Heikki Anttila Herding Bias and Its Impact on Long-Term Fund Performance: A Human vs. AI Comparison Evidence from the US Vaasa 2025 School of Accounting and Finance Master’s thesis in Finance Master’s Degree Programme in Finance 2 UNIVERSITY OF VAASA School of Accounting and Finance Author: Heikki Anttila Title of the thesis: Herding Bias and Its Impact on Long-Term Fund Performance: A Hu- man vs. AI Comparison : Evidence from the US Degree: Master of Science in Economics and Business Administration Discipline: Finance Supervisor: Veda Fatmy Year: 2025 Pages: 92 ABSTRACT : The growing role of artificial intelligence in asset management (AI) has raised questions on how it compares to human fund managers, especially considering behavioral biases and their effect on fund performance. Herding bias is a well-studied topic in financial literature, however, its implications in comparison to AI-managed funds are yet to be explored. The objective of this thesis is to examine the herding bias and its impact on the long-term fund performance of human-managed funds compared to AI-managed funds. The main focus of the paper is to study whether human-managed funds exhibit herding behavior and whether such behavior affects their performance relative to AI-managed funds. The comparison focuses on U.S.-based fund categories from 01.01.2019 to 24.09.2024. The key method for detecting herding behavior among the funds during the whole sample pe- riod and the subperiods is the Cross-Sectional Absolute Deviation (CSAD) model, which esti- mates return dispersion relative to market movements. The empirical study uses daily fund-level returns on both human-managed and AI-managed funds. In addition to the whole sample pe- riod, the study examines four subperiods: the COVID-19 pandemic, the Russia-Ukraine war, the Silicon Valley Bank collapse, and the Israel-Gaza conflict. To compare performance between funds, mean return, cumulative return, and the Sharpe ratio are utilized. The two hypotheses are based on previous literature and address both the whole sample period and the subperiods. The findings of this thesis do not provide statistically significant herding behavior among the human-managed fund categories during the whole sample period or the subperiods. Despite the absence of herding, the fund performance metrics reveal that AI-managed funds offer stable risk-adjusted returns during the sample period. However, human-managed fund categories Growth and Balanced offer superior returns and risk-adjusted returns during most of the peri- ods. These results indicate that the performance differences are based more on the strategy and context than on herding behavior. When considering the results of the thesis as a whole, the evidence of herding behavior remains inconsistent with previous literature. However, the findings are consistent with the Efficient Market Hypothesis, which states that markets portray all available information. The varying per- formance differences highlight the relevance of fund strategies and introduce the fact that AI funds’ advantage lies in consistency. This thesis contributes to the limited body of research that directly compares behavioral biases and performance differences between human and AI-man- aged funds. The study highlights the importance of fund types, philosophies, and market condi- tions in forming outcomes. The results also pave the way for future research that utilizes up- dated measures for detecting herding and focuses on different markets and behavioral biases. KEYWORDS: Behavioral Finance, Herding Behavior, Artificial Intelligence, Fund Management, Cross-Sectional Absolute Deviation. 3 TIIVISTELMÄ : Tekoälyn kasvava rooli omaisuudenhoidossa on herättänyt kysymyksiä siitä, miten se vertautuu ihmisten hallinnoimiin rahastoihin, erityisesti käytöksellisten harhojen ja niiden vaikutuksien nä- kökulmasta. Laumakäyttäytyminen on laajalti tutkittu aihe rahoituskirjallisuudessa, mutta sen vaikutuksia verrattaessa tekoälyn hallinnoimiin rahastoihin ei vielä ole tutkittu. Tämän pro- gradu tutkielman tavoitteena on tutkia laumakäyttäytymisen esiintymistä ja sen vaikutusta ih- misjohtoisesti hallinnoitujen rahastojen pitkän aikavälin tuottoon verrattuna tekoälypohjaisiin rahastoihin. Tutkimuksen pääpaino on selvittää, esiintyykö ihmisjohtoisissa rahastoissa lauma- käyttäytymistä ja heikentääkö tämä mahdollisesti niiden suhteellista tuottavuutta verrattuna AI- rahastoihin. Vertailun kohteena ovat Yhdysvalloissa toimivat rahastokategoriat ajanjaksolla 1.1.2019–24.9.2024. Keskeisenä menetelmänä laumakäyttäytymisen tunnistamisessa käytetään Cross-Sectional Ab- solute Deviation (CSAD) -mallia, joka mittaa tuottojen hajontaa suhteessa markkinaliikkeisiin. Empiirisessä tutkimuksessa käytetään päivätason rahastotuottoja sekä ihmisen että tekoälyn hallinnoimista rahastoista. Koko otosjakson lisäksi tutkimuksessa tarkastellaan neljää osajaksoa: COVID-19-pandemiaa, Venäjä–Ukraina-sotaa, Silicon Valley Bankin romahdusta ja Israel–Gaza- konfliktia. Rahastojen tuottavuuden vertailussa käytetään keskimääräistä tuottoa, kumulatii- vista tuottoa sekä Sharpen lukua. Tutkielman kaksi hypoteesia pohjautuvat aiempaan kirjallisuu- teen ja käsittelevät sekä koko otosjaksoa että osajaksoja. Tutkimuksen tulokset eivät osoita tilastollisesti merkittävää laumakäyttäytymistä ihmisjohtoi- sissa rahastoissa koko otosjakson tai osajaksojen aikana. Laumakäyttäytymisen puuttumisesta huolimatta tuottolukujen tarkastelu osoittaa, että tekoälyrahastot tarjoavat vakaampia riskikor- jattuja tuottoja tarkastelujakson aikana. Kuitenkin ihmisjohtoiset Kasvu- ja Tasapainotettu-ra- hastot saivat parempia kokonaistuottoja ja riskikorjattuja tuottoja useimmissa tarkastelluissa jaksoissa. Tulokset viittaavat siihen, että tuottoerot perustuvat enemmänkin rahaston strategi- aan ja markkinatilanteeseen kuin laumakäyttäytymiseen. Tarkasteltaessa tutkielman tuloksia kokonaisuutena, laumakäyttäytymisen empiiriset todisteet ovat ristiriidassa aiemman kirjallisuuden kanssa. Tulokset ovat kuitenkin linjassa tehokkaiden markkinoiden hypoteesin kanssa, jonka mukaan markkinahinnat heijastavat kaikkea saatavilla olevaa informaatiota. Rahastojen vaihtelevat tuottoprofiilit korostavat strategioiden merkitystä ja osoittavat, että tekoälyrahastojen etu piilee johdonmukaisuudessa. Tutkielma täydentää har- valukuista tutkimuskenttää, joka vertailisi suoraan käytöksellisiä harhoja ja tuottoeroja ihmis- ja tekoälypohjaisten rahastojen välillä. Tämä tutkimus korostaa rahastotyyppien, sijoitusfilosofioi- den ja markkinatilanteiden merkitystä lopputuloksiin. Tutkielma luo myös pohjaa jatkotutkimuk- selle, joka voisi hyödyntää kehittyneempiä metodeja laumakäyttäytymisen tunnistamiseksi sekä keskittyisi eri markkinoihin ja käytöksellisiin harhoihin. AVAINSANAT: Behavioristinen Rahoitus, Laumakäyttäytyminen, Tekoäly, Rahastojen Hallin- nointi, Cross-Sectional Absolute Deviation 1 1 In the making of this thesis, artificial intelligence tools were utilized for enhancement purposes. Gram- marly was employed to elevate the quality of writing. ChatGPT was used for consultancy and idea purposes. 4 Table of Contents 1 Introduction 7 1.1 Purpose of the Study 8 1.2 Contribution to the Existing Literature 9 1.3 Structure of the Thesis 10 2 Theoretical Background 11 2.1 Traditional Finance 11 2.1.1 Efficient Market Hypothesis 12 2.1.2 Expected Utility Theory 14 2.1.3 Modern Portfolio Theory 16 2.2 Behavioral Finance 19 2.2.1 Herding in Financial Markets 20 2.2.2 Dimensions of Herding Behavior 23 2.2.3 Drivers of Herding Behavior 24 2.3 Artificial Intelligence in Finance 27 2.3.1 Overview of Artificial Intelligence 28 2.3.2 Artificial Intelligence’s Applications in Fund Management 32 2.3.3 Artificial Intelligence’s Limitations and Biases 35 3 Literature Review 37 3.1 Overview of Herding Behavior in Financial Markets 37 3.2 Herding Behavior in Volatile Market Conditions 39 3.3 Herding and Fund Management 41 3.4 Human vs. Artificial Intelligence in Fund Management 43 3.5 Hypothesis Development 46 4 Methodology 49 4.1 Measures of Herding Behavior 49 4.2 Fund Performance Metrics 51 5 Data 53 5.1 Data Collection 53 5 5.2 Descriptive Statistics 54 6 Empirical Results 59 6.1 Herding Behavior During the Whole Sample Period 59 6.2 Herding Behavior During Subperiods 63 7 Limitations and Suggestions for Future Research 75 7.1 Sample and Data Limitations 75 7.2 Herding Measurement Limitations 76 7.3 Other Limitations 77 7.4 Suggestions for Future Research 77 8 Conclusions 78 References 82 Appendices 91 Appendix 1. Fund List 91 Figures Figure 1. Dimensions of Select Data Techniques (Choi et al., 2020). 29 Figure 2. Main AI Methods in Asset Management (Bartram et al., 2020) 33 Figure 3. Relationship between fund CSADs and Market Return (Rm) 57 Tables Table 1. Descriptive Statistics of Cross-Sectional Absolute Deviations 54 Table 2. Regression results of herding behavior during the entire sample period 60 Table 3. Fund performance metrics for the full sample period 62 Table 4. Regression results of herding behavior during the COVID-19 period 64 Table 5. Regression results of herding behavior during the Russia-Ukraine War period66 Table 6. Regression results of herding behavior during the SVB Collapse period 67 Table 7. Regression results of herding behavior during the Israel-Gaza Conflict period 69 Table 8. Fund performance metrics for the subperiods 71 6 Abbreviations EMH Efficient Market Hypothesis EUT Efficient Utility Theory MPT Modern Portfolio Theory CAPM Capital Asset Pricing Model CSSD Cross-Sectional Standard Deviation CSAD Cross-Sectional Absolute Deviation AI Artificial Intelligence ML Machine Learning 7 1 Introduction Behavioral finance rose as an approach after traditional finance could not explain certain trading behaviors through traditional economic theories (Barberis & Thaler, 2003). Be- havioral biases or cognitive biases are central to behavioral finance, which incline hu- mans to use shortcuts while making decisions (Jain et al., 2015). Behavioral biases have been recognized as critical factors influencing investor behavior and financial markets. One of the most studied biases is herding behavior, where investors mimic the collective actions of other investors instead of relying on their analysis or information (Christie & Huang, 1995). Herding mentality often leads to inefficiencies in the market and can re- sult in the creation of bubbles, mispricing of securities, and crashes. Herding is particu- larly present during periods of high volatility and uncertainty (Devenow & Welch, 1996). In these times, investors tend to follow others’ trading behavior out of fear of missing out or concern about major losses. This behavior has been documented in human-man- aged funds where behavioral biases play a role. As behavioral finance rose against traditional finance, now artificial intelligence (AI) is rising to battle the cognitive biases that the human brain possesses. With the rise of AI in finance and fund management, the question arises whether AI-powered funds, driven by algorithms and data rather, are less prone to human biases and can produce superior returns in the long term. AI's development in asset management can potentially reduce or even eliminate behavioral biases. AI-driven funds utilize machine learning and data- driven strategies, which are said to make decisions based on objective market indicators and prediction models (Kolanovic & Krishnamachari, 2017). In contrast to human man- agers, AI systems do not show fear, greed, or other emotional reactions that frequently result in irrational decisions (Hirshleifer, 2020). AI may outperform human-managed funds, especially during high volatility, since it can process large volumes of data and execute transactions quickly. However, the degree to which AI-powered funds outper- form humans and whether they can completely disregard behavioral biases is yet to be answered. 