1 

Iina Wilenius 

Utilization of Artificial Intelligence in Investment 
Decisions Under Market Volatility 

Manager vs. Machine 

 

 

 

 

 

 

 
  

Vaasa 2023 

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:    Iina Wilenius 
Title of the Thesis:  Utilization of Artificial Intelligence in Investment Decisions Un-
der Market Volatility: Manager vs. Machine 
Degree:    Master of Science in Economics and Business administration 
Programme:   Master’s Degree Programme in Finance 
Supervisor:   Anupam Dutta  
Year:    2023 Pages: 87 

ABSTRACT: 
The adoption of artificial intelligence (AI) in the financial sector has consistently increased, driven 
by the AI boom that began in 2015. However, the amount of prior study of AI-powered instruments 
is quite limited, especially in volatile market conditions. Furthermore, the number of empirical stud-
ies comparing the performance of AI portfolio managers to their human counterparts is notably 
scarce. Thus, the study aims to fill this gap in the existing literature and determine whether AI out-
performs a traditional portfolio manager under volatile market conditions, considering the Ukrain-
ian conflict and Silicon Valley Bank (SVB) collapse. The study divided the sample data into two cat-
egories: AI-managed funds and human-managed funds. The event study method was selected as 
the research approach, with the aim of identifying possible abnormal returns during the events. 
Abnormal returns were calculated for ±20 days around the event date. Additionally, cumulative and 
holding period returns for eight separate observation periods were determined. All returns were 
risk-adjusted using the S&P500 index. Furthermore, a systematic literature review was conducted 
to examine previous empirical studies. The purpose was to answer a specific research question 
about psychological and other factors that influence the investment decision-making process and 
its outcomes between AI and human. 
 
The results indicate that neither management approach, AI nor a human portfolio manager, con-
sistently outperformed the other. AI funds exhibited higher abnormal returns during the Ukrainian 
conflict, while human-managed funds had higher abnormal returns during the SVB collapse. Con-
versely, AI funds demonstrated higher long-term performance after both events. Furthermore, the 
findings imply that the selection of better management approach in the investment decision-mak-
ing depends on specific circumstances. The results highlight that neither decision-making method—
AI or human-driven—is mutually exclusive; instead, they serve different purposes, each with dis-
tinct strengths and weaknesses. A hybrid approach, combining the strengths of both AI and human 
portfolio managers, could optimize performance across various investment situations, results state. 
 
Moreover, the study introduced two hypotheses: H0, the null hypothesis, assumed that the utiliza-
tion of artificial intelligence has no impact on investment performance in volatile market conditions. 
Alternatively, H1, the alternative hypothesis, posited that the utilization of artificial intelligence 
does impact investment performance in volatile market conditions. Based on the results, the null 
hypothesis, H0, was rejected and while the results suggested strong support to the alternative hy-
pothesis, H1. In addition, it should be noted that the sample size of this thesis could have been 
larger, but that would have been challenging due to the comprehensive research approach focusing 
on each individual fund. As a further research proposal, the number of samples and the observation 
period should be increased in order to create more significant results. Also, another proposal is to 
enhance this research by integrating AI algorithms into investment decisions at a practical level. 
However, the implementation can bring challenges due to companies' data encryption principles. 
 

KEYWORDS: Event Study, Abnormal Return, Performance, Artificial Intelligence, Portfolio 
Manager, Exchange Traded Fund, Investment Decision-making 



3 

Contents 
 
 

1 Introduction 6 

1.1 Purpose of the study 8 

1.2 Hypotheses 9 

1.3 Motivation 10 

1.4 Previous literature 11 

1.5 Structure of the study 13 

2 Artificial Intelligence 14 

2.1 Definition and Overview of Artificial Intelligence 14 

2.2 Historical Overview of AI 19 

2.3 AI in Finance 22 

3 Traditional Funds 24 

3.1 Overview of traditional funds 24 

3.2 Investment decision-making process of human 26 

3.2.1 The portfolio manager's decision-making framework and techniques 28 

3.2.2 Psychological factors and investment decision-making process 31 

4 AI-powered funds 36 

4.1 Overview of AI-powered funds 36 

4.2 Investment decision-making process of AI 38 

4.2.1 AI techniques and models for investment decision-making 42 

4.2.2 Absence of psychological factors 47 

5 Data and Methodology 50 

5.1 Data description 50 

5.2 Methodology 51 

5.2.1 Return calculations 52 

6 Results and analysis 56 

6.1 Returns and performance 56 

6.1.1 Ukrainian Conflict 56 

6.1.2 Silicon Valley Bank collapse 60 



4 

6.1.3 Summary of returns and performance results 64 

6.2 Decision-making process between human and AI 65 

7 Discussion and conclusions 69 

7.1 Interpretation of the results 69 

7.2 Limitations and future research 73 

7.3 Implications for policy and practice 74 

7.4 Further considerations 74 

References 76 

Appendices 87 

Appendix 1. Event study calculations 87 

  



5 

List of pictures, figures and tables  
 
 
Pictures 
 
Picture 1. AIEQ’s investment process (ETF Managers Trust, 2023b). 40 

 
 
Figures 
 
Figure 1. AI workflow. 15 

Figure 2. Example of machine learning prediction. 16 

Figure 3. Positioning of Artificial Intelligence, Machine Learning, Deep Learning, Neural 

Networks and Natural Language Processing. 18 

Figure 4. Dual Process Theory (Kahneman, 2011; Posner & Snyder; 1975). 27 

Figure 5. Example of decision tree algorithm. 44 

Figure 6. The timeline of the event study. 53 

Figure 7. AARs in the Ukrainian conflict event window. 59 

Figure 8. AARs in the SVB collapse event window. 63 

 
 
Tables 
 
Table 1. Examples of AI applications in financial market activities (OECD 2021b). 23 

Table 2. Descriptive Statistics. 50 

Table 3. AARs, CAARs and BHAARs in the Ukrainian conflict event window. 58 

Table 4. Comparison of CAARs and BHAARs during and after the Ukrainian conflict. 59 

Table 5. AARs, CAARs and BHAARs in the SVB collapse event window. 62 

Table 6. Comparison of CAARs and BHAARs during and after the SVB collapse. 63 

Table 7. Weaknesses in decision-making processes between human and AI. 66 

Table 8. Weaknesses in decision-making processes between human and AI. 67 

 
 



6 

1 Introduction 

Investing is not as rational as many of us believe. People have a variety of biases that can 

affect investment decisions. Human investors are likely to make decisions based on emo-

tions, biases, a lack of expertise, or patience, which can result in suboptimal investment 

returns. In a world where markets fluctuate and information is readily available, it is difficult 

to avoid panic sales and impulsive purchases. Considerable amount of finance literature is 

based on research which indicates biases in human judgment and decision-making to ex-

plain investor behaviour and market anomalies, such as studies by Daniel et al. (1998), De 

Bondt (1998) and Gärling et al. (2009) which states that investors do not act rationally what 

comes to investing. 

