000)
0.240***
(0.000)
-0.012**
(0.045)
0.624
0.605
150.25
4.954

2,592
0.014***
(0.000)
0.255***
(0.000)
-0.019**
(0.040)
0.673
0.651
149.15
4.881

2,581
0.011***
(0.000)
0.260***
(0.000)
-0.017**
(0.043)
0.654
0.631
150.10
4.862

693
0.018***
(0.000)
0.285***
(0.000)
-0.021***
(0.004)
0.714
0.691
145.35
5.804

687
0.019***
(0.000)
0.295***
(0.000)
-0.026***
(0.003)
0.747
0.720
143.50
5.525

695
0.017***
(0.000)
0.290***
(0.000)
-0.023***
(0.003)
0.734
0.717
144.30
5.991

252
0.012***
(0.000)
0.265***
(0.000)
-0.015**
(0.028)
0.667
0.649
148.20
5.005

249
0.016***
(0.000)
0.275***
(0.000)
-0.020**
(0.022)
0.681
0.665
146.80
5.042

252
0.013***
(0.000)
0.270***
(0.000)
-0.019**
(0.025)
0.697
0.675
147.90
5.226

22.35
0.000
2.05
21.50
0.000

21.70
0.000
2.02
20.80
0.000

22.00
0.000
2.04
21.30
0.000

24.20
0.000
2.10
23.00
0.000

25.00
0.000
2.12
24.30
0.000

24.50
0.000
2.11
24.00
0.000

23.50
0.000
2.07
22.50
0.000

24.10
0.000
2.09
23.70
0.000

23.70
0.000
2.08
23.40
0.000

Note: Pre-pandemic: January 2010 to March 2020. Pandemic: March 2020 to December 2022. Post-pandemic: January 2023 to December 2023. Significance levels: **, *** indicate significance at the 5%, and 1% levels,
respectively.

Source: Authors' compilation
308

Ooi & Candra

Figure 1(a). CSAD for U.S., Malaysia, and Indonesia (Non-Machine Learning)

Figure 1(b). CSAD for U.S., Malaysia, and Indonesia (Machine Learning)
Note: Figure 1 compares the CSAD values for the U.S., Malaysia, and Indonesia from January 2010 to December 2023 using non-machine learning (Figure 1a) and machine learning methods (Figure 1b). The machine
learning approach shows lower CSAD values and faster normalization after major macroeconomic events,
indicating a better adaptation to dynamic market conditions.

Source: Authors' compilation

309

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

Country
USA

Malaysia

310

Table 5. Adaptive Herding in U.S., Malaysia, and Indonesia
Variable
250_ML 500_ML 250_
500_
NoML
NoML
a0
1.231***
1.452***
3.123***
3.457***
(0.001)
(0.001)
(0.007)
(0.006)
a1
2.734***
3.154***
2.513***
2.845***
(0.001)
(0.001)
(0.007)
(0.006)
a3
-8.526*** -10.038*** -11.825*** -13.416***
(0.001)
(0.001)
(0.009)
(0.008)
MS
1.239***
1.451***
1.126***
1.341***
(0.001)
(0.001)
(0.008)
(0.007)
NS
-1.238*** -1.454*** -1.128*** -1.344***
(0.002)
(0.002)
(0.009)
(0.008)
IH
1.453***
1.673***
1.342***
1.563***
(0.001)
(0.001)
(0.010)
(0.009)
TV
2.345***
2.567***
2.231***
2.456***
(0.001)
(0.001)
(0.011)
(0.010)
R-squared
0.654
0.683
0.635
0.671
Adjusted R-squared
0.643
0.674
0.624
0.662
Log-likelihood
110.113
120.234
105.456
115.789
Hannan-Quinn criterion 5.256
4.112
5.124
4.987
F-statistic
108.623
670.823
102.412
640.345
Prob (F-statistic)
0.000
0.000
0.000
0.000
Durbin-Watson stat
1.473
1.534
1.492
1.523
Wald F-statistic
110.345
465.012
105.345
480.234
Prob (Wald F-statistic)
0.000
0.000
0.001
0.002
a0
1.334***
1.562***
3.342***
3.567***
(0.002)
(0.001)
(0.007)
(0.006)
a1
2.536***
3.032***
2.515***
2.913***
(0.001)
(0.001)
(0.007)
(0.006)
a3
-9.236*** -11.341*** -12.936*** -14.613***
(0.002)
(0.002)
(0.009)
(0.008)
MS
1.348***
1.564***
1.235***
1.451***
(0.001)
(0.001)
(0.008)
(0.007)
NS
-1.341*** -1.563*** -1.235*** -1.451***
(0.002)
(0.002)
(0.009)
(0.008)
IH
1.564***
1.783***
1.451***
1.673***
(0.001)
(0.001)
(0.010)
(0.009)
TV
2.456***
2.671***
2.345***
2.567***
(0.001)
(0.001)
(0.011)
(0.010)