8 The inhumane nature of AI does have some disadvantages in comparison with humans, which might give the edge to human-managed funds. Inaccurate data may cause AI sys- tems to make unreasonable investment decisions (Barocas et al., 2023). Barocas contin- ues that humans usually do not make these unreasonable choices and usually rely on moral significance among the used attributes. Additionally, using flawed training data while creating the algorithm can lead to poor decision-making (IBM, 2023). The U.S. financial market provides a diverse environment for detecting these dynamics, given its liquidity, various market participants, and presence of human- and AI-managed funds. In recent years, the world of stocks has experienced multiple periods of significant volatility, including the COVID-19 pandemic, the Russia-Ukraine war, the Silicon Valley Bank collapse, and the Israel-Gaza war. Studying and potentially detecting herding during these periods offers valuable information through which the comparison between hu- man- and AI-managed funds can be closely examined. 1.1 Purpose of the Study This study primarily aims to examine the impact of herding behavior on the long-term performance of human-managed funds compared to AI-powered funds within the U.S. financial markets. More specifically, it focuses on periods of high market volatility from 2019-2024, including COVID-19, the Russia-Ukraine war, the Silicon Valley Bank collapse, and the Israel-Gaza conflict. The purpose of the study is to evaluate whether herding behavior is more prevalent among human fund managers and whether AI-powered funds can diminish emotional biases and deliver superior long-term performance. Investors are increasingly debating whether AI in fund management can eliminate cog- nitive and emotional heuristics that traditionally have affected human fund managers (Kolanovic & Krishnamachari, 2017). According to Hirshleifer (2020), AI has the potential to process larger volumes of data and make quick decisions without being emotionally influenced. However, it is still unclear whether AI-driven funds are completely free from 9 biases or if they show systematic behavior resembling herding, particularly in highly vol- atile market conditions. This study attempts to analyze the levels of herding during turbulent markets and then compare the returns between human and AI-managed funds. The paper will explore whether avoiding trading in the same direction has produced better long-term returns and overall outcomes by utilizing metrics such as mean return, cumulative return, and Sharpe ratio. For detecting herding behavior, the Cross-Sectional Absolute Deviation (CSAD) model from seminal work by Chang et al. (2000) will be utilized. 1.2 Contribution to the Existing Literature This thesis offers a new perspective to the existing literature by investigating whether herding behavior influences the performance of funds managed by humans compared to those managed by AI, which should be unaffected by such biases. Despite the growing popularity of AI in fund management, there is a scarcity of academic studies that directly compare its performance with that of human managers. Existing research on AI in fi- nance typically centers around areas such as algorithmic trading, predictive modeling, or risk management, rather than providing a head-to-head performance comparison with human managers facing behavioral constraints (Chen & Ren, 2022; Grobys et al., 2022). By examining human-managed funds, which may be influenced by herding bias, along- side AI-managed funds that are anticipated to function solely on data-driven methods, this thesis delivers a comparative analysis that is currently absent in existing literature. While herding and related behavioral biases have been widely examined in human-man- aged funds (Lakonishok et al., 1992; Brown et al., 2014; Jiang & Verardo, 2018), no pre- vious research has systematically evaluated the performance differences between hu- man and AI fund management with a focus on herding. This research addresses a signif- icant gap by analyzing how herding among human fund managers impacts their long- 10 term performance when compared to AI-managed funds, which operate without emo- tional or psychological factors. 1.3 Structure of the Thesis The structure of the study is the following. The first chapter presents the introduction for the study by outlining its purpose and contribution to the prevailing literature. The theoretical background is introduced in the second part of the thesis by explaining key theories from traditional finance, behavioral finance, and artificial intelligence in fund management. The third chapter is dedicated to the literature review, which includes pre- vious relevant research and hypothesis development. This chapter establishes the framework upon which the research hypotheses are built. The fourth chapter outlines the methodology used in the empirical part of the study. It explains the models and for- mulas applied to measure herding activity and to calculate the performance metrics of human and AI-driven funds. In the fifth chapter, the data collection process and descrip- tive statistics of the dataset are presented. The sixth chapter presents the empirical re- sults of the study. The results are discussed and evaluated with an emphasis on the re- search hypotheses. The seventh chapter provides a discussion about the prevailing limi- tations of the study and suggestions for future research. Lastly, the final chapter con- cludes the findings and explores their possible implications and overall contribution. 11 2 Theoretical Background Commonly, finance has been divided into traditional and behavioral finance. Traditional finance aims to explain phenomena using rational means, while the new approach, be- havioral finance, argues that some events can be better explained using irrational agents (Barberis & Thaler, 2003). According to Holtfort (2018), traditional finance assumes per- fect market conditions, and behavioral finance contradicts this view by assuming that psychological factors affect investors’ decision-making. The theoretical background part of the study is divided into three subsections. First, this thesis examines traditional fi- nance. Then, light is shed on the evolution of finance to the recognition of behavioral finance. Lastly, the latest and most revolutionary area of finance, artificial intelligence, is introduced. This chapter provides the necessary theories and hypotheses to understand the topic of the study better. 2.1 Traditional Finance Traditional finance, often recognized as standard finance, is based on core theories and principles (Kumar and Goyal, 2015). These include the arbitrage principles developed by Miller and Modigliani, Markowitz’s portfolio theory, the Capital Asset Pricing Model (CAPM) proposed by Sharpe, Lintner, and Black, and the option-pricing model of Black, Scholes, and Merton. According to Kumar and Goyal, at the core of the aforementioned is the Efficient Market Hypothesis (EMH). The EMH relies on the assumption that the asset prices incorporate all available information. Additionally, the EMH also suggests that market participants act rationally and have access to accurate information. This pre- vents asset mispricing and opportunities for arbitrage. In addition to the Efficient Market Hypothesis, the other core principle in traditional finance is the Expected Utility Theory (EUT). Next, the two mentioned theories are examined in more depth. The Modern Port- folio Theory by Markowitz is also introduced to frame the fund management aspect of the thesis. 12 2.1.1 Efficient Market Hypothesis Efficiency is a key principle in finance. For many years, scholars and economists have examined its relevance in capital markets, leading to the Efficient Market Hypothesis (EMH) becoming a central topic in financial research (Ţiţan, 2015). The study by Ţiţan indicates that the EMH has sparked discussion, with many scholars either supporting it or questioning its legitimacy. It is essential to comprehend the roots of this idea and the development of research on market efficiency, especially in recent years, to understand the current debates and varying viewpoints on the EMH. The Efficient Market Hypothesis aligns with the Random Walk Theory (RWT) (Malkiel, 2003). As explained by Malkiel, the RWT describes a sequence of prices where all future prices deviate from previous prices randomly. The reasoning for the RWT is that if infor- mation flows freely and is instantly shown in stock prices. THerefore the stock price of tomorrow will only show tomorrow’s information and will not correlate with today’s price changes. However, news is unpredictable. Price changes must also appear random. Malkiel states that this implies prices to fully portray all available information. This lets even uninformed investors purchase a diversified portfolio at current market prices and obtain the same rates of return as financial experts. The generalized version of the Efficient Market Hypothesis was first introduced in 1970 by Eugene Fama. In his research, Fama clarifies that the primary function of the capital market is to distribute ownership of the economy's capital assets. Ideally, this market would operate in a way where prices accurately signal resource allocation; that is, firms can make informed decisions regarding production and investment, while investors can choose from securities that represent ownership of these companies' activities under the assumption that security prices at any given moment "fully portray" all available in- formation. A market characterized by prices that consistently "fully portray" the available information is considered "efficient." 13 According to Fama (1970), the EMH divides asset prices into three subsections. The first category, weak form tests, evaluates whether prices solely depend on past price infor- mation. Following that, semi-strong form tests examine how effectively prices reflect publicly accessible information, like earnings reports or stock splits. Finally, strong form tests investigate whether specific investors or groups possess exclusive access to vital information that may affect pricing. According to the author, the results suggest that, with a few notable exceptions, the efficient markets hypothesis largely stands up to em- pirical examination. In her study, Ţiţan (2015) discusses various research examining the three forms of the Efficient Market Hypothesis. Most of this research does not support the semi-strong and strong forms of the EMH, as financial data fails to validate these theories. Opinions on the weak form of the EMH, which includes the random walk theory, are mixed. Some studies within the weak form suggest that abnormal returns result primarily from ran- dom chance, with overreactions and underreactions occurring at approximately equal probabilities, providing support for the weak form of the EMH. A commonly observed pattern in financial data is that anomalies often diminish or dis- appear as analytical models improve (Ţiţan, 2015). This suggests that these anomalies may stem from the methodologies used. Ţiţan emphasizes that most research in this field employs event studies; some of these studies analyze the immediate price response to announcements, arguing that financial assets quickly adjust to new information, thus supporting the idea of market efficiency. However, other studies take a longer-term per- spective, noting a gradual adjustment of prices to new information. This challenges the Efficient Market Hypothesis (EMH) by indicating potential inefficiencies over medium- and long-term periods. As noted by Ţiţan (2015) in the first paragraph of this subsection, the EMH has faced both support and criticism. Critics argue that certain market phenomena, such as price 14 anomalies and behavioral biases, challenge the notion that markets incorporate all avail- able information efficiently. Studies conducted by Fama and French (1988) and Poterba and Summers (1988) found evidence of mean reversion in stock returns. These findings indicate that returns can deviate from efficiency over the long term due to overreactions and subsequent corrections. Behavioral finance researchers DeBondt and Thaler (1995) and other have connected these patterns to psychological factors. Such factors include investor overconfidence and sentiment cycles, which can lead to temporary mispricings. Moreover, the concept of limits to arbitrage, where rational investors are unable to fully correct mispricings caused by irrational behavior, challenges the EMH. This concept sug- gests that irrationality can have lasting effects on market prices (Malkiel, 2003; Schwert, 2001). 