 

According to the findings of an U.S. survey (McNair, 2021), 66 percent of investors have 

made a spontaneous or emotionally charged investing decision that they later regretted. 32 

percent of investors have traded while drunk, survey states.  Furthermore, consumers who 

manage their own portfolios have a more difficult time keeping emotions out of investing 

than those who use a financial advisor, the findings reveal.  

 

Human limitations are one of the most serious issues in investing. Fortunately, there is a 

potential solution available. Artificial intelligence (AI) provides a systematical approach for 

addressing these issues. AI's primary function is to reduce exposure to psychological traits 

that can result in negative outcomes. AI can defeat many human limitations by analysing 

massive amounts of financial data, making objective investment decisions based on data-

driven insights, and constantly learning and adapting to changing market circumstances.  

 

New financial technology and innovations have transformed the finance industry by provid-

ing more value for investors' time and effort, as well as new ways for organizations to gen-

erate more returns. Technological advancements and the use of artificial intelligence are 

changing service delivery across multiple industries at the moment. Knowledge of how to 

apply AI and the ability to extract insights from available data will enable organizations in 



7 

financial services and other industries to improve their competitiveness, which in turn will 

help the organizations to increase their revenues and market share. 

 

Institutions in the financial industry face intense competition, which inevitably increases 

the operational and survival requirements of financial institutions. As a result, it is critical 

to understand about the possibilities of artificial intelligence, as it may be a determining 

factor in the company's performance in the future. Artificial intelligence has the potential 

to become a valuable tool in the investment decision-making process for institutional in-

vestors. In the future, the use of AI in the investment process may result in completely new 

and innovative ways of investing, thereby creating new opportunities and enhancing the 

financial sector's global development. 

 

The financial markets have seen major transformations as a result of significant advances 

in computing power, data science, and telecommunications. In this new environment of 

financial markets, computers have the power to gather and analyse massive amounts of 

data while executing trades in milliseconds. All this autonomously and without human in-

tervention. Even though, AI is still a relatively new phenomenon in human history and the 

use of artificial intelligence in investment decisions is even more recent. Pioneers in their 

field who have already begun to use artificial intelligence in financial decision-making have 

started to see positive results by now (OECD, 2021a). This technological revolution has re-

sulted in a fundamental reorganization of financial markets. Therefore, it is important to 

examine the possibilities of the utilization of AI in investment process.  

 

Artificial intelligence is considered as a key technology of the future. That is why it is critical 

to understand what AI is and how it can be applied to our needs. As artificial intelligence 

becomes more prevalent, it is vital to understand not only the benefits it provides, but also 

the disadvantages. When used correctly, artificial intelligence can provide new profit op-

portunities, ultimately leading to greater market share and a substantial advantage over 

competitors. 

 



8 

1.1 Purpose of the study 

The purpose of this study is to examine the profitability and effectiveness of using artificial 

intelligence in investment decisions, particularly under volatile market conditions. This 

study conducts research into the performance of AI-powered funds in comparison to tradi-

tional human-managed funds. In this study, the research problem of comparing the perfor-

mance of AI and humans is examined primarily by analysing the returns. Furthermore, the 

study extends to investigate the psychological and other elements that influence the out-

comes. In general, is it possible to increase profits by engaging AI into investment decisions?  

 

The research problem will be examined through the following research questions: 

 How does AI perform under volatile market conditions, and how profitable are AI-

powered funds during that time versus traditional funds managed by humans? 

 What are the psychological factors and other elements that influence the invest-

ment decision-making process of an AI versus human?  

 

The study examines AI-driven funds and traditional funds’ performance in volatile market 

conditions using the event study method, which quantifies the financial impact of particular 

events on asset prices. Through the use of the event study method, this study intends to 

examine the impact of the Ukrainian conflict and the collapse of Silicon Valley Bank on the 

performance of both AI-driven and human-managed ETF funds. 

 

First, the study focuses on the period of conflict in Ukraine, which began several months 

after the deployment of military bases near Ukraine's border (Zafra & McClure, 2023).  After 

that, Russia invaded Ukraine on February 24, 2022. At the time, the S&P 500 index fell more 

than 10 percent from its recent high and the US stocks closed at their lowest level since 

June 2021 (Horner & Wursthorn, 2022). The ongoing military action raised concerns, espe-

cially over the loss of lives and destruction of property, but also about its potential impact 

on the global economy and the reactions of global financial markets.  

 



9 

Moreover, the study examines the period of Silicon Valley Bank (SVB) collapse. The story 

began in 2020, when SVB changed its strategy and transformed itself as an investment com-

pany. In response to the impact of the Covid-19 pandemic on interest rate decreases, SVB 

channeled its interest-free deposits into securities, while cash flows from the booming tech 

industry soared for its core customer, technology companies (Ivantsov, 2023). However, 

SVB's securities investments together with the Federal Reserve's decision to raise interest 

rates to combat inflation, resulted to a dramatic decrease in the investments (Ivantsov, 

2023). Fears about SVB's solvency prompted a mass withdrawal of deposits, leading to SVB's 

collapse on March 10, 2023, marking the second-largest bank failure in US history after 

Washington Mutual during the 2007–2008 global financial crisis (Ivantsov, 2023). This crisis 

had far-reaching consequences, leading Swiss authorities to intervene in Credit Suisse just 

five days later. The collapse of SVB also significantly impacted Credit Suisse, a major global 

asset manager and systemically important bank, causing a rapid decline in its shares (Staff, 

2023 which raised concerns about potential spillover effects on the global financial system.  

 

To conclude, the aim of the study is to determine an answer to the research problem of 

whether AI outperforms a traditional portfolio manager under volatile market conditions, 

in the context of the Ukrainian conflict and the collapse of Silicon Valley Bank. In addition, 

the study attempts to identify the underlying causes for this potential outperformance in 

the decision-making process. Also, the goal is to provide a broad overview of artificial intel-

ligence, offering the reader a thorough understanding of what AI is and how it operates. 

 

1.2 Hypotheses 

The efficient-market hypothesis (EMH), assumes that asset prices reflect all available infor-

mation, suggesting an incapability to consistently generate excess returns in the markets 

(Fama 1970). The presumption of the study is that markets incorporate information effi-

ciently. The hypotheses focus on examining the influence of artificial intelligence on invest-

ment decisions, particularly in volatile market scenarios. A direct comparison is made be-

tween AI and human capabilities, specifically during volatile market conditions. Based on 



10 

EMH, the null hypothesis states that the use of artificial intelligence has no effect on invest-

ment outcomes in such conditions: 

 

𝐻଴: The utilization of artificial intelligence has no impact on investment performance in vol-

atile market conditions. 