Ooi & Candra

R-squared
Adjusted R-squared
Log-likelihood
Hannan-Quinn criterion
F-statistic
Prob (F-statistic)
Durbin-Watson stat
Wald F-statistic
Prob (Wald F-statistic)
Indonesia a0
a1
a3
MS
NS
IH
TV
R-squared
Adjusted R-squared
Log-likelihood
Hannan-Quinn criterion
F-statistic
Prob (F-statistic)
Durbin-Watson stat
Wald F-statistic
Prob (Wald F-statistic)

0.667
0.603
0.652
0.629
0.654
0.591
0.639
0.617
112.345
125.512
108.123
118.789
5.278
4.056
5.321
4.098
110.512
680.623
123.612
632.133
0.001
0.002
0.002
0.002
1.514
1.554
1.523
1.549
115.234
475.345
112.456
470.812
0.001
0.002
0.001
0.002
1.453***
1.673***
3.453***
3.673***
(0.003)
(0.002)
(0.006)
(0.005)
2.912***
3.234***
2.843***
2.963***
(0.001)
(0.001)
(0.007)
(0.006)
-10.142*** -12.813*** -13.516*** -15.234***
(0.002)
(0.002)
(0.009)
(0.008)
1.453***
1.674***
1.342***
1.563***
(0.001)
(0.001)
(0.008)
(0.007)
-1.453*** -1.674*** -1.342*** -1.563***
(0.002)
(0.002)
(0.009)
(0.008)
1.674***
1.893***
1.563***
1.782***
(0.001)
(0.001)
(0.010)
(0.009)
2.563***
2.781***
2.453***
2.674***
(0.001)
(0.001)
(0.011)
(0.010)
0.462
0.423
0.445
0.419
0.451
0.412
0.434
0.407
115.784
130.934
110.672
125.123
5.198
4.123
5.145
4.176
109.456
675.734
104.567
655.123
0.005
0.015
0.004
0.014
1.524
1.564
1.534
1.574
112.534
470.812
107.567
460.123
0.005
0.015
0.005
0.015

Note: This table presents the rolling regression analysis for the Cross-Sectional Absolute Deviation (CSAD)
in the U.S., Malaysia, and Indonesia, comparing machine learning (ML) and non-machine learning (NoML)
methods with 250-day and 500-day rolling windows. Significance level: *** indicates significance at the 1%
level, respectively.

Source: Authors' compilation

In the U.S. market, the a2 coefficient for the 250-day ML model is -8.526 (p <
0.001), compared to -11.825 (p < 0.009) in the NoML model. Similarly, for the 500-day
window, the ML model shows an a2 of -10.038 (p < 0.001) versus -13.416 (p < 0.008)
in the NoML model, indicating ML models' superior sensitivity to market dynamics. In
311

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

Malaysia, the 250-day ML model has an a2 of -9.236 (p < 0.002) compared to -12.936 (p
< 0.009) in the NoML model, and for the 500-day window, the ML model's a2 is -11.341
(p < 0.002) versus -14.613 (p < 0.008) in the NoML model. In Indonesia, the 250-day ML
model shows an a2 of -10.142 (p < 0.002) compared to -13.516 (p < 0.009) in the NoML
model, and the 500-day ML model's a2 is -12.813 (p < 0.002) versus -15.234 (p < 0.008) in
the NoML model.
Furthermore, the ML models demonstrate higher coefficients and greater significance for dynamic variables, such as market sentiment (MS), news sentiment (NS), investor happiness (IH), and trading volume (TV). In the Malaysian market using the 500-day
window, MS has a coefficient of 1.564 (p < 0.001) in the ML model versus 1.451 (p < 0.007)
in the NoML model. This indicates that ML models can better integrate sentiment analysis
to predict herding behavior. Similarly, in the Indonesian market, the 500-day ML model
showed an IH coefficient of 1.893 (p < 0.001) compared to 1.782 (p < 0.009) in the NoML
model.
The results of this study indicate that ML models outperform NoML models in
capturing herding behavior, which is consistent with previous research. Gu et al. (2020)
demonstrated that machine-learning models significantly enhance predictive accuracy in
asset pricing compared to traditional econometric models. Similarly, Fischer and Krauss
(2018) found that neural networks and other ML techniques outperform conventional
models in forecasting stock prices and managing financial risk. The higher coefficients and
greater significance for dynamic variables, such as market sentiment, news sentiment, investor happiness, and trading volume in our ML models, align with the findings of Athota
et al. (2023), Oliveira et al. (2016), and Zhang et al. (2011) who highlight the importance
of integrating sentiment analysis into predictive models.