2.1.2 Expected Utility Theory The Expected Utility Theory (EUT) is a fundamental framework for decision-making un- der uncertainty (Fishburn, 1981). Originally formalized by Von Neumann and Morgen- stern in 1944, this theory offers insights into how to make rational choices when the outcomes of our actions are uncertain (Briggs, 2014). Its core principle is to select the action that yields the highest expected utility. As noted by Ritter (2003), the EUT is viewed as normative rather than descriptive, highlighting that it places greater emphasis on the level of wealth itself rather than on variations in that wealth. In his research, Briggs also treats the EUT as how the decisions should be made rather than how they are done, of which the latter represents how the EUT is used in classical economics. According to Briggs' research, the expected utility of a particular action is the average of the utilities of its possible outcomes, weighted by their probabilities. In this context, the utility of each outcome indicates the degree to which that outcome is favored compared to other alternatives. Moreover, the utility of each outcome is modified by the probabil- ity that the action will lead to that specific result, states Briggs. In the study, Briggs sim- 15 plifies the EUT using an example; suppose you are going for a walk, and you are wonder- ing whether to bring your umbrella on a sunny day. You would rather not carry it for nothing, but it would be better to have it, in case of rain. Briggs continues that the abovementioned scenario can be formalized using three types of elements. First, the outcomes, meaning things that are directly preferred or not. There are three in the problem. Staying dry and not carrying an umbrella, staying dry but bur- dened with the umbrella, or getting wet. Next, the states, which are the factors beyond your control that affect the decision’s outcome. Here, the states are rain or no rain. Lastly, the acts, indicating things that the decision-maker prefers and can do. In this case, the acts are carrying the umbrella or leaving it at home. According to the example provided in Briggs’ study, the EUT helps rank these choices on the choice worthiness: the more expected utility you gain, the more reasonable it is to select that option. Similar to the Efficient Market Hypothesis, the Expected Utility Theory also faces views both for and against. The EUT faces criticism mainly from two perspectives. First, the EUT is considered irrational (Malkiel, 2003; Briggs, 2014). Secondly, according to Briggs, some critics claim it to be impossible. As Briggs (2014) mentioned, multiple authors have provided evidence that the EUT ap- pears to offer incorrect guidance. Some examples suggest that rational preferences de- viate from the EUT, implying that maximized expected utility might not be essential for rational behavior. Briggs also provides cases where the EUT allows irrational preferences, indicating that maximizing expected utility might not be sufficient for rationality. How- ever, Malkiel (2003) contradicts the criticism of the EUT being irrational by stating that even if the market participants act irrationally, the market can remain efficient. The au- thor emphasizes that efficiency can persist despite stock prices sometimes showing more volatility than what fundamentals like dividends and earnings justify. 16 As stated, the EUT has also received objections on the fact that it is impossible. In his study, Feldman (2006) notes that utilitarians are drawn to the notion that an action is morally right if it results in the most favorable outcome. However, the author states that critics have been arguing that in numerous instances, the agents are unable to identify which of the options would produce the best result. Therefore, the traditional principle cannot be relied on to guide our decisions. According to Briggs (2014), similar points have been made by Smith (2010) and March and Simon (1958). Furthermore, McGee (1991) argues that it is mathematically impossible to maximize expected utility, even with a perfect computer equipped with infinite memory. McGee explains that to maxim- ize expected utility, a rational agent should accept any wager based on a true arithmetic statement while refusing wagers related to false statements in arithmetic. However, since the set of all true arithmetic statements is not recursively enumerable, no Turing machine can determine the truth or falsity of all such arithmetic statements. 2.1.3 Modern Portfolio Theory Modern Portfolio Theory (MPT) was pioneered by Harry Markowitz in 1952. It is re- garded as one of the most compelling works of traditional finance. The paper by Marko- witz (1952) revolutionized the investment process and finance by introducing a quanti- tative framework for portfolio building that maximizes expected return for a given level of risk. The MPT relies on the rules of diversification and efficient frontier optimization, allowing investors to make decisions based on statistical measures of risk and return (Markowitz, 1952). At first, MPT attracted minimal attention, but over time, the financial community fully embraced the concept (Fabozzi et al., 2002). Now, financial models de- rived from those foundational ideas are continuously being updated to include all the new insights generated from that groundbreaking research. In the center of Modern Portfolio Theory lies Markowitz's Portfolio Selection Model, fo- cusing on diversification as a means to reduce risk. According to Fabozzi et al. (2002), 17 MPT measured the idea of diversification by introducing the statistical concept of covar- iance or correlation. Essentially, this principle suggests that investing all your funds in assets that are likely to fail simultaneously. This means that their returns are closely re- lated, regardless of how low the probability is that any individual investment will fail. Fabozzi et al. point out that this is because if one particular investment fails, it is highly probable that its strong correlation with the other investments will also lead to those investments failing. Eventually resulting in the collapse of the entire portfolio. The paper by Markowitz proposes that investors can optimize portfolios by balancing the trade-off between risk and return or mean-variance to construct an efficient frontier of optimal portfolios (Markowitz, 1952; Fabozzi et al., 2002). Fabozzi et al. (2002) noted that this frontier is efficient because each point on it represents a portfolio that achieves the highest expected return for a given level of risk or the lowest risk for a specific level of expected return. The portfolios located on this frontier establish the set of efficient portfolios. Expanding on Markowitz's research, Sharpe (1964), Lintner (1965), and Mossin (1966) developed the Capital Asset Pricing Model (CAPM), which builds upon Modern Portfolio Theory to quantify the link between systematic risk and expected returns (Fama & French, 2004). The CAPM operates under the premise that investors maintain a diversi- fied portfolio to eliminate unsystematic risk, indicating that only systematic risk (market risk) should affect asset prices. The model's formula forecasts an asset's expected return based on the risk-free rate, the asset's beta (which measures its sensitivity to the market), and the market risk premium (Fabozzi et al., 2002; Sharpe, 1964). The CAPM strengthened the MPT concept of the risk-return trade-off, highlighting that investors should receive compensation solely for assuming systematic risk since unsys- tematic risk can be mitigated through diversification in an efficient portfolio (Fama & French, 2004). Sharpe (1964) notes that the CAPM serves as a standard for assessing 18 asset performance and reinforces the principle that assets with higher risk (indicated by beta) are expected to provide greater returns over time. Despite its widespread adoption, Modern Portfolio Theory has faced criticism, largely regarding its assumptions about behavioral aspects and market conditions. A major crit- icism is that MPT depends on past correlations and returns to predict future perfor- mance, which might not represent actual market behavior (Jagannathan & Ma, 2003). The authors Jagannathan and Ma (2003) contend that applying specific limitations on portfolio weights can improve diversification and decrease risk, which questions MPT’s dependence on optimization without constraints. DeMiguel et al. (2009) also studied an optimally diversified portfolio against the “naïve” equally weighted portfolio (1/N strategy) and found that the sample-based mean-vari- ance portfolio performed comparably or even worse. According to the authors, MPT- optimized portfolios may lead to errors when applied to real-world data, making them less effective than the 1/N strategy. Furthermore, MPT has also been questioned by behavioral economists, arguing that in- vestors often behave irrationally and might not optimize the risk-return relationship in the way MPT assumes. Research in behavioral finance suggests that cognitive biases, such as overconfidence and loss aversion, can impact investment decisions, leading to deviations from MPT’s predictions (Malkiel, 2003). Despite receiving criticism from vari- ous perspectives, MPT has significantly influenced the portfolio management industry. Following the traditional finance theories, this study will next present the theoretical framework for the behavioral finance component. 19 2.2 Behavioral Finance The conventional structure of finance is attractively simple, and it would be gratifying if its forecasts were validated by the data (Barberis & Thaler, 2003). After extensive re- search, it became regrettably evident that fundamental truths regarding the overall stock market, the variation of average returns, and excessive market volatility are not easily interpreted within this framework (Shiller, 2003; Barberis & Thaler, 2003). This suggests that psychological and emotional factors play a major role in the financial market. During the 1980s, the field of behavioral finance emerged, integrating behavioral and psychological elements into the processes of economic and financial decision-making (Kumar & Goyal, 2015). This discipline questions the efficient market hypothesis and aids in understanding the reasons behind the specific behaviors of investors when investing in financial assets. According to Barberis & Thaler (2003), behavioral finance represents a new perspective on financial markets that has developed as a reaction to the chal- lenges encountered by the conventional paradigm. It suggests that certain financial events can be more effectively interpreted using models where some agents are not en- tirely rational. To be precise, it investigates the implications of loosening one or both of the two principles that form the basis of individual rationality. The authors Barberis and Thaler (2003) state that behavioral finance consists of two ar- eas: limits to arbitrage and psychology. Limits to arbitrage posit that rational investors may struggle to correct the fluctuations caused by less rational investors. The psycholog- ical building block of behavioral finance, on the other hand, captures the various types of deviations from complete rationality that are anticipated to be encountered. A key factor contributing to the fast acknowledgment of behavioral finance is the evolu- tion of the prospect theory. Kahneman and Tversky (1979) introduced the prospect the- ory, which serves as an alternative to Expected Utility Theory in understanding decision- making in uncertain situations (Kumar & Goyal, 