 

Considering cumulative data, the study evaluates the overall impact of events during the 

event windows. While AI excels at real-time data prediction and analysis, human compe-

tencies involve complex reasoning based on nuanced contextual knowledge and reliance 

on intuition in uncertain situations. Despite these differences, it is reasonable to presume 

that the use of AI influences investment decisions, potentially yielding either negative or 

positive effects. Consequently, the alternative hypothesis is expressed: 

 

𝐻ଵ: The utilization of artificial intelligence does impact investment performance in volatile 

market conditions. 

 

The validity of the alternative hypothesis is contingent on the statistical significance of the 

impact observed in returns during events. This hypothesis aims to explore the potential 

influence of AI on investment performance under volatile market conditions. 

 

1.3 Motivation 

The urge for conducting this study arises from the growing significance and relevance of 

artificial intelligence (AI) as well as the prevailing AI boom. Furthermore, there is a critical 

need for further study of AI’s impact from a financial perspective.  

 

The main motivation of this study is that there is a significant gap in scientific research when 

it comes to comparing the performance of artificial intelligence and human managers, es-

pecially under volatile market conditions. There are just few studies about comparing AI 

and human portfolio managers in general, but none regarding volatile market environment, 

to my knowledge.  



11 

Although, the motivation for research arises not only from the limited amount of previous 

literature, but also from the factors of the current global situation, such as the growing 

popularity of artificial intelligence. When OpenAI was released in 2015, the AI boom started 

to form, and its popularity has grown at an exponential rate ever since. At the moment, 

artificial intelligence is conquering the world and establishing its position globally at ex-

tremely rapid speed. It has entered every aspect of the people’s life and is expanding so 

quickly that no one can forecast its limits.  Based on a report by Grand View Research (2023), 

AI's current market size of 200 billion dollars is predicted to increase to almost two trillion 

dollars by 2030, demonstrating the industry's rising tendency. 

 

In addition, despite the fact that artificial intelligence is a relevant topic with almost limit-

less possibilities, the potential of AIs performance in a quantitative form that describes its 

financial capability versus humans, has received little attention in the field of scientific re-

search as well as the aspect of AI in financial decision making. 

 

1.4 Previous literature  

The existing literature and articles in this field offer limited insights into the distinctions in 

investment decisions when comparing the financial performance of AI and humans. While 

many studies in the financial literature evaluate the performance of hedge fund managers 

in particular, just a few compare AI driven exchange traded funds (ETFs) against those man-

aged by humans. For example, Harvey et al. (2017) investigated the performance differ-

ences between discretionary and systematic hedge funds (no daily intervention by humans) 

and discovered similar results in performance of both categories. In contrast, Grobys (2022) 

and Niang (2021) found that hedge funds with higher levels of automation outperform 

those with a higher level of human engagement. Furthermore, Chen et al. (2022) studied 

AI-powered mutual funds, revealing that they outperform funds managed by humans. Nev-

ertheless, none of these studies compared the performance of AI to human managers un-

der volatile market conditions or examined ETFs. 

 



12 

A closer look at Eugene Fama's (1970) Efficient Market Hypothesis (EMH) theory asserts 

that in efficient markets, prices include all available information. This fundamental theory 

is applied to quantify how certain events affect the value on securities. The financial impact 

of an event can be measured through event studies, a widely used statistical method in 

economics, and finance (Binder, 1998). The method was first introduced in the 1960s by 

Ball & Brown (1968) and Fama, Fisher, Jensen & Roll (1969), establishing the methodology 

in its current form. It has since been applied in various empirical studies to examine the 

influence of an important occurrence or contingent event on the value of a security. 

 

The effect of specific events, such as war, geopolitical shocks, or bank failures in the securi-

ties market, have previously been studied in the empirical literature. The selection of these 

study’s events was based on the assumption that increased market volatility often arises in 

response to certain occurrences such as geopolitical conflicts and banking crises. Like study 

by Bredin & Fountas (2018) asserts, conflicts, war, and banking crisis tend to increase eco-

nomic uncertainty. Furthermore, prior literature supports the assumption that that higher 

economic uncertainty is positively related to increased market volatility (Antonakakis et al., 

2013; Bansal et al., 2014; Tong et al., 2023;). In addition, Gray and Kucher (2000), investi-

gated the effects of World War II on government bond prices, whereas studies on the 2008 

global financial crisis examined the effects of financial crises and banking shocks on global 

financial markets, see, e.g., Grammatikos & Vermeulen (2012) and Bénétrix et al. (2015). 

This study contributes to the existing literature by examining the potential impacts of in-

creased market volatility particularly on AI-based securities focusing on recent events, such 

as the collapse of Silicon Valley Bank and the conflict in Ukraine. However, while previous 

research has already examined at the effects of the Ukrainian conflict on the stock market 

(Boungou & Yatié, 2022; Izzeldin et al., 2023) and the effects of the SVB collapse (Martins, 

2023; Pandey et al., 2023), none of these studies have focused on AI-based instruments, 

underlining the focus of this research.  

 

In addition, the field of behavioural finance which examines financial psychology to analyse 

investors' actions, has been widely studied in the financial literature, and it owes much to 

the concept's primary founders, Amos Tversky, Daniel Kahneman, and Richard Thaler. 



13 

Despite being a widely researched field, it has very limited or hardly any research on the 

psychological differences between artificial intelligence and humans in the context of in-

vesting decisions. Most research on AI financial decision-making focuses entirely on the AI 

perspective, often disregarding direct comparisons with human decision-making. One of 

the early studies on the subject by Pomerol (1997) compared two aspects of AI and humans 

in decision-making: diagnosis and look-ahead, indicating AI's competence in diagnosis but 

a lack of attention on look-ahead thinking. Ren's study in 2021 highlighted the strategic 

relevance of implementing AI technology in finance, whereas Chen and Ren's study in 2022 

presented an overview of AI and human behavioural elements with no specific focus on 

investment decision-making or in-depth analysis based on prior literature. In contrast, this 

study provides a broad overview for examining the psychological factors that influence in-

vestment decisions between AI and humans through prior literature. 

 

To my knowledge, no empirical studies have assessed the impact of bank failures, conflicts, 

or volatile market conditions on general AI-driven investment performance. Additionally, 

there is a lack of empirical research comparing AI and human portfolio managers in volatile 

market conditions. Despite extensive study in behavioural finance, a notable gap exists in 

understanding psychological distinctions between AI and human managers in the context 

of investment decisions. This research aims to address these gaps in the current literature. 

 

1.5 Structure of the study 

The paper comprises seven chapters, each serving a distinct purpose. Chapter one intro-

duces the study, outlining objectives, motivation, and hypotheses. Chapter two presents a 

comprehensive overview of artificial intelligence, including historical development. Chapter 

three observes traditional human-managed funds, focusing on investment decision-making 

and psychological factors. In contrast, chapter four delves into AI-powered funds, examining 

their functioning. Chapter five introduces the empirical part, presenting data and research 

methodology. Chapter six covers key findings and analysis of AI-powered and traditional 

funds. Finally, chapter seven presents’ discussions, conclusions, and analyses the results' 

importance, reliability, and validity while addressing the study's contribution. 