Quantile-on-Quantile Analysis

Figure 2 employs quantile-on-quantile analysis to determine the impact of market sentiment, news sentiment, and investor happiness on CSAD. In Figures 2(a), 2(c), and 2(e),
which represent the ML models, there is a marked presence of strong positive herding
behavior, as evidenced by the predominance of red areas in the 3D surfaces. This suggests
that ML models are more sensitive to variations in sentiment and investor emotions, allowing them to detect more pronounced herding tendencies when market participants
collectively react to market news, sentiment, or happiness levels. Conversely, Figures 2(b),
2(d), and 2(f), corresponding to the NoML models, show significantly less red, with more
diffuse patterns of herding behavior. The surfaces are flatter, indicating that NoML models are less effective in capturing the intensity of herding behavior under similar market
conditions. The weaker and more scattered herding signals in these figures imply that
traditional models may fail to fully capture the complex, nonlinear relationships that drive
collective investor actions, particularly during periods of market stress or emotional extremes.
The results of this study align with and build on previous research findings that
investor sentiment significantly influences herding behavior, especially during downmarket periods (Chiang & Zheng, 2010; Omane-Adjepong et al., 2021; Youssef & Waked,
2022). This study also highlights that adverse herding occurs in low-trading volume and
low-volatility periods, which aligns with the distinct patterns observed in the ML models. Similarly, the findings demonstrate that investor sentiment can significantly impact
312

Ooi & Candra

stock market volatility, and their findings indicate that higher sentiment correlates with
increased herding behavior. This finding supports the observed strong gradients and significant responses in the quantile estimates of the ML models for market sentiment, news
sentiment, and investor happiness.

Figure 2 (a). Market Sentiment (ML) Figure 2 (b). Market Sentiment (NoML)

Figure 2 (c). News Sentiment (ML) Figure 2 (d). News Sentiment (NoML)

Figure 2 (e). Happiness (ML)

Figure 2 (f). Happiness (NoML)

Note: Figure 2 uses a color gradient where red indicates strong positive herding behavior, signifying a higher
likelihood of collective market actions based on sentiment or emotion, while blue represents weak or negative
herding behavior, indicating lower or opposite reactions. The ML models show more extensive red areas, capturing the intensity of herding more effectively, whereas the NoML models predominantly display blue, reflecting
a weaker ability to detect such behavior. This contrast underscores the superior performance of ML models in
identifying and modeling complex market dynamics driven by investor sentiment and emotions.

Robustness: Granger Causality

Table 6 examines the robustness of the relationship between key sentiment indicators and
313