2015). In their paper, Kahneman and Tversky oppose the EUT for being utilized as a descriptive model, even though it has been 20 welcomed as a normative theory. Therefore, the model presumed that all rational indi- viduals would prefer to follow the theory's axioms and that, most of the time, the ma- jority of people generally do so. The authors describe various instances where the pref- erences of individuals contradict the assumptions of the Expected Utility Theory. For ex- ample, the paper illustrates that individuals tend to put weight on outcomes that are regarded as foolproof compared to solely probable. In this instance, the preferences were examined only between positive probabilities. Additionally, to support their claim, Kahneman and Tversky also portray that the EUT is similarly violated when the probabil- ities are switched to negative counterparts. The research conducted by Kahneman and Tversky has led to greater acceptance of the cognitive aspects of behavioral finance. This has resulted in the introduction of various behavioral biases, with herding behavior being the primary focus of this thesis. A few other common behavioral biases include overconfidence, mental accounting, disposition effect, and heuristics (Ritter, 2003). In addition to behavioral biases, agency theory ad- dresses the most favorable form of contract to handle relationships. In agency theory, one ‘principal’ (e.g., investor) delegates responsibility to another ‘agent’ (e.g., fund man- ager) (De Camargo Fiorini et al., 2018). While behavioral biases explain cognitive actions that impact decision-making, agency theory provides an additional explanation based on incentive structures. This relationship in the fund management context may lead to biased behavior, such as herding. Although agency theory is not behavioral, it can create similar irrational outcomes and is relevant when comparing human-managed funds to AI alternatives. Next, the main topic of herding will be further explored. 2.2.1 Herding in Financial Markets The concept of herding is observed in a variety of fields, ranging from neurology and zoology to sociology, psychology, economics, and finance. In the areas of economics and finance, herding or herd behavior typically refers to the tendency of economic agents to copy one another's actions and/or make decisions influenced by the actions of others 21 (Spyrou, 2013). In the fields of finance and economics, herding behavior has been largely examined and documented in various market conditions (Alexakis et al., 2023). Accord- ing to the authors Alexakis et al. (2023), herding is a fairly typical behavioral bias that involves trading actions that are correlated and driven by imitation, regardless of indi- vidual insights, opinions, or evaluations. The context of herding behavior dates back to 1936 when Keynes first rationalized the decision-making process through psychological aspects (Bekiros et al., 2017). According to the paper by Bekiros et al. (2017), Keynes outlines herd behavior as an act mimicking “animal spirits” or “beauty contests.” The work by Keynes emphasizes that when facing uncertainty or asymmetric information, individuals tend to choose assets with high short-term market values instead of those with solid fundamentals. Additionally, the the- ory of “animal spirits” suggests that during uncertain times, fear, and panic, people are subject to their instincts rather than reason, leading to irrational behavior. Moreover, under uncertainty, herding potentially creates irrational optimism that may evolve into a bubble in the market. Thus, it can be stated that herd activity enables the growth of volatile markets. There are various studies conducted on the topic of herding. In each of the studies, an own definition is provided. For instance, herding can refer to a group of investors who are buying or selling in the same direction at the same time (Nofsinger & Sias, 1999). Avery and Zemsky (1998) note that it can also be market participants who disregard their previous analysis and trade by mimicking the previous trade’s trend. There may also be mutual imitation or significant consensus among analyst forecasts (Welch, 2000; DeBondt & Forbes, 1999). Moreover, herding can show in exhibiting a certain level of correlated behavior (Hwang & Salmon, 2004). Sias (2004) notes that a group of investors following one another into or out of the same securities can also be considered as herd behavior. According to the studies mentioned above, the definition of herding remains mostly the same, based on the main focus of the study. 22 In addition to conventional herding, recent empirical research has shown instances of reverse herding (Bekiros et al., 2017). In their paper, Bekiros et al. (2017) detail that in reverse herding, investors behave contrary to the prevailing market consensus. Accord- ing to the authors, this results in an unusually high degree of variation in cross-sectional returns. In the mentioned scenario, market participants may have overlooked or under- valued their own insights and information in favor of distinct price movement signals coming from the market. This subsequently aligns their beliefs with a prevalent "domi- nant view" (a specific group of assets), resulting in significant herding behavior in and out of positions. The authors Bekiros et al. highlight that this pattern suggests price rises (falls) occur in an "excessive manner," leading to a systematic increase in the dispersion of market returns beyond what would be considered rational pricing. As stated, herding behavior has implications in a variety of fields. In a more recent study, Alexakis et al. (2023) highlight that herding behavior significantly impacts asset pricing, market efficiency, portfolio diversification, and overall market stability in the financial markets. According to Chiang and Zheng (2010), asset prices can deviate from their fun- damental values due to imitative trading activities. This leads to a breakdown in market efficiency. Furthermore, this type of herding can lead to inadequately diversified portfo- lios, increasing exposure to difficult-to-hedge risks. Consequently, market participants must recognize such correlated trading behaviors to modify their asset allocation strate- gies accordingly. Alexakis et al. (2023) affirm that this herding tendency is often amplified during crises, potentially causing cross-market herding and a contagion effect. The behavior of imitating others and trading in the same direction might be perceived as rational on an individual level (Hwang & Salmon, 2004). However, such behavior might not necessarily result in efficient market outcomes, as stated earlier. The study con- ducted by Bekiros et al. (2017) finds that herd behavior is often heightened during ex- treme market situations. The authors observe that during periods of significant volatility, herding is detected at the start and turns to unimportant closer to the end of the crisis. Similarly, in their study, Chiang and Zheng (2010) reinforce that a crisis initiates herding 23 behavior in the country where it originates, leading to a contagion effect that extends the crisis to adjacent nations. It can be stated that herding behavior is largely related to uncertainty in the market but not the root cause. 2.2.2 Dimensions of Herding Behavior In their study, Bikhchandani and Sharma (2000) classify herding into two distinct types in the financial markets, depending on the way investors respond to new information. Spurious herding happens when groups facing similar decision-making issues and infor- mation are inclined to arrive at similar decisions. Intentional herding, on the other hand, arises from a clear intention among investors to imitate the behaviors of their peers. The authors Bikhchandani and Sharma state that spurious herding represents an efficient outcome, while intentional herding may not yield the same efficiency. However, it is im- portant to note that empirically differentiating between spurious herding and inten- tional herding is often more challenging than it seems. The differentiating might even be unattainable, as a variety of factors can influence an investment decision. Spurious herding, often referred to as unintentional herding, takes place when investors independently arrive at similar trading choices (Indārs et al., 2019). These choices are influenced by common information or external circumstances without directly mimick- ing others’ behavior. According to Indārs et al. (2019), this type of herding emerges from "homogenous trade decisions in line with company fundamentals", where investors re- spond comparably to the same information set, such as earnings reports or macroeco- nomic indicators. In these instances, herding is fueled by fundamental factors and the decisions are based on objective and rational information. Nonetheless, spurious herd- ing can also arise from information inaccuracies, behavioral biases, and typical trading patterns that lead investors to behave similarly. This can happen even when their deci- sions are derived from misleading or incomplete data (Indārs et al., 2019). This non-fun- damental variant of spurious herding highlights how cognitive biases may lead investors 24 to interpret information. Similarly, this results in synchronized market actions that are not directly founded on fundamentals. Intentional herding, on the other hand, refers to a deliberate choice made by investors to copy the behavior of other market participants. This phenomenon is marked by a "strong willingness of investors to replicate the actions of others" (Indārs et al., 2019). Intentional herding behavior could be further divided into either rational or irrational behavior among economic actors (Wang et al., 2021). The study by Indārs et al. (2019) notes that investors are regarded as engaging in rational herding when they base their imitation of investment choices on fundamental information. This rational dimension of intentional herding is driven by informational, compensational, or reputational motiva- tions. Investors believe that imitating informed traders will yield favorable results (Indārs et al., 2019). Rational herding contributes to the incorporation of new information into prices, bringing them closer to their fundamental values (Wang et al., 2021). Conversely, intentional herding can also take on an irrational form when investors replicate decisions based on non-fundamental elements, such as noise or speculative trends. This happens without considering the true value of assets. Indārs et al. (2019) note that this version of herding can result in "erroneous investment decisions being copied." Irrational herding can also lead to price deviations from fundamental values, causing increased volatility and reducing market efficiency. In the worst case, possibly leading to the creation of as- set bubbles (Wang et al., 2021; Indārs et al., 2019). 