14 

2 Artificial Intelligence  

This chapter provides an overview of artificial intelligence and its history. Furthermore, how 

artificial intelligence is applied in the financial industry is discussed. The purpose is to pro-

vide a broad overview of artificial intelligence, and its applications in the financial sector. 

 

2.1 Definition and Overview of Artificial Intelligence 

The ways AI and humans use to operate are entirely different, yet the goal is the same. 

Human-like intelligence is an empirical science related to psychology that includes experi-

ments, human behaviour, and mental processes, whereas AI's rational approach incorpo-

rates a combination of mathematics, engineering, and statistics.  

 

Artificial intelligence is a part of computer science that attempts to develop machines or 

computer systems capable of doing activities that would normally need human intelligence. 

The fundamental objective of AI is to create machines that can think, learn, reason, and 

adapt in the same way that humans do.  

 

According to Russell and Norvig (2022), AI systems can be basically assembled into two cat-

egories: Narrow AI systems designed for narrow tasks, such as recommendation systems 

and online personal assistants, and Artificial General Intelligence, also known as General AI 

(AGI), which refers to systems with human-like cognitive capabilities and can perform well 

in a number of areas. AGI systems can understand, learn, and gain knowledge in a way that 

is similar to human intellect, they continue. In contrast from Narrow AI, AGI is not limited 

to certain tasks and may adapt and succeed in several kind of tasks (Russell & Norvig, 2022). 

Moreover, aside from Narrow AI and AGI, there is a third hypothetical AI system category 

known as Artificial Super Intelligence (ASI). As stated by Russell & Norvig (2022) ASI is a 

theoretical artificial intelligence system that outperforms even the most talented humans 

in terms of cognitive capacity. They explain that because of the significant ethical and safety 

considerations surrounding its relationship with humanity, ASI is still purely speculative.  

 



15 

Artificial intelligence uses algorithms that allow it to learn, analyse data and make informed 

decisions. Lowe & Lawless (2021) assert that an algorithm is a procedure or set of rules that 

a computer must follow in calculations or other problem-solving operations. Basically, an 

algorithm is a set of instructions designed to accomplish a given task, they define. In prac-

tice, it might be described using a simple method like multiplication equations: computers 

only execute it in binary form, while humans use decimals (Lowe & Lawless, 2021). Figure 

1. below depicts the general AI operation process by observing the AI process in action step 

by step. 

 

 
Figure 1. AI workflow. 
 

This study concentrates on the most relevant AI algorithms within the subject of the study. 

The first technique observed is machine learning (ML), which is certainly among the most 

important AI techniques. Machine learning is a core AI technology and a subgroup of AI. As 

Ahmed et al. (2022) states, ML is an artificial intelligence technology that allows systems to 

learn patterns from data without being explicitly programmed for it. ML gains knowledge 

from experiences or data sets rather than from instructions alone. Ahmed et al. (2022) con-

tinues that, in general, machine learning focuses on explicitly identifying a problem that a 

computer is able to solve. The problem must be described mathematically, in a form that 

can be solved by an algorithm. ML models are frequently composed of a set of rules, pro-

cedures, or sophisticated "transfer functions" that can be applied to identify intriguing data 

1. Data management
- Collect  & explore data

- Cleanse data
- Prepare data 
- Format data 

2. Model Training
- Select learning task
- Engineer features

- Select algorithms and 
models

3. Feature Extraction
- Identificate the most 

essential features of the 
data that contribute to the 

current assignment.

4. Optimization
- Adjust the internal

parameters to minimize the
errors and maximize the

performance on the
training data

5. Model Evaluation
- Test the model with fresh 
test data set to assess the  

generalization ability
- Performance measure 
- Compare alternatives

6. Deployment
- Integration of the AI 

model
- Monitor outcomes

- Improve Model



16 

patterns or predict behaviour. ML utilizes data to make predictions about uncertain events 

in the future, and incorporates methods from statistics, neural networks, operations re-

search, and physics. It employs these methods to uncover hidden patterns in data without 

being specifically programmed for what to observe or discover (Ahmed et al., 2022) 

 

Machine learning employs algorithms to create models. ML algorithm is a process that is 

executed on data to produce a ML model, so the ML Model is the result of a machine learn-

ing algorithm applied to the data (Ahmed et al.,2022). In addition, as asserted by Russell & 

Norvig (2022) machine learning means that it must learn to predict, classify, or find patterns 

based on certain data. In order for the machine to learn these skills, it has three different 

learning styles at its disposal. The three types of ML are supervised, unsupervised and re-

inforcement learning (Russell & Norvig, 2022). 

 

Next, Russell & Norvig (2022) talk about different AI learning methods. Supervised learning 

is the process by which an artificial intelligence system is trained on pairs of input and out-

put data to generate predictions or classifications. Unsupervised Learning, on the other 

hand, is the process through which an AI system identifies patterns, clusters, or correlations 

in data, without any explicit guidance, they explain. Moreover, Reinforcement Learning 

means as learning through interaction with an environment. In this method the AI agent 

receives feedback in the form of incentives or penalties in order to optimize its actions (Rus-

sell & Norvig, 2022). Figure 2. shows the learning process in its simplest form. 

 

 
Figure 2. Example of machine learning prediction. 
 



17 

Deep Learning (DL) is one of the most essential AI techniques and a special subset of AI and 

ML, that teaches machines to make intelligent decisions on their own. According to Ahmed 

et al. (2022) and Lowe & Lawless (2021), deep learning is a class of algorithms, and it in-

volves a higher level of automation than typical ML models. It is basically a three- or more-

layered neural network. They state that, these neural networks seek to imitate human brain 

behaviour by learning from substantial amounts of data. Whereas a single-layer neural net-

work is capable of producing approximations, additional hidden layers of deep learning can 

assist in accuracy optimization and improvement, they summarise. 

 

Several AI products and services employ deep learning in order to improve automation and 

conduct analytic and physical activities without human involvement. Deep learning tech-

nology is at the core of everyday products and services like voice-controlled electronic de-

vices and credit card fraud detection, as well as emerging technologies like self-driving ve-

hicles (Alzubaidi et al., 2021; Russell & Norvig, 2022). 

 

As explained by Ahmed et al. (2022), a neural network (NN) is a ML model and a subgroup 

of machine learning which serves as the foundation for deep learning algorithms. NN is a 

series of algorithms that uses interconnected neurons in a layered structure to communi-

cate between each other and process information in response to external inputs, they add. 

Basically, it is a set of different techniques or algorithms that determine the relationship 

between several underlying factors and process the data in a similar way to the human brain, 

they explain. NN is ultimately a mathematical version based on the biological brain. 