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

herding behavior through Granger causality tests. The purpose is to determine whether
past values of these sentiment indicators can predict future herding behavior, thereby providing insights into the directional causality between these variables.
Table 6. Granger Causality Test
Causality Direction
FPRemarks
Statistic Value
Market Sentiment → Herding
4.52
0.0021 Reject null, Market Sentiment
Behavior
Granger causes Herding Behavior
Herding Behavior → Market
2.03
0.1210 Do not reject null, No Granger
Sentiment
causality
News Sentiment → Herding
5.78
0.0005 Reject null, News Sentiment GrangBehavior
er causes Herding Behavior
Herding Behavior → News
1.67
0.1980 Do not reject null, No Granger
Sentiment
causality
Investor Happiness → Herding
3.89
0.0047 Reject null, Investor Happiness
Behavior
Granger causes Herding Behavior
Herding Behavior → Investor
2.50
0.0876 Do not reject null, No Granger
Happiness
causality
The results reveal that market sentiment, news sentiment, and investor happiness
significantly Granger cause herding behavior, as indicated by the F-statistics of 4.52, 5.78,
and 3.89, respectively, and corresponding p-values of 0.0021, 0.0005, and 0.0047. These
low p-values lead to the rejection of the null hypothesis, confirming that changes in these
sentiment variables precede and potentially predict shifts in herding behavior.
Conversely, the reverse relationships—herding behavior predicting changes in
market sentiment, news sentiment, and investor happiness—do not exhibit significant
Granger causality. The F-statistics for these directions are 2.03, 1.67, and 2.50, with p-values of 0.1210, 0.1980, and 0.0876, respectively, all above the common significance threshold of 0.05. These results suggest that while sentiment indicators are strong predictors of
herding behavior, the influence does not appear to flow in the opposite direction, reinforcing the role of external sentiments as drivers of collective market actions rather than
outcomes influenced by those actions.

Conclusion

This study has compated human investors and machine learning models to determine if
adaptive herding occurs in different markets and conditions. The study focuses on developed countries like the U.S. and developing markets like Malaysia and Indonesia. It used
the Thomson Reuters MarketPsych market sentiment indices, the Bloomberg sentiment
analysis news sentiment scores, and the Hedonometer investor pleasure indicators from
January 2010 to December 2023. CSAD was used for static and adaptive herding assessments, along with real-time sentiment analysis and several machine learning models, including single and multi-layer neural networks.
This supports Hypothesis 1, showing that growing markets' relative inefficiency and
changing regulatory frameworks induce more herding. Machine learning models pre314

Ooi & Candra

dicted herding behavior better than statistical methods, especially in poor market conditions, supporting Hypothesis 2. The machine learning models were more sensitive to
market swings, capturing herding dynamics more precisely and restoring market conditions faster after severe economic disruptions like the COVID-19 epidemic. Hypothesis
3 was supported, showing that market sentiment, news sentiment, and investor contentment strongly influence herding. The study indicated that positive market and news sentiment and investor contentment reduced herding, but negative sentiment increased it.
They integrated these sentiment indicators well, showing that machine-learning models
can account for real-time market fluctuations better than traditional models. The study
confirmed Hypothesis 4, showing that machine learning models could reduce behavioral
biases like herding speeding market stabilization following economic shocks. The models'
fast adaptation to market conditions and feelings illustrates their potential to improve
trading tactics and market resilience.

Theory, practice, policy

This study examines financial market herding with advanced machine learning. Static
models ignore investor change. Market conditions, news, and investor pleasure explain
adaptive herding. Herding behavior fluctuates with market conditions, undermining the
EMH's market efficiency premise. ML models better capture these discrepancies, underscoring the need for flexible financial theories that account for herding.
This study helps portfolio managers and investors. In unstable markets, ML models predicted herding better than traditional methods. Machine learning may improve
risk-reward judgments. ML models can assist financial organizations in improving portfolio performance and stability by altering trading strategies faster using real-time sentiment data. The findings imply dynamic herding regulations, especially in emerging
nations like Malaysia and Indonesia. ML algorithms predict herding, helping regulators
monitor markets and act quickly. Market transparency and accurate information could
reduce herding and stabilize financial markets.

Policy, theoretical, and practical implications

The findings of this study challenge the assumptions of the EMH by demonstrating that
herding behavior is both dynamic and sentiment-sensitive. Across the sample period,
herding intensified significantly during the pandemic regime, with CSAD values rising
from 0.015 to 0.021 in the U.S., 0.018 to 0.026 in Malaysia, and 0.019 to 0.023 in Indonesia (Table 4). These patterns suggest that market participants react collectively in times
of uncertainty, deviating from rational pricing mechanisms assumed under EMH. This
reinforces the adaptive market hypothesis, which posits that market efficiency is not static
but evolves in response to changing environments. The pronounced herding in emerging
Shariah-compliant markets, particularly during crises, reflects structural differences in
regulatory frameworks, informational asymmetries, and investor composition.
From a practical standpoint, the results indicate that machine learning models
offer superior predictive performance over traditional statistical approaches. In adaptive
herding estimation (Table 5), ML-based models recorded significantly lower CSAD coefficients during high-volatility periods—for instance, −8.526 (250-day window) in the U.S.,
compared to −11.825 in the non-ML model. Similar trends were observed in Malaysia
(−9.236 ML vs. −12.936 NoML) and Indonesia (−10.142 ML vs. −13.516 NoML), con315