2.2.3 Drivers of Herding Behavior The motivations for herd behavior can vary widely (Spyrou, 2013). As indicated earlier, market players might draw conclusions based on the actions of earlier participants, or investors may respond to the emergence of fundamental information. Also, analysts might follow the crowd to safeguard their reputation, and institutional investors may conform for reasons tied to compensation. Some investors might just act irrationally, and herd behavior can emerge due to psychological or social norms. As stated in the chapter 25 on dimensions of herding, intentional herding includes rational and irrational herding behavior. Rational herding stems from inadequate information, reputational concerns, and the structure of compensation. In turn, irrational herding is driven by psychological factors, such as behavioral biases. Spurious or unintentional herding, conversely, is driven by characteristics and relative homogeneity (Holmes et al., 2013). One of the primary factors driving rational herding is informational cascades. As ex- plained by Bikhchandani et al. (1992), an informational cascade happens when it be- comes advantageous for a person, having observed the choices of those before him, to mimic the behavior of the previous individual without considering their private infor- mation. Devenow and Welch (1996) enhance this explanation with an illustration: An investor who possesses the most negative private information might still be influenced to buy if they observe that three other investors have already made purchases (the knowledge of these three transactions may outweigh their negative private information). However, afterward, all following investors will lack any additional positive information and will behave similarly, even though everyone recognizes that these purchases (or fu- ture ones) provide no informative value. This model can account for widespread herding (clumping) based on an erroneous decision, and low consensus (when surveyed). It can also lead to vulnerability (where a small amount of public information can overturn long- standing trends since cascades can form from very little publicly shared data), and a strong reliance on initial conditions (specifically, the purchase choices of the first few investors). In addition to informational cascades, rational herding can be driven by the reputational concerns of the investor (Bikhchandani & Sharma, 2000). According to Bikhchandani and Sharma (2000), reputational concerns or career concerns stem from doubts regarding a specific manager's skills or abilities. The fundamental concept, originally introduced by Scharfstein and Stein, is that if both an investment manager and the employer are unsure of the manager's capability to select the right stocks, clinging to the actions of other investment professionals maintains the uncertainty about the manager's capacity to 26 handle the portfolio. This situation works favorably for the manager, and if other invest- ment professionals find themselves in a comparable position, herding behavior is likely to take place. Furthermore, Bikhchandani and Sharma (2000) address the third driver of rational herd- ing, compensation-based herding. If the earnings of an investment manager rely on their performance relative to that of their competitors, this distorts the manager's motiva- tions, resulting in a portfolio that is ultimately inefficient. The addressed situation might lead to herding behavior. The authors Maug and Naik (1995) examine compensation- based herding utilizing an example: an investor (the agent) who is risk-averse and whose remuneration is tied to their performance in comparison to a benchmark, which might be the returns of other investors or a market index. Both the agent and the benchmark possess limited private information regarding stock returns. The benchmark investor makes the first move, and the agent, upon observing these actions, has a motivation to imitate. According to Maug and Naik, this leads to them modifying their portfolio to more closely match that of the benchmark. Furthermore, the structure of the agent's compensation incentivizes them to replicate the benchmark's portfolio instead of solely basing their trading decisions on their account. Maug and Naik (1995) clarify that the compensation imposes penalties for underperformance against the benchmark. Informational cascades, reasons related to reputation, and herding based on compensa- tion indicate that herding is conducted with intention (Holmes et al., 2013). Contrarywise, herding might branch from relative homogeneity among investment professionals (for instance, similar educational and social backgrounds). This could lead them to interpret the same information signals similarly. Likewise, characteristic herding includes institu- tions being drawn to investments that possess specific characteristics, such as high his- torical returns (for example, momentum trading). According to Holmes et al., the argu- ments regarding relative homogeneity and characteristics imply that herding may occur without intent. This leads to a thought-provoking and applicable question arising about 27 whether herding behavior results from a deliberate choice to follow or if it happens un- intentionally. 2.3 Artificial Intelligence in Finance The landscape of asset management is constantly changing, driven by advancements in technology and investors' changing preferences (Stier, 2022; Bartram et al., 2020). In the article by Deloitte, Stier states that traditional methods of differentiation in investment management are becoming more standardized. This is when artificial intelligence (AI) offers fresh opportunities that go beyond merely lowering costs and improving opera- tional efficiency. According to Bartram et al. (2020), AI has enhanced portfolio management, trading, and risk management practices. The evolution has been achieved by boosting efficiency, ac- curacy, and compliance. The authors Bartram, Branke, and Motahari specify that AI methodologies assist in portfolio construction by providing more accurate forecasts of risk and returns while having more sophisticated limitations. According to the three au- thors, trading algorithms leverage AI to create innovative trading signals and execute trades with reduced transaction expenses. Moreover, Bartram et al. continue that AI also enhances risk modeling and forecasting by deriving insights from new data sources. Ul- timately, the success of robo-advisors is largely attributed to AI techniques. However, Bartram et al. emphasize that the application of AI also introduces new risks and chal- lenges generated from model opacity, complexity, and dependence on data integrity. To begin the chapter about artificial intelligence, an overview of AI is presented to pro- vide a foundation for the finance context. From there, the study moves on to examine AI’s applications in the decision-making of investment as well as its biases and limitations to support the viewpoint of this thesis. 28 2.3.1 Overview of Artificial Intelligence To thoroughly grasp the advantages and limitations of AI in fund management, it is cru- cial to delve into the mechanics and framework of AI. By investigating the core algo- rithms and decision-making frameworks employed by AI systems, it is possible to assess the dependability and efficiency of AI-driven strategies in fund management. Gaining insight into the mechanics of AI also allows for the tackling of issues related to transpar- ency, interpretability, and ethical considerations linked to the application of AI in finance. The first idea for artificial intelligence dates back to 1956, when a team of computer scientists suggested that computers could be programmed to reason as well as think (Choi et al., 2020). According to Choi et al. (2020), this suggests that every part of learn- ing or any other trait of intelligence could theoretically be described in such detail that a machine could be developed to replicate it. This idea was referred to as “artificial in- telligence”. Artificial intelligence can be defined as a focus on creating machines that exhibit intelligent behavior (Stone et al., 2022). According to the authors Stone et al. (2022), intelligence refers to the ability that allows an entity to operate effectively and with anticipation in its surroundings. The study refers to AI as a science and a combina- tion of computational technologies. These areas are encouraged by the ways humans utilize their physical bodies and nervous systems to perceive, learn, reason, and act. However, Stone et al. emphasize that even though AI draws inspiration from human na- ture, its operation differs from it. Although advancements in AI have been uneven and difficult to forecast, considerable progress has been made since the field was established sixty years ago. In general, AI can be categorized into two types: deductive and generative (Arora et al., 2023). Deductive AI refers to computer systems that evaluate extensive datasets to iden- tify patterns and draw conclusions. On the other hand, in their recent study, Arora et al. mention that generative AI learns from current data to create new material. From these two types, generative AI has gathered a lot more attention in the past couple of years. As generative AI continues to evolve, speculation around its potential for self-design and 29 iterative algorithm improvement has gained traction, raising the prospect of artificial general intelligence (AGI) outpacing human intelligence and influencing the future or- ganization of work. Arora et al. (2023) highlight that these concepts have traditionally belonged to the realm of science fiction. However, the capacity of large language models to create computer code that exceeds the skills of many data scientists has heightened both enthusiasm and apprehension regarding AGI. Choi et al. (2020) describe that AI is part of data science and includes both traditional programming and machine learning (ML). Machine learning consists of various models and techniques, which include deep learning (DL) and artificial neural networks (ANN). This division is portrayed in the figure below: Figure 1. Dimensions of Select Data Techniques (Choi et al., 2020). In the field of Machine Learning (ML), a computer receives a set of inputs (variables and datasets) along with an output that results from those input variables (Kolanovic & Krish- namachari, 2017). The machine then identifies or ‘learns’ a relationship that connects the input to the output. Ultimately, Kolanovic and Krishnamachari (2017) state that the effectiveness of this learning process is evaluated ‘out of sample’. This assesses its capa- 30 bility to acquire valuable insights into the relationships among variables and forecast re- sults in previously unseen scenarios. ML can further be categorized into supervised learning, unsupervised learning, and reinforcement learning (Lotfi & Bouhadi, 2021). In the first category, Supervised Learning (SL), the aim is to find a rule or an ‘equation’ that could be used to forecast a variable. For example, the idea might be to identify a momentum indicator that can most effectively forecast future market performance. Ko- lanovic and Krishnamachari (2017) mention that this can be achieved by utilizing sophis- ticated regression models to evaluate which model demonstrates greater predictive ca- pability and holds the most resilience to changes in the regime. Furthermore, in Unsu- pervised Learning (UL), the focus is on revealing the construct of data. For reference, market returns could be gathered to distinguish the main market drivers. Kolanovic and Krishnamachari explain the matter further by providing an example: A thriving model might discover that, at a certain moment, the market is influenced by factors such as momentum, energy prices, the value of the USD, and an emerging factor that could be associated with liquidity. Lastly, according to Latfi and Bouhadi (2021), Reinforcement Learning (RL) enables agents to acquire knowledge on how to act within an environment filled with error. To put it simply, the algorithm analyzes the outcomes of its actions and learns from the mistakes made when undesirable results occur. This concept is referred to as the reward principle. The reward principle seeks the optimal actions to maximize the rewards. Latfi and Bouhadi state that this learning approach is fundamentally differ- ent from classification or planning, as it operates in an unknown and potentially chang- ing environment, and it is achieved without any supervision. Deep Learning (DL) is a part of Machine Learning, as illustrated in Figure 1 above. As described by Kolanovic and Krishnamachari (2017), deep learning is a method of ML that processes data through several layers of learning, hence the term ‘deep’. It typically be- gins by understanding simpler concepts and then integrates these to grasp more com- plex ideas and abstract concepts. It is often said that automation’s objective is to carry out tasks that are simple to outline but tedious to execute for humans. Alternatively, the 31 purpose of DL AI systems is to tackle tasks that are challenging for humans to specify yet straightforward to perform. The authors Kolanovic and Krishnamachari