 

Furthermore, Natural Language Processing (NLP) is an element of AI that stands for the 

ability of a system or machine to learn, perceive, and understand human language as it is 

delivered (Ahmed et al.,2022). The majority of NLP methods use ML and DL-based tech-

niques to obtain insights from human language and NLP enables machines to understand 

human speech (Ahmed et al.,2022). For example, speech recognition is one of the things 

that can be implemented with the help of natural language processing. According to Russell 

& Norvig (2022), NLP can detect fake news, spam, as well as provide responses. It is also 



18 

used in applications like as language translation and chatbots (Russell & Norvig, 2022). Fig-

ure 3. visualizes how the AI and its subsets are positioned in relation to each other. 

 

 
Figure 3. Positioning of Artificial Intelligence, Machine Learning, Deep Learning, Neural Networks 
and Natural Language Processing.   
 

As Russell & Norvig (2022) explain, AI needs agents to achieve its goals in the best possible 

way. They define that, an artificial intelligence agent can be viewed using an example, in 

which the AI agent perceives and acts in the environment. The function of an agent specifies 

the action that the agent takes in response to any sequence of observations, they add. They 

also clarify that agent's goal is to observe its environment and act in accordance with it, in 

order to achieve its goals. The main purpose is to act as optimally as possible in relation to 

the environment in order to obtain the best result. An agent's adequate shape and structure 

are determined by its surrounding environment (Russell & Norvig (2022). 

 

ML can be utilized to empower an AI agent or product to learn from experience, such as 

how to complete a task and achieve a goal. Given the information offered by the environ-

ment and knowledge built into the agent, a rational agent should take an action that is 

predicted to maximize its performance (Lowe and Lawless, 2021; Russell & Norvig, 2022). 

According to Russel and Norvig (2022) the performance measure assesses the agent's be-

haviour in a given environment. They continue that the aim of a rational agent is to maxim-

ize the expected value of a performance measure, and there is a high risk that the agent 

will optimize the wrong objective if the performance measurement cannot be performed 

or is difficult to define. 



19 

Additionally, there are also different types of agents which are roughly classified into differ-

ent categories: first, model-based reflex agents attempt to track features of the world that 

are not visible in existing observations, whereas simple reflex agents respond directly to 

observations, as noted by Russell & Norvig (2022). Then again, goal-based agents looking 

for to achieve a specific goal, whereas utility-based agents seek to maximize their own “sat-

isfaction”, they conclude. 

 

2.2 Historical Overview of AI 

It all began in 1943, when Warren McCulloch and Walter Pitts developed a computational 

model for neural networks (NNs), which served as the cornerstone for foundations of arti-

ficial intelligence (Lowe & Lawless, 2021). Then again, in 1950, Alan Turing published an 

article “Computer Machinery and Intelligence” in the academic journal Mind, in which he 

debated how to create intelligent machines and introduced a method how to test their in-

telligence. The procedure came to be known as the "Turing Test," and it was a form of ex-

perimentation to see if a computer could exceed a human (Cowell, 2019). In the experiment 

he utilized a human interrogator as part of his strategy to ask questions in order to find out 

whether the responses came from a computer or a human (Cowell, 2019). Turing's research 

was the most influential and significant contribution to artificial intelligence at the time, 

and the name "artificial intelligence" was adopted by the scientific community six years 

later (Lowe & Lawless, 2021).  

 

Today, Alan Turing has been recognized as one of the great fathers of artificial intelligence 

and one of the prominent codebreakers of World War II, whose cryptology provided infor-

mation that was considered to speed up the Allied victory, as asserted by Cowell (2019). He 

embraced the first visions of modern computing, and his intellect created revolutionary in-

sights into what is today referred to as artificial intelligence (Cowell, 2019).  

 

1951, Marvin Minsky envisioned a neural network simulator machine that evolved into sto-

chastic neural analog amplification calculator, also known as SNARC and it is considered as 

the first neural network machine (Russell & Norvig, 2022). Later, in the summer in 1956, 



20 

John McCarthy and Marvin Minsky hosted the first academic conference on artificial intel-

ligence, “Dartmouth Summer Research Project on Artificial Intelligence” in New Hampshire, 

which had a significant impact of AI research over the next twenty years (Lowe & Lawless, 

2021; Russell & Norvig, 2022). Artificial intelligence thrived from 1950s to early 1970s as 

computers could hold more data and got faster, cheaper, and more accessible (Rockwell, 

2017). Machine learning algorithms developed as well, and people knew which algorithm 

to use for which task. Early experiments, such as General Problem Solver by Newell and 

Simon and Joseph Weizenbaum's ELIZA, one of the first chatterbots, were promising in the 

areas of problem solving and spoken language interpretation (Rockwell, 2017). 

 

In 1973, as stated by Lowe & Lawless (2021) most funding resources were rejected from AI 

research, owing mostly to Sir James Lighthill's report on the state of artificial intelligence, 

where he underlined that traditional AI problems could be solved by other sciences and 

general artificial intelligence is impossible to accomplish. As a result, funding for AI research 

dropped and the years that followed were known as the "AI Winter," which lasted from 

1974 to 1980, they continue. Gladly, it didn’t last forever. After the "AI Winter" in 1986, 

David Rumelhart and James McClelland published a book in which they developed ideas 

about parallel distributed processing and neural network models and created computer 

simulations to giving computer scientists the first testable neural processing models (Lowe 

& Lawless, 2021). 

 

The 1980s similarly witnessed the rise of robots, with numerous scientists arguing that for 

AI to be useful, it must have a body, which resulted in the development of sensor-motor 

abilities (Lowe & Lawless, 2021). Likewise, Edward Feigenbaum pioneered expert systems 

that replicate the decision-making process of a human expert, and the program was de-

signed to ask an expert in the subject how to act in a certain situation, and once the ma-

chine learnt the answers, non-experts could obtain guidance from that program (Lowe & 

Lawless, 2021; Rockwell, 2017). Afterwards, in the 1990s, a new paradigm known as "intel-

ligent agents" gained widespread acceptance in the AI field (Rockwell, 2017). In addition, 

1997 was a significant year for the development of AI when IBM's Deep Blue computer 



21 

game system defeated human for the first time. The game was played between an IBM 

supercomputer and the reigning world chess champion Garry Kasparov (Rockwell, 2017). 

 

The 2010s also saw a lot of development. In 2002, iRobot released the Roomba robot vac-

uum cleaner, which could navigate independently and also avoid obstacles, according to 

Lowe & Lawless (2021). In 2004, the United States Department of Defence Research Organ-

ization (DARPA) launched a challenge to develop vehicles capable of traveling more than 

150 miles autonomously, they continue. Similarly, DARPA announced the Urban Challenge 

for Autonomous Vehicles initiative in 2007, and after that, Google constructed its first self-

driving car, making its entry into the self-driving autonomous car market in 2009 (Lowe & 

Lawless (2021). 