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

firming the ability of ML models to detect behavioral shifts more sensitively. Furthermore,
ML models demonstrated stronger responsiveness to real-time sentiment inputs such as
investor happiness (1.673 in the U.S. and 1.893 in Indonesia), news sentiment (−1.454
in Malaysia), and market sentiment (1.674 in Indonesia). The quantile-on-quantile surfaces (Figure 2) further illustrate how ML models captured high herding intensity under extreme sentiment conditions—represented by dense red regions—unlike the flatter,
less responsive surfaces from traditional models. These results imply that ML-enhanced
trading systems could provide institutional investors with a competitive edge in adjusting
portfolio allocations dynamically during market disruptions.
Moreover, the Granger causality analysis (Table 6) revealed that market sentiment
(F = 4.52, p = 0.0021), news sentiment (F = 5.78, p = 0.0005), and investor happiness (F =
3.89, p = 0.0047) all Granger-cause herding behavior, while the reverse causality was not
statistically significant. This one-directional causality highlights the critical role of external
sentiment as a leading indicator of irrational collective behavior. In response, policymakers—particularly in sentiment-sensitive emerging markets—should implement dynamic
herding regulations and invest in real-time market monitoring systems. ML-based tools
that incorporate social mood and sentiment metrics can serve as early warning systems
for regulators. Furthermore, ensuring timely and reliable financial information through
responsible media dissemination could mitigate the contagion effects of negative sentiment, contributing to improved market stability and investor protection.

Limitation

This study provides valuable insights but has limits that need more study. The Thomson
Reuters MarketPsych and Bloomberg sentiment analyses may not represent market sentiment fully, especially in places where alternative media outlets impact investor behavior.
Include social media and regional news outlets to get a fuller view of market sentiment
and herding. The U.S., Malaysia, and Indonesia focus provides useful comparative insights
but restricts generalizability. To study herding behavior dynamics, future research should
include marketplaces with varied regulatory and market structures. Comparative studies
of developed and emerging markets would clarify how market maturity and regulation
affect investor behavior.
CRediT authorship contribution statement
Ooi Kok Loang: Conceptualization; Methodology; Validation; Formal Analysis; Resources; Data Curation; Writing -Original Draft; Writing -Review & Editing; Visualization;
Supervision. Sevenpri Candra: Conceptualization; Validation; Data Curation; Writing
-Original Draft; Writing -Review & Editing; Project Administration.

References

Amanda, C., & Pradipta, A. D. (2024). Risk-Based Premiums of Insurance Guarantee
Schemes: A Machine-Learning Approach. Journal of Indonesian Economy and
Business, 39(2), Article 2. https://doi.org/10.22146/jieb.v39i2.9323
Athota, V. S., Hasan, Z., Vaz, D., Désiré, S. S. M., & Pereira, V. (2023). Can Artificial Intelligence (AI) Manage Behavioral Biases Among Financial Planners? J. Glob. Inf.
316