state that DL es- sentially mirrors the way humans acquire knowledge. This represents a true effort to artificially replicate the intelligence of humans. However, in the context of finance, Dixon and Halperin (2019) state that in areas other than natural language processing or text mining, DL is often not essential for most applications in finance. It rather serves as a helpful convenience. The last segment of Figure 1, Artificial Neural Networks (ANN) are machine learning al- gorithms (Choi et al., 2020). According to Chen et al. (2019), ANNs are based on the functionality and structure of biological neural networks that can draw knowledge from intricate or inaccurate data. The authors further clarify that within the climate of wireless communications, ANNs can be utilized to forecast and explore networks as well as the behavior of users. Thus, it provides user data for unraveling various wireless networking issues. For instance, ANNs can solve problems related to cell association, allocation of computational resources, spectrum management, and replacement of sealed content. Furthermore, recent advancements in the smart mobile applications industry have con- siderably accelerated the level of interaction with mobile systems for human users. Chen et al. inform that an educated ANN can be categorized as a “professional” in operating with human-related information. Thus, employing ANNs to gather insights from the user’s surroundings can enable a wireless network to anticipate users’ future actions. As a result, this can formulate an optimal strategy to enhance the overall quality of service and reliability. Lastly, Chen et al. point out that there is a diverse set of ANNs. Every kind of ANN is according to the paper suitable for a specific learning task. The authors men- tion Modular Neural Networks, Recurrent Neural Networks, Generative Adversarial Net- works, Deep Neural Networks, Spiking Neural Networks, Feedforward Neural Networks, and finally Physical Neural Networks. 32 2.3.2 Artificial Intelligence’s Applications in Fund Management Artificial intelligence in the field of finance has been a topic of interest in research for decades (Cao, 2021). For instance, financial markets, the banking industry, the insurance industry, risk management, regulation, and trading have evolved to the next generation of Financial Technology (FinTech). According to the study by Cao, FinTech has made smart digital currencies, lending, asset and wealth management, and accounting possible. In addition to FinTech, economics and finance (EcoFin) are also being linked with AI at a growing pace. Among the sectors impacted by FinTech, asset management is arguably the most significantly affected and is anticipated to experience the greatest number of job reductions in the near future (Bartram et al., 2020). Many asset management firms are currently employing AI and statistical models to operate trading and investment plat- forms. The growing reliance on AI for a variety of tasks within asset management neces- sitates a more thorough analysis of the different techniques and applications. Addition- ally, it demands the associated opportunities and challenges they present to the industry. As outlined in the overview section, Artificial intelligence predominantly focuses on ma- chine learning. Machine learning currently stands as the most widely used method in the field (Bartram et al., 2020). As stated by Bartram et al. (2020), there is a significant number of applications for ML. Additionally, most ML techniques in asset management and finance primarily depend on a few key methods. These encompass artificial neural networks, cluster analysis, decision trees, random forests, evolutionary algorithms, least absolute shrinkage and selection operator (LASSO), support vector machines, and natu- ral language processing (NLP). Following in Figure 2, the key characteristics of these methods are: 33 Figure 2. Main AI Methods in Asset Management (Bartram et al., 2020) As outlined by Bartram, Branke, and Motahari (2020), the area of asset management includes three main elements in which AI has an impact. The elements are portfolio man- agement, trading, and risk management, all of which play a crucial role when it comes to fund management. In portfolio management, AI plays a pivotal role in generating al- pha and managing sigma. It also transforms portfolio optimization. In the area of trading, AI applies to algorithmic trading, transaction cost analysis, and trade execution. Lastly, 34 in portfolio risk management, Bartram et al. state that AI contributes to the contexts of market and credit risk. In portfolio management, AI methods can be utilized to conduct advanced fundamental analysis and to enhance optimal asset allocations (Bartram et al., 2020). Considering the various difficulties faced by traditional portfolio optimization techniques, AI methods frequently yield improved estimations of both returns and covariances. These can then be applied within classic portfolio optimization models. AI applications can predict ex- pected returns via LASSO, neural networks, and support vector machines. Also, using neural networks and support vector machines, AI can estimate variances. Additionally, Bartram et al. point out that the traditional covariance matrix can be substituted with a tree structure utilizing hierarchical clustering. Furthermore, AI can be directly used for making asset allocation decisions to build portfolios that more accurately achieve per- formance objectives. Artificial intelligence employs genetic algorithms to address opti- mization challenges within intricate constraints. Finally, neural networks are capable of directly generating optimal portfolios or those that, for instance, replicate an index. AI technologies are increasingly applied in trading (Bhuyan & Singh, 2022). Due to the rising speed and complexity of transactions, AI methods are becoming essential in the trading arena. As stated earlier, AI has implications in algorithmic trading, transaction cost analysis, and trade execution. One of the most attractive aspects of AI is its ability to process large amounts of data and generate trading signals. As a result of these signals, a new field known as algorithmic trading has developed. According to Bhuyan and Singh, algorithmic trading can be programmed to carry out trades automatically. Additionally, AI techniques can further reduce transaction costs by automatically evaluating the mar- ket and identifying the optimal timing, size, and placement for trades. In trade execution, AI systems consider market dynamics, liquidity limitations, and execution risks to guar- antee that trades are executed effectively (Bartram et al., 2020). Through continuous learning from market feedback, AI systems can adjust their execution strategies to opti- mize cost efficiency and sustain a competitive edge. 35 AI’s contributions to portfolio risk management are transformative, particularly in its ability to model and forecast risks more precisely (Bartram et al., 2020). In the area of market risk, AI systems analyze complex datasets to predict market volatility and assess macroeconomic trends. According to Aziz and Dowling (2018), ML is especially applica- ble in market model stress testing to distinguish either accidental or developing risks in trade behavior. Aziz and Dowling remark that due to insufficiencies in conventional credit risk management techniques, there is a growing interest in the application of AI and ML methods. Bartram et al. (2020) explain that AI provides valuable resources for assessing the financial stability of counterparties and forecasting the likelihood of defaults. AI models deliver detailed insights into credit exposures by examining past data, financial reports, and even unstructured information like news articles. 2.3.3 Artificial Intelligence’s Limitations and Biases Artificial Intelligence has significantly shaped the financial sector by creating new possi- bilities and offering advancements in efficiency, risk management, and customer service. However, the integration of AI does not come without challenges, especially concerning the limitations of AI and biases that can impact its decision-making processes. The im- plementation of AI also poses risks in the frontiers of ethicality, fairness, and technicality. According to an article by IBM (2023), artificial intelligence bias, often known as machine learning bias or algorithmic bias, pertains to AI systems that yield biased outcomes. This mirrors and continues existing human biases in society, including both historical and con- temporary social inequalities. Bias may arise from the original training data, the algo- rithm itself, or the predictions generated by the algorithm. In training data bias, AI sys- tems develop decision-making capabilities based on the training data they receive. Thus, if the data is flawed to start with, the bias presents itself in the decision made by the AI system. 36 The article by IBM explains that the algorithmic bias stems from the usage of flawed training data. This may lead to algorithms that continuously generate mistakes and un- just results or even exacerbate the biases present in the inaccurate data. The article clar- ifies that algorithmic bias can also stem from programming mistakes. In this scenario, a developer might disproportionately weigh certain factors in the algorithm's decision- making due to their own conscious or subconscious biases. For instance, variables like income or vocabulary could be utilized by the algorithm, unintentionally leading to dis- crimination against individuals of a particular race or gender. The article also informs that an AI bias can arise from human errors. For example, when people handle information and make decisions based on it, they are unavoidably af- fected by their personal experiences and preferences. Therefore, these biases might be built into the AI systems via the selection process or the weight of data. Similarly to be- havioral finance, this is referred to as a cognitive bias. The IBM article visualizes that this cognitive bias may cause favoring of certain geographic areas, for example, while gath- ering datasets. Human decision-makers rarely prioritize maximizing predictive accuracy above all else. They often consider whether the attributes used for predictions have moral significance (Barocas et al., 2023). This is one reason to approach comparisons that suggest the su- periority of statistical decision-making with caution. As mentioned by Barocas et al. (2023), humans are also less likely to make unreasonable choices. However, such occur- rences could arise in automated decision-making, possibly due to inaccurate data. These differences between human and automated decision-making highlight why machine learning-based decision-making systems may be unjust. 37 3 Literature Review In this section of the research, the thesis provides a thorough summary of relevant pre- vious studies. It offers a comprehensive literature review that plays a crucial role in shap- ing the hypothesis. This review will identify gaps and limitations in the existing literature and illuminate areas where further investigation is needed. This paper delves into stud- ies focusing on herding behavior. It explores how collective actions influence individual decision-making, alongside research related to the advancements and implications of artificial intelligence. Moreover, the analysis will extend to studies on fund management and the decision-making processes involved. This ensures a well-rounded perspective on the overarching themes presented in the thesis. Through this detailed examination, this thesis aims to lay a solid foundation for the development of the hypotheses. 