 

Between 2011 and 2014, Smartphone apps Siri, Google Now, and Microsoft Cortana were 

released, and they used natural language to answer questions, make recommendations, 

and conduct activities, as stated by Lowe & Lawless (2021). Moreover, SCHAFT Inc., a 

Google company, produced the HRP-2 robot in 2014, which was capable of driving a vehicle, 

walking over debris, climbing a ladder, removing debris, walking through doors, cutting 

through a wall, closing valves, and attaching a hose, they explain. Moreover, finally, in 2014, 

the Turing Test was passed when chatbot Eugene Goostman - a fake 13-year-old boy from 

Odessa, Ukraine, who didn't speak English fluently - tricked 33 percent of the jury into think-

ing he was a real child during a five-minute interview (Pulakkat, 2014).  

 

Later, in 2016, Google's DeepMind AlphaGo supercomputer won world champion Lee Se-

dol in one of the world's most complex strategy games, and the AI learned the game in 30 

hours using unsupervised learning (Lowe & Lawless, 2021). Similarly, AlphaGo Zero won the 

world's greatest chess computer program in 2017, and the AI taught itself to play chess in 

less than four hours (Lowe & Lawless, 2021). 

 

In 2015, OpenAI, an artificial intelligence lab, was founded with the purpose of developing 

"artificial intelligence," or AGI, or software that is as intelligent as humans (Verma, 2023). 

After couple years, in 2020, OpenAI presented GPT-3 "the API" which was meant for 



22 

answering inquiries in natural language but can also translate between languages and pro-

cess writing (Verma, 2023). Then, in 2021, OpenAI introduced DALL-E, a deep learning 

model that could create graphics based on human instructions (Verma, 2023). Year later, 

OpenAI launched a free preview of ChatGPT, the latest AI chatbot built on GPT-3.5 (Kay, 

2023). Ultimately, in 2023, OpenAI announced the fourth version of its multimodal language 

model, GPT-4, which is delivering safer and more reliable responses as well as solving chal-

lenging problems with improved accuracy (Kay, 2023).  

 

The development of artificial intelligence has come a long way from the 1950s to today. AI 

development has accelerated in recent decades, owing to enhanced computers and in-

creased computer capacity.  Currently, the AI market includes a wide range of sectors, in-

cluding technology, healthcare, finance, retail, manufacturing, transportation, marketing, 

education, agriculture, energy, entertainment, government, and so on. These sectors have 

been implementing and integrating AI into their operations for applications ranging from 

medical imaging and fraud detection to personalized learning experiences and self-driving 

automobiles. The adaptability of AI is driving continuous growth and innovative benefits, 

enhancing decision-making and productivity across many industries.  

 

2.3 AI in Finance  

One of the few heavily automated businesses is the financial sector. According to Bartoletti 

et al. (2020), computerized systems have automated all elements of financial service func-

tions, resulting in massive volumes of data. AI has the potential to be widely used in finance, 

given the business is essentially focused on formulas, statistics, and strategies, they note. 

On a broad scale, AI is able to improve nearly all areas of the financial sector.  

 

According to Bartoletti et al. (2020) AI has a significant impact in the financial sector. On 

practical level, AI can be utilized for example in lending, making payments and deposits, 

insurance, investments, and wealth management and it is already being used for credit 

scoring by banks and financial organisations, they explain. AI is a significant tool especially 

in the finance industry's customer service operations, and for example, AI can generate 



23 

consumer enquiries and provide assistance, such as a chatbot or answering machine for a 

bank (Bartoletti et al., 2020). AI can also perform credit scoring: whereas the traditional 

method relies on static variables and historical data, artificial intelligence-based credit scor-

ing evaluates a wide range of data points using machine learning algorithms (Bartoletti et 

al., 2020). Furthermore, artificial intelligence is employed as a tool in the identification of 

fraud by financial organisations, because it is capable of recognizing unusual patterns in 

activities such as credit card transactions (Hilpisch, 2020). 

 

Likewise, portfolio managers can benefit considerably from AI techniques in portfolio man-

agement. AI can assist portfolio managers with data analysis, risk assessment, asset alloca-

tion and performance tracking, and among other things, it can even compose and rebalance 

portfolios (Hilpisch 2020). Furthermore, intelligent technology can assist with investing op-

erations such as trade execution: for example, AI is capable of learn how to perform large-

block deals while minimizing transaction costs, he continues. AI can even perform deriva-

tives hedging, where AI is taught to optimally execute hedge transactions against specific 

derivative instruments or portfolios (Hilpisch 2020). 

 

Table 1. Examples of AI applications in financial market activities (OECD 2021b).  

 
 

Table 1. above presents real life examples of different AI applications in financial activities. 

There are several ways artificial intelligence can be employed in different functions of the 

financial sector, and the integration of artificial intelligence into financial operations has 

already spread to various financial functions. 

As
se

t M
an

ag
em

en
t 

Al
go

rit
hm

 tr
ad

in
g 

Cr
ed

it 
in

te
rm

ed
ia

tio
n 

Bl
oc

kc
ha

in
-b

as
ed

 fi
na

nc
e 

Back Office Middle Office Front Office 

Post-trade processing Risk management Asset allocation 
Trading P&L, reconciliations KYC checks Robo-advisors, Chatbots 
Reporting and record man-
agement 

Compliance Biometric authentica-
tion 

Data analytics Control functions 
/ processes 

Trade execution 

Credit scoring / risk under-
writing 

AML / CFT Personalised recom-
mendations 

IT / infrastructure Anti-fraud Customer service 



24 

3 Traditional Funds 

3.1 Overview of traditional funds 

An investment fund i.e., mutual fund, in general is an instrument that invests in various 

types of assets and is owned by investors who have purchased fund units, which are in fact 

shares of the fund. The fund’s assets may consist of cash, stocks, loans, tangible or intangi-

ble assets, and most funds have been established to hold multiple types of assets. Basically, 

a fund means any combination of assets and broadly, almost any economic collection or 

pooling can be considered as a fund (Hudson, 2014; Kallunki et al., 2019). In general, the 

fund's revenue is based on the income generated by its investments, which include interest 

income, dividends and the increase or decrease in the value of the invested capital (Kallunki 

et al., 2019).  

 

The assets of the fund are owned by the persons, entities and foundations who invested in 

it, in proportion to the size of their investment. One of the main features of a traditional 

fund is the presence of a professional fund manager who manages and advises the fund. 

Traditional funds select stocks and make investment decisions mainly through human judg-

ment. Portfolio managers are responsible for managing funds’ assets and deciding how to 

invest them. As asserted by Kallunki et al. (2019) and Hudson (2014) the fund may invest in 

multiple types of instruments in accordance with its approved investment policy. The capi-

tal of the fund varies according to the fluctuation in the value of its investments and how 

investors buy and sell fund shares, they continue. This affects the calculation of the value 

of the shares, as the value of each individual share is calculated based on the fund's worth 

(Hudson, 2014; Kallunki et al., 2019). 