Ooi & Candra

Manage., 31(2), 1–18. https://doi.org/10.4018/JGIM.321728
Baker, M., & Wurgler, J. (2007). Investor Sentiment in the Stock Market. Journal of Economic Perspectives, 21(2), 129–152. https://doi.org/10.1257/jep.21.2.129
Barberis, N., & Thaler, R. (2003). Chapter 18 A survey of behavioral finance. In Handbook of the Economics of Finance (Vol. 1, pp. 1053–1128). Elsevier. https://doi.
org/10.1016/S1574-0102(03)01027-6
Bihari, A., Dash, M., Muduli, K., Kumar, A., Mulat-Weldemeskel, E., & Luthra, S. (2023).
Does cognitive biased knowledge influence investor decisions? An empirical investigation using machine learning and artificial neural network. VINE Journal of
Information and Knowledge Management Systems, ahead-of-print(ahead-of-print).
https://doi.org/10.1108/VJIKMS-08-2022-0253
Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades. Journal of Political Economy,
100(5), 992–1026. https://doi.org/10.1086/261849
Bollen, J., Mao, H., & Zeng, X. (2011). Twitter mood predicts the stock market. Journal of
Computational Science, 2(1), 1–8. https://doi.org/10.1016/j.jocs.2010.12.007
Bouri, E., Gupta, R., & Roubaud, D. (2019). Herding behavior in cryptocurrencies. Finance Research Letters, 29, 216–221. https://doi.org/10.1016/j.frl.2018.07.008
Chang, E. C., Cheng, J. W., & Khorana, A. (2000). An examination of herd behavior in equity markets: An international perspective. Journal of Banking & Finance, 24(10),
1651–1679. https://doi.org/10.1016/S0378-4266(99)00096-5
Chiang, T. C., & Zheng, D. (2010). An empirical analysis of herd behavior in global stock
markets. Journal of Banking & Finance, 34(8), 1911–1921. https://doi.org/10.1016/j.
jbankfin.2009.12.014
Choi, J., & Park, H. J. (2023). Financial Distress and Audit Report Lags: An Empirical
Study in Korea. Gadjah Mada International Journal of Business, 25(3), Article 3.
https://doi.org/10.22146/gamaijb.72251
Da, Z., Engelberg, J., & Gao, P. (2015). The Sum of All FEARS Investor Sentiment and
Asset Prices. The Review of Financial Studies, 28(1), 1–32. https://doi.org/10.1093/
rfs/hhu072
Dodds, P. S., Harris, K. D., Kloumann, I. M., Bliss, C. A., & Danforth, C. M. (2011). Temporal Patterns of Happiness and Information in a Global Social Network: Hedonometrics and Twitter. PLOS ONE, 6(12), e26752. https://doi.org/10.1371/journal.
pone.0026752
Engelberg, J. E., & Parsons, C. A. (2011). The Causal Impact of Media in Financial
Markets. The Journal of Finance, 66(1), 67–97. https://doi.org/10.1111/j.15406261.2010.01626.x
Fama, E. F. (1970). Efficient Capital Markets: A Review of Theory and Empirical Work.
The Journal of Finance, 25(2), 383–417. https://doi.org/10.2307/2325486
Feng, G., Giglio, S., & Xiu, D. (2019). Taming the Factor Zoo: A Test of New Factors (SSRN
Scholarly Paper No. 2934020). https://doi.org/10.2139/ssrn.2934020
Fischer, T., & Krauss, C. (2018). Deep learning with long short-term memory networks
for financial market predictions. European Journal of Operational Research, 270(2),
654–669. https://doi.org/10.1016/j.ejor.2017.11.054
Gao, Y., Zhao, C., & Wang, Y. (2024). Investor sentiment and stock returns: New evidence
from Chinese carbon-neutral stock markets based on multi-source data. International Review of Economics & Finance, 92, 438–450. https://doi.org/10.1016/j.
317

Gadjah Mada International Journal of Business - September-December, Vol. 27, No. 3, 2025