3.1 Overview of Herding Behavior in Financial Markets As the theoretical background section already insinuated, herding behavior has gathered a lot of attention from finance researchers and other fields of science. Even though herd- ing in the financial markets is a widely discussed topic among academics, the empirical evidence is surprisingly scarce (Welch, 2000). In his study on an overview of herding- related research, Spyrou (2013), in a similar way, states that there are many issues when delving deeper into evidence and limitations of herding behavior. As previously dis- cussed, the theory of herding is categorized into models that suppose rational or close to rational behavior and models that suppose irrational agents. Spyrou continues that this literature gap between rational and non-rational models should be narrowed by uti- lizing models that interact between sources of herding. Meanwhile, various models pro- vide implications that are hard to process empirically with current databases. Secondly, Spyrou remarks that in addition to the difficulties in the models of herding, the empirical evidence tends to be inaccurate. For instance, although numerous significant papers identify little evidence of herding among institutional investors, other research presents conflicting results. Similarly, while several studies report findings that align with analyst 38 herding for various reasons, more recent research suggests that analysts tend to be “anti-herd.” Lastly, Spyrou states in his study that the primary empirical methods used to assess herding possess limitations that might prevent economists from completely grasping the herding process. However, despite the conflicting results considering herding research, various seminal papers have created a foundation for this area. One of these is the paper by Bikhchan- dani and Sharma (2000) that has been addressed previously in this study. The authors distinguish a division between “spurious herding” and “intentional herding.” In addition to Bikhchandani and Sharma (2000), other important works examining the root causes of herding include papers by Banerjee (1992), Hirshleifer and Teoh (2003), and Avery and Zemsky (1998), to name a few. Some other influential studies on herding are conducted by Lakonishok et al. (1992), Christie and Huang (1995), and Chang et al. (2000). These three provide methods for measuring herding behavior in the financial markets. First, Lakonishok et al. (1992) cre- ated the LSV method. Next, Christie and Huang (1995) developed the Cross-Sectional Standard Deviation (CSSD) method a few years later. Lastly, Chang et al. (2000) enhanced the CSSD method by taking the Cross-Sectional Absolute Deviations of returns at the millennium's break. In their groundbreaking study, Christie and Huang (1995) investigate whether returns on equity suggest the prevalence of investor herding behavior during market volatility. To capture herd activity, the authors predict the cross-sectional stand- ard deviation (CSSD) of returns. Contrary to rational asset pricing models predicting in- creasing dispersions, the CSSD model assumes dispersions to be rather low. They dis- cover no consistency with herding behavior in daily or monthly returns during major price deviations. According to Christie and Huang, the expectancy of herding is the high- est during significant down-market swings. However, the extent of the rise in the disper- sion of actual returns corresponds to the rise in the variability of predicted returns de- rived from a rational asset pricing model. 39 In their study, Chang et al. (2000) investigate the herding behavior of investors within varying international markets (i.e., Japan, South Korea, Hong Kong, the US, and Taiwan). While they detect no presence of herding in Hong Kong or the US, they do find robust evidence from the emerging markets of South Korea and Taiwan. However, the main contribution of the study by Chang et al. (2000) is the invention of the Cross-Sectional Absolute Deviation model. Further continuing the work of Chang et al. (2000), Chiang and Zheng (2010) propose an enhanced model to detect herding behavior. In compari- son, their model takes asymmetric investor behavior into account. Using their method, Chiang and Zheng investigate herding in global markets. They detect herding in both up and down markets in advanced and Asian markets. However, the US market does not portray herding during the study period. Evidence of herding behavior is documented by Chiang and Zheng solely during the period of the financial crisis. Similarly, authors BenSaïda et al. (2015) do not find evidence of herd activity in the US market, even though the study period ranges from 2000 to 2014. As the previous literature shows, herding behavior varies significantly based on market conditions, geographic locations, and methodologies used. Key studies, like those by Christie and Huang (1995) and Chang et al. (2000), indicate that herding is more preva- lent in emerging markets than in developed markets like the US. The maturity of the market and investor behavior play essential roles in how herding manifests. While some research links herding to market volatility, others, like BenSaïda et al. (2015), find its ab- sence in the US market. Such inconsistencies highlight the influence of data periods, model assumptions, and unique market dynamics. This suggests that herding is a com- plex phenomenon shaped by specific contexts and a blend of rational and irrational fac- tors. 3.2 Herding Behavior in Volatile Market Conditions Researchers have shown significant interest in the US market, particularly regarding herding during various bubble and crisis periods (Alexakis et al., 2023). As noted earlier, 40 herding behavior is most common during market turbulence. Nonetheless, Alexakis et al. (2023) state that the empirical findings are ambiguous and vary primarily based on the examined period and the methodology used. In their study, Bekiros et al. (2017) examine herd behavior under uncertain market con- ditions and investigate whether it impacts volatility. Utilizing a modified version of the Cross-Sectional Absolute Deviation model, the authors observe signs that herding be- havior tends to intensify during extreme circumstances. Bekiros et al. state that at the onset of a crisis, herding behavior is noted, but it becomes negligible as the crisis pro- gresses. According to their observations, the herding tendencies in the US market display dynamic trading patterns that vary over time. These tendencies can be linked to factors such as overconfidence or a pronounced “flight to quality,” particularly evident after the global financial crisis. Additionally, implied volatility shows asymmetric trends and sig- nificantly influences irrational behavior. In a more recent study, Alexakis et al. (2023) examine herding activity during the COVID- 19 pandemic. Also, utilizing the CSAD method, the authors investigate the volume of trading, herding dynamics, and events to detect herding under market stress. Similarly to Bekiros et al., the authors find evidence indicating that investors portray herding ac- tivity during significant periods of the crisis. The outbreak of the COVID-19 pandemic instigated fear among investors and created overall uncertainty in the markets, making herding behavior more prevalent. Post-pandemic, there is a plethora of research on the effect of herding in financial markets during COVID-19. In a similar way to Alexakis et al., Ferreruela and Mallor (2021) show herding behavior on days with high volatility during the pandemic sub-period in both Spain and Portugal. Exploring the major Asian stock markets, Jiang et al. (2022) also discover evidence of herding behavior during the pan- demic. The authors detect a clear rise in the magnitude of herding around the crash of the 2020 spring. 41 In their research paper, Belhoula and Naoui (2011) investigate herd behavior and posi- tive feedback trading in the US market during times of significant volatility. The pair finds evidence of both positive feedback trading and herding behavior. Galariotis et al. (2015) also find evidence that indicates herding behavior among US investors due to non-fun- damentals and fundamentals during various market crises. The authors Galiatoris et al. also document herding spill-over impacts from the US market to the UK in times of fi- nancial crises. In a later study, Galiatoris, Krokida, and Spyrou (2016) find further evi- dence of herd mentality regarding high liquidity stocks in the US from 2000 to 2015, including multiple shorter periods. The trio states that, particularly during crisis and post- crisis periods, the variance in the average liquidity of the equity market is influenced by the clustering of returns. The studies above indicate that herding behavior becomes more pronounced during pe- riods of market instability, particularly at the beginning of crises, as seen during the global financial crisis and the COVID-19 pandemic. Nevertheless, herding behavior typi- cally lessens as markets regain stability over time (Bekiros et al., 2017; Alexakis et al., 2023). Volatility acts as a major factor influencing herding, with fear and uncertainty heightening irrational decision-making. Additionally, cross-market spillovers, such as those observed between the US and the UK during crises (Galariotis et al., 2015), further illustrate the interconnected nature of herding during chaotic periods, emphasizing its considerable effect on market dynamics. 3.3 Herding and Fund Management Herding behavior is widely studied in the financial markets from diverse perspectives. Herding has also been extensively researched from a fund management and perfor- mance perspective, which aligns with this thesis topic. In their influential 1992 research, Lakonishok et al. studied tax-exempt funds to detect herd behavior and positive feed- back trading. Their paper revealed that pension fund managers tend not to participate 42 in positive feedback trading or herd activity. However, Lakonishok et al. did discover herd behavior in smaller stocks, although no impact on stock price movement was found. Herding behavior among mutual funds has also been studied through analyst recom- mendations and its impact on stock prices. The research by Brown, Wei, and Wermers (2014) indicates that mutual funds tend to "herd" by making similar trades in response to analyst upgrades that are widely supported. Additionally, they collectively withdraw from stocks that receive broad consensus downgrades from analysts. As stated by Brown et al., the impact of changes in analyst recommendations on fund herding is more pro- nounced in cases of downgrades and among managers who have heightened career con- cerns. According to Brown et al., changes in analyst recommendations prompt herding behavior among career-focused fund managers, resulting in price instability as mutual fund ownership of stocks rises. Continuing with the review of fund managers’ herding behavior, Wylie (2005) similarly documents herding activity among mutual fund managers in the U.K. The author notes that the observed herding behavior reflects a rise in the number of managers trading a specific stock over time. This increase is more pronounced for both the smallest and largest stocks. Wylie points out that the documented herding is similarly found among studies of U.S.-based pension and fund managers. According to the study, herding is more pronounced at the stock level rather than at the industry level. The study by Wylie indicates that herding behavior among fund managers does not necessarily vary depend- ing on the geographical location. Moreover, the study notes that the investment man- agement industry in the U.K. is comparable to that in the U.S. In a more recent study, Jiang and Verardo (2018) investigate mutual funds and the pos- sible skill behind herding behavior. The study reveals an inverse relationship between herding behavior and skill within the mutual fund sector. Jiang and Verardo observe that fund managers tend to mimic the trading activities of institutional investors. The authors 43 indicate that funds exhibiting herding behaviors outperform their antiherding counter- parts by more than 2% annually. Variations in skill are the reason behind this perfor- mance disparity. Antiherding funds excel in making investment choices, even for stocks that institutions do not frequently trade, and they can foresee the actions of the crowd. Additionally, the performance difference between herding and antiherding funds is en- during, more pronounced when skill holds greater value, and is larger among managers who have heightened career concerns. The literature suggests that herding behavior in fund management can have positive and negative effects. Its influence on fund performance is significantly dependent on factors like managerial expertise, prevailing market conditions, and outside influences. These results highlight the necessity of delving deeper into the consequences of herding, es- pecially when comparing human-managed funds with those driven by AI. Next, this the- sis delves deeper into the use of AI in fund management and its performance compared to human-managed funds. 