 

The price of a fund unit is always the same for both the new investor and the existing inves-

tor who sells the share, i.e., the value of the unit is the fair market value of the fund's in-

vestments divided by the number of fund shares currently in circulation (Kallunki et al., 

2019). When fund units are subscribed or redeemed, the fund's capital increases or de-

creases, Kallunki et al. (2019) explains. Fund units are not typically traded on the stock 



25 

exchange but exchange traded funds (ETF’s) are publicly listed funds whose fund units are 

traded on the stock exchange like a stock, and their price is determined according to supply 

and demand (Kallunki et al., 2019). 

 

Kallunki et al. (2019) continues to explain about different types of funds. They state that 

there are many different types of funds with different characteristics. Short-term invest-

ments funds, for example, invest in short-term money market securities such as govern-

ment bonds and corporate bonds. The Long-term investment funds, on the other hand, 

invest primarily in bonds and other interest instruments with loan terms of more than one 

year. There are also mixed funds, which invest in both bonds and stocks, they add. The 

weightings of bonds and stocks may vary between mixed funds. In contrast, equity funds 

invest principally in stocks, Kallunki et al. (2019) asserts. 

 

In addition, as declared earlier, ETF investment funds are publicly listed funds that trade 

similarly to stocks. ETFs' investment policies are stated in the fund's policy, and investment 

policies between ETFs might differ significantly, Kallunki et al. (2019) says. Most ETF funds 

are index funds, and their return is equal to the target index's performance minus the fund's 

expenses. In its research, this paper focuses on exchange-traded funds. 

 

Following, this thesis observes at the funds that serve as benchmarks for traditional funds 

for this paper. The funds are carefully chosen based on their characteristics. 

 

The first traditional fund utilized in this research is Dimensional U.S. Core Equity 2 ETF. The 

following paragraph is based on data sourced from the fact sheet and fund prospectus pro-

vided by Dimensional Fund Advisors LP (2023). The fund is actively managed exchange-

traded fund and trades under the ticker symbol DFAC on NYSE Arca. As confirmed by Di-

mensional Fund Advisors LP (2023.), the ETF seeks long-term capital appreciation while ad-

dressing the federal income tax implications of investing decisions. The DFAC is intended to 

invest in a wide and diverse range of securities of U.S. companies. Moreover, the portfolio 

invests in companies of different sizes, with a focus on companies with lower capitalization, 

lower relative price, and higher profitability than their competitors in the US Market. The 



26 

fund Advisor, in the function of portfolio manager, makes investment decisions with specific 

characteristics when necessary. The ETF does not attempt to replicate the performance of 

a particular index and under normal conditions, DFAC will invest at least 80 percent of its 

net assets in securities of U.S. companies.  

 

Another fund considered in this study is the Avantis U.S. Equity ETF, which trades under the 

ticker AVUS on Nyse Arca. The information in this paragraph is drawn from the fact sheet 

and fund prospectus supplied by American Century Proprietary Holdings (2023). The ETF is 

actively managed and incorporated in the United States. The fund does not seek to track 

the performance of a specific index. The main objective of the ETF is to seek long-term 

capital appreciation by primarily investing in a wide selection of US companies of all market 

capitalizations, sectors, and industries. The fund places a strong emphasis on small-cap 

companies that are expected to have higher returns, better profitability, and attractive 

value characteristics. In contrast, AVUS tends to reduce or exclude investments mainly in 

larger companies that are expected to have lower returns, profitability, and less attractive 

value characteristics. The fund's advisor is American Century Investment Management. 

Portfolio managers constantly analyze market and financial data to make buy and sell deci-

sions with desired characteristics. In addition, they regularly evaluate the portfolio inclusion 

criteria. Under typical market conditions, the fund invests at least 80 percent of its assets 

in equity securities issued by US companies. 

 

3.2 Investment decision-making process of human 

Humans make nearly 35 000 decisions every day (Krockow, 2018). Decisions can range from 

small to huge, like what to wear today, what to eat for lunch, whether to buy a house from 

city or countryside, and so on. Decisions can be performed in many ways, with fast or in-

depth consideration (Krockow, 2018). These choices can be minor or life changing. 

 

Human decision-making is a complex and multidimensional phenomenon that has been 

researched from various fields such as psychology, neuroscience, economics, and behav-

ioural science. While there is no single model that can generally describe every human-



27 

made decision, the Kahneman’s ground-breaking theory (2011), based on the dual-process 

model by Posner & Snyder (1975) is a commonly accepted paradigm for understanding hu-

man decision-making. The Kahneman’s model divides the decision-making process into two 

segments: System 1: fast, intuitive thinking, which happens automatically and with minimal 

effort (instincts, habits, past experiences) and System 2: slower analytical thinking which is 

conscious and logical (reflection, planning, and problem solving). The dual-process model 

theory works as a benchmark theory for behavioural and brain sciences that can be refor-

mulated and adapted to fit to almost any human behavioural context. 

 

 
Figure 4. Dual Process Theory (Kahneman, 2011; Posner & Snyder; 1975). 
 

In practice, Kahneman's theory (2011) could be described in the following way. Imagine 

you're a coffee enthusiast, and every morning you instinctively head to your favourite local 

coffee shop without much consideration. You know the barista by name and your normal 

order is nearly a natural instinct. It's a routine that requires minimal cognitive effort because 

it's become rooted in your daily life. However, one day you arrive at the coffee shop only to 

discover that it is closed for renovation. You're now facing an unexpected decision-making 

process and must consider your options: Should you try a new coffee shop close by? Is it 

worth driving a little further for your favourite brew? Maybe you could make your own 

coffee at home, but that would take more time. Consequently, in this scenario, the habitual 

decision (system 1) to go to your regular coffee shop is disrupted, prompting you to consider 

and evaluate alternative options (system 2). 

 

In addition, Engel and Singer (2008) give explanation about advanced human abilities in the 

context of decision-making. Moral, ethical, and emotional understanding contribute to hu-

man decision-making, laying the groundwork for nuanced decisions, they state. These abil-

ities with creativity, enable humans to generate innovative solutions to problems with 

System 1 

Fast/ Automatic
Impulses

Habits
Beliefs

System 2

Slow / Effortfull
Reflections

Planning
Problem Solving



28 

efficiency, they add. Moreover, the basic human abilities like, self-awareness, intuition, and 

instinct, was well as the ability to read and interpret situations, contribute to the under-

standing of context required for making nuanced decisions (Engel & Singer, 2008). Further-

more, human decision-making extends to considering social and cultural factors, and the 

ability to consider the impact of decisions on others, based on Engel and Singer (2008). 

 

3.2.1 The portfolio manager's decision-making framework and techniques 

The role of a portfolio manager requires constant decision-making on many different areas. 