iref.2024.02.049
Goodell, J. W. (2020). COVID-19 and finance: Agendas for future research. Finance Research Letters, 35, 101512. https://doi.org/10.1016/j.frl.2020.101512
Gopal, S., & Loang, O. K. (2024). The Rationality Conundrum: Exploring Herd Mentality
among Individual Investors in the Indian Stock Market. Indian Journal of Finance,
18(6), Article 6. https://doi.org/10.17010/ijf/2024/v18i6/173967
Gu, S., Kelly, B., & Xiu, D. (2020). Empirical Asset Pricing via Machine Learning. The
Review of Financial Studies, 33(5), 2223–2273. https://doi.org/10.1093/rfs/hhaa009
Henrique, B. M., Sobreiro, V. A., & Kimura, H. (2019). Literature review: Machine learning techniques applied to financial market prediction. Expert Systems with Applications, 124, 226–251. https://doi.org/10.1016/j.eswa.2019.01.012
Hu, J., Sui, Y., & Ma, F. (2021). The Measurement Method of Investor Sentiment and Its
Relationship with Stock Market. Computational Intelligence and Neuroscience,
2021(1), 6672677. https://doi.org/10.1155/2021/6672677
Hwang, S., & Salmon, M. (2004). Market stress and herding. Journal of Empirical Finance,
11(4), 585–616. https://doi.org/10.1016/j.jempfin.2004.04.003
Kliegr, T., Bahník, Š., & Fürnkranz, J. (2021). A review of possible effects of cognitive biases on interpretation of rule-based machine learning models. Artificial Intelligence,
295, 103458. https://doi.org/10.1016/j.artint.2021.103458
Krauss, C., Do, X. A., & Huck, N. (2017). Deep neural networks, gradient-boosted trees,
random forests: Statistical arbitrage on the S&P 500. European Journal of Operational Research, 259(2), 689–702. https://doi.org/10.1016/j.ejor.2016.10.031
Kumbure, M. M., Lohrmann, C., Luukka, P., & Porras, J. (2022). Machine learning techniques and data for stock market forecasting: A literature review. Expert Systems
with Applications, 197, 116659. https://doi.org/10.1016/j.eswa.2022.116659
Lo, A. W. (2004). The Adaptive Markets Hypothesis. The Journal of Portfolio Management,
30(5), 15–29. https://doi.org/10.3905/jpm.2004.442611
Mahadwartha, P. A., Ismiyanti, F., & Zunairoh, Z. (2023). The Gambler’s Fallacy, the Halo
Effect, and the Familiarity Effect Based on Risk Profile: Bullish and Bearish Market in Indonesia Stock Exchange. Gadjah Mada International Journal of Business,
25(2), Article 2. https://doi.org/10.22146/gamaijb.64801
Mnif, E., Jarboui, A., & Mouakhar, K. (2020). How the cryptocurrency market has performed during COVID 19? A multifractal analysis. Finance Research Letters, 36,
101647. https://doi.org/10.1016/j.frl.2020.101647
Nguyen, Y. V. B., & Vo, A. H. K. (2023). Herding behavior before and after COVID-19
pandemic: Evidence from the Vietnam stock market. Journal of Economic Studies,
51(2), 357–374.
Noda, A. (2016). A test of the adaptive market hypothesis using a time-varying AR
model in Japan. Finance Research Letters, 17, 66–71. https://doi.org/10.1016/j.
frl.2016.01.004
Nugroho, R., & Pratiwi, N. R. (2023). Revisiting Islamic Leadership: Clarifying Theoretical Issues. Journal of Leadership in Organizations, 5(2), Article 2. https://doi.
org/10.22146/jlo.78412
Oliveira, N., Cortez, P., & Areal, N. (2016). Stock market sentiment lexicon acquisition
using microblogging data and statistical measures. Decision Support Systems, 85,
62–73. https://doi.org/10.1016/j.dss.2016.02.013
Omane-Adjepong, M., Paul Alagidede, I., Lyimo, A. G., & Tweneboah, G. (2021). Herding
318

Ooi & Candra

behavior in cryptocurrency and emerging financial markets. Cogent Economics &
Finance, 9(1), 1933681. https://doi.org/10.1080/23322039.2021.1933681
Schmeling, M. (2009). Investor sentiment and stock returns: Some international evidence.
Journal of Empirical Finance, 16(3), 394–408. https://doi.org/10.1016/j.jempfin.2009.01.002
Soni, P., Tewari, Y., & Krishnan, D. (2022). Machine Learning Approaches in Stock Price
Prediction: A Systematic Review. Journal of Physics: Conference Series, 2161(1),
012065. https://doi.org/10.1088/1742-6596/2161/1/012065
Tetlock, P. C. (2007). Giving Content to Investor Sentiment: The Role of Media in the
Stock Market. The Journal of Finance, 62(3), 1139–1168. https://doi.org/10.1111/
j.1540-6261.2007.01232.x
Urquhart, A., & McGroarty, F. (2016). Are stock markets really efficient? Evidence of the
adaptive market hypothesis. International Review of Financial Analysis, 47, 39–49.
https://doi.org/10.1016/j.irfa.2016.06.011
Youssef, M., & Waked, S. S. (2022). Herding behavior in the cryptocurrency market during COVID-19 pandemic: The role of media coverage. The North American Journal
of Economics and Finance, 62, 101752. https://doi.org/10.1016/j.najef.2022.101752
Zhang, X., Fuehres, H., & Gloor, P. A. (2011). Predicting Stock Market Indicators Through
Twitter “I hope it is not as bad as I fear.” Procedia - Social and Behavioral Sciences,
26, 55–62. https://doi.org/10.1016/j.sbspro.2011.10.562

319