3.4 Human vs. Artificial Intelligence in Fund Management The comparison between human- and AI-managed funds has raised questions in recent years’ literature. AI was introduced to fund management in 2017 when the AIEQ-named fund was founded (Chen & Ren, 2022). The fund raised over 70 million USD in a few weeks due to its popularity. This was the first step of AI’s inclusion in the asset manage- ment industry. From then on, the amount of AI-powered funds has grown rapidly. In their paper on the performance of AI-powered mutual funds, Chen and Ren (2022) compare the advantages and disadvantages between human- and AI-powered funds. Ac- cording to the authors, the benefits of AI in fund management are well established, em- phasizing its advantages over human managers. AI can quickly and efficiently process large amounts of data, facilitating advanced predictive modeling and decision-making, 44 as noted by Krauss et al. (2017) and Adcock & Gradojevic (2019). Unlike humans, AI op- erates without the limitations of cognitive biases, optimizing decision-making through data learning (Bazley et al., 2020; Linnainmaa et al., 2021; D'Acunto et al., 2019). The decline in the performance of human-managed funds and fewer skilled managers has increased interest in the potential of AI technologies (Barras et al., 2010; Ratanaban- chuen & Saengchote, 2020; Chen & Ren, 2022). However, Chen and Ren (2022) argue that AI funds face challenges. Particularly, chal- lenges in questioning whether they genuinely outperform traditional financial models in predicting stock returns. While machine learning models can identify useful predictive signals, their additional predictive strength compared to traditional methods is often limited (Gu et al., 2020). Furthermore, the high trading frequency associated with AI portfolios raises concerns about transaction costs and net returns. Research by Carhart (1997) indicates that actively managed funds often underperform after accounting for these costs, prompting the need to evaluate whether AI funds achieve an optimal bal- ance between returns and turnover. These factors reveal the potential and limitations of AI in fund management, highlighting the need for further investigation into its perfor- mance across various market conditions. Grobys, Kolari, and Niang (2022) study the confrontation between artificial intelligence and human involvement in hedge fund management. The trio employs partially hand- collected data and forms sample hedge funds into different categories based on the level of automation. The results of the paper state that the hedge funds with the least amount of human involvement perform best. Additionally, the authors observe that a zero-cost strategy contrasting man with machine, where a long position is taken that utilizes the most automation and a short position where human involvement is the highest, pro- duces a notably significant spread of at least 50 basis points each month. Grobys et al. conclude that the degree of automation is a crucial factor in the profitability of the hedge fund sector. 45 In a slightly more recent study, Cao et al. (2024) find conflicting evidence between AI and humans in stock picking. The authors state that an AI analyst designed to process corpo- rate disclosures, industry trends, and macroeconomic data outperforms most analysts in predicting stock returns. However, Cao et al. emphasize that humans excel in scenarios where institutional knowledge is critical, such as with intangible assets and financial dif- ficulties. AI tends to excel when dealing with transparent yet large volumes of infor- mation. Human input offers considerable added value in a combined AI and human ap- proach, which also significantly minimizes extreme errors. When “alternative data” be- comes available, human analysts can catch up to machines if their organizations invest in AI capabilities. Recorded relationships between machines and humans demonstrate how people can leverage their strengths to better adapt to the increasing growing abili- ties of AI. Harvey et al. (2017) examine and compare the performance of discretionary versus sys- tematic hedge funds. Systematic funds implement strategies based on rules, with mini- mal or no daily human involvement. From 1996 to 2014, the performance of both sys- tematic and discretionary managers appears similar after accounting for volatility and factor exposures. Similarly, Abis (2020) investigates the discrepancies in performance between quantitative and discretionary funds. Abis remarks that quantitative funds tend to hold more stocks, focus on stock selection, and participate in more crowded trades. On the other hand, discretionary funds invest in less widely known stocks, alternate be- tween picking and timing, and tend to outperform quantitative funds during economic downturns. According to the earlier studies mentioned, the comparison between funds managed by AI and those managed by humans has become a more pertinent topic in financial re- search, especially as AI-managed funds have expanded rapidly since their launch in 2017. AI's capability to analyze large volumes of data quickly and remove cognitive biases pro- vides it with a competitive advantage in predictive modeling and decision-making, which 46 fuels interest in its potential to surpass human managers. Nevertheless, there are ongo- ing concerns about AI's actual predictive capabilities compared to traditional models, its dependence on historical data, and how high-frequency trading might affect net returns. Recent research presents mixed findings on whether AI is superior. Hedge funds that are most automated generally perform better than those with significant human oversight, indicating that automation is a key factor in profitability. On the other hand, AI analysts excel over human analysts in stock selection that involves heavy data analysis, yet human insight remains crucial for assessing intangible assets and financial risks. Likewise, an analysis of systematic versus discretionary hedge funds indicates scattered evidence. While AI-driven quantitative funds gain from scalability and stock picking, human-man- aged funds have an edge in handling economic downturns. These insights emphasize both the advantages and limitations of AI and human fund management, suggesting that a combined approach may provide the most efficient strategy for optimizing perfor- mance. 3.5 Hypothesis Development As discussed before, herding behavior is a well-researched topic in the financial markets in varying contexts and conditions. However, as previous literature has proven, the evi- dence on herding behavior is inconclusive. There are multiple potential reasons for this. There could be several reasons for these fragmented findings, considering herd activity. According to Spyrou (2013), the models used to detect herding are allocated into two classes: models assuming rational agents and models assuming non-rational agents. There is yet one that considers both. Secondly, Spyrou states that the evidence of herd- ing behavior among institutional investors is scattered. Some studies find no evidence, and then some studies find consistent evidence. This stems from the reasons behind herding, whether it is spurious or intentional. Lastly, Spyrou mentions the limitations of empirical methods. The author questions the ability of economists to not completely comprehend the procedure of herding behavior. 47 Although research into herding behavior is increasing, a significant gap persists regarding its effects on fund performance, specifically when comparing human-managed to AI- managed funds. Most existing studies have concentrated on conventional fund manage- ment frameworks, leaving the exploration of herding tendencies in AI-driven funds, which should operate without behavioral biases, largely unaddressed. To fill this gap, this thesis leverages established herding detection methods to maintain empirical con- sistency and allow for comparison with previous research. By employing these method- ologies on both human- and AI-managed funds, this study aims to uncover fresh insights regarding the performance consequences of herding in fund management, especially during periods of market volatility. The first hypothesis is based on the studies by Jiang and Verardo (2018), Brown et al. (2014), and Chen and Ren (2022). Jiang and Verardo find that fund managers who herd tend to underperform their antiherding peers. This indicates that herd-driven decisions often result in suboptimal long-term returns. This supports the idea that human-man- aged funds, affected by biases like herding, may lag behind AI-managed funds, which are free from such influences. Furthermore, the research by Brown, Wei, and Wermers (2014) shows that fund managers often follow analyst recommendations, especially dur- ing downgrades. This type of herding behavior can lead to price fluctuations and ineffi- ciencies, suggesting it contributes to weaker long-term fund performance. Chen and Ren (2022) provide evidence that AI-managed funds outperform human-managed ones by efficiently processing large amounts of data and avoiding biases like herding. This indi- cates that AI funds may achieve better long-term performance. Therefore, according to these research findings, the first hypothesis is as follows: H1: Herding behavior negatively impacts the long-term performance of human-man- aged funds compared to AI-managed funds. 48 As noted, herding behavior tends to be most prevalent during turbulent market condi- tions. During financial crises and the COVID-19 pandemic, herding intensifies during pe- riods of high market volatility (Bekiros et al., 2017; Alexakis et al., 2023; Ferreruela and Mallor, 2021; Jiang et al., 2022). This indicates that market stress intensifies herding ac- tivity, resulting in inefficiencies in human-managed funds. Chang et al. (2000) and Chiang and Zheng (2010) also document herding in global markets, with evidence suggesting that herding is particularly strong in emerging markets and during downturns. Grobys et al. (2022) remark that hedge funds with the least human involvement tend to outper- form during periods of volatility. This implies that AI-managed funds might be more effi- cient in turbulent market conditions compared to human peers. Backed by these findings, the second and final hypothesis is as follows: H2: Herding behavior is more prevalent during times of significant market volatility, leading to a greater performance divergence between human-managed and AI-man- aged funds. 49 4 Methodology Next, an overview of the methodology used to produce this thesis' empirical results will be provided. The methodology consists of methods for measuring herding behavior and methods for calculating returns to compare the differences between human- and AI- managed funds. 4.1 Measures of Herding Behavior The Cross-Sectional Standard Deviation (CSSD) model by Christie and Huang (1995) can be considered the first relevant way to measure herding behavior across the market. Later, many varying models have been developed for different types of data and market conditions. To study market-wide herding, the authors Christie and Huang stress the sub- ject by utilizing the following equation: 𝐶𝑆𝑆𝐷! = %∑ #$!,#%$$,#& %& !'( '%( , (1) W