Portfolio management requires an in-depth knowledge of the strengths and weaknesses, 

opportunities, and risks of various investment instruments. A portfolio manager oversees 

all of the trades executed during the day by the investment fund or portfolio by making final 

decisions on the securities involved.  

 

Moreover, portfolio management can be either passive or active. Passive management in-

volves replicating the performance of a specific market index without active trading, while 

active management requires constant decision-making by the portfolio manager who ac-

tively buys and sells securities to often outperform a specific index, such as the S&P 500 

(Bodie et al., 2023).  This study focuses on active portfolio management approach. 

 

The success of the fund depends on the fund manager's ability to make the right decisions, 

based on thorough market research, market forecasts and the manager's expertise. Portfo-

lio managers are constantly tracking market trends, economic events, political movements, 

and corporate news, which guide the timing of trade execution decisions to profit from 

anomalies (Bodie et al., 2023). Furthermore, decision-making is required from several areas 

such as asset allocation and diversification, which are also critical elements of portfolio 

management.  

 

First, asset allocation is a critical concept that involves making decisions regarding the com-

bination of assets within a portfolio. Different asset classifications include stocks, bonds, 

cash equivalents, and derivatives (Kallunki et al., 2019). Recognizing that different asset 

classes exhibit varying degrees of volatility; managers strategically balance their portfolios 



29 

based on their risk tolerance and financial targets (Kallunki et al., 2019). The asset allocation 

decision entails selecting from these general asset classes, whereas the security selection 

decision involves selecting which specific stocks to hold within each asset class (Bodie, et 

al., 2023).  Second, diversification is a fundamental concept of portfolio management, 

which also demands constant decision-making by managers. Diversification is based on 

Modern portfolio theory (Markowitz, 1952) according to which owning a portfolio of differ-

ent asset classes is less risky than owning a portfolio of similar assets. Portfolio managers 

seek to capture the total returns of various sectors over time while reducing volatility by 

creating a diversified portfolio of investments stretching across asset classes, industries, 

and geographical areas (Bodie et al., 2023).  

 

Third, rebalancing involves regularly checking and adjusting the investment allocation of 

the portfolio to its initial target allocation, which helps to lower the risk and volatility in the 

portfolio, often leading to improved returns, as stated by Malkiel & Ellis (2020). The re-

balancing process allows the manager to capture profits and enhance growth opportunities 

in sectors with favourable outlooks, all while staying within the original risk-return profile 

of the portfolio, Malkiel & Ellis (2020) continue. Fourth, cost efficiency is an important part 

of portfolio management that requires a comprehensive strategy involving a number of 

critical components and a significant amount of decision-making effort. Tax efficiency, min-

imizing transaction costs, optimizing operational processes, and involving diversification 

strategies, are all important (Malkiel & Ellis, 2020). The primary objective is to maximize 

returns while minimizing unnecessary costs, consequently increasing the overall perfor-

mance of the investment portfolio.  

 

Portfolio management also include making risk-related decisions. Managers determine the 

overall level of risk in the portfolio by allocating weights to various asset classes. The man-

ager's decisions are subject to a range of risks, such as the selection of securities, allocation 

choices, and investing style, whether based on a value or growth, or small-cap versus large-

cap investment method (Bodie et al., 2023). Risk management is an important part of the 

investment decision-making process. It includes identifying and evaluating the risks con-

nected with an investment, followed by deciding whether to accept the identified risk in 



30 

the context of the expected return (Malkiel & Ellis, 2020). For example, standard deviation, 

Sharpe ratio, beta, value at risk (VaR) and conditional value at risk (CVaR) are all common 

risk measurements (Bodie et al., 2023; Malkiel & Ellis, 2020), and portfolio managers make 

numerous financial decisions based on the outcomes of these measurements. 

 

Furthermore, portfolio managers employ a wide range of tools to assist them in making 

difficult investment decisions. Financial analysis software is essential for assessing individ-

ual securities, industries, and markets through the analysis of financial statements, in addi-

tion to previous performance (Maham, 2023). Usual analysis types include, for instance, 

ratio analysis, which calculates financial ratios for assessing a company's performance; 

trend analysis, delving into historical financial data to discern patterns; and common-size 

analysis, displaying financial information as a percentage of total sales (Maham, 2023). Ac-

cess to comprehensive market research reports is critical when analysing trends, economic 

indicators, and industry-specific data (Maham, 2023). Additionally, conventional risk man-

agement models, such as standard deviation and beta, enable the evaluation and mitigation 

of various investment risks (Bodie et al., 2023; Malkiel & Ellis, 2020). Meanwhile, quantita-

tive, and qualitative models, often based on mathematical algorithms, assist in the analysis 

of potential investments and the forecasting of market movements (Hayes, 2023). Asset 

allocation tools are critical for improving asset allocation in accordance with investor objec-

tives, risk tolerance, and market conditions. All these different kinds of tools have a sub-

stantial impact on the decision-making process of the portfolio manager. 

 

To enhance the competence of their investment decisions, portfolio managers can leverage 

various software precisely made for effective portfolio management. Prominent software 

options include Bloomberg Terminal, providing real-time financial information; FactSet, of-

fering financial analytics and portfolio management; and Morningstar Direct, facilitating in-

vestment analysis, while Refinitiv Eikon provides financial information and analytics 

(Olumofe, 2023). Comprehensive systems like Charles River IMS and BlackRock's Aladdin 

incorporate all aspects of investment management, including portfolio and risk manage-

ment (Whyte, 2019). Furthermore, MSCI RiskManager and RiskMetrics function as risk 

management and portfolio analysis tools (Whyte, 2019). Several programs seamlessly 



31 

integrate financial analysis, risk assessment, and market research, providing portfolio man-

agers with access to real-time data and comprehensive insights to support informed deci-

sions. In the end, for human portfolio managers, the entire portfolio management is based 

on the manager's own decisions, which are made within the framework of human limita-

tions. 

 

3.2.2 Psychological factors and investment decision-making process 

In the field of asset management, we often narrow our focus to evaluating the performance 

of investments, while the broader context in which investment decisions are made receives 

little attention. As a result, the importance of psychology in financial decisions is under-

stated. Behavioural finance investigates investor behaviour to understand how people make 

decisions. Psychology is often defined as the study of mind or mental behaviour. Mental 

behaviour includes the area of thinking, and thinking contains actions like making decisions.  

 

As a pseudonymous writer of Money Game, Adam Smith (1968, p.25–26) once said: 

It sounds simplistic to say the first thing you have to know is yourself, and of course 
you are not necessarily out to become a professional money manager. But if you stop 
to think about it, here is one authority saying there are not formulas which can be 
automatically applied. If you are not automatically applying a mechanical formula, 
then you are operating in this area of intuition, and if you are going to operate with 
intuition- or judgement- then it follows that the first thing you have to know is your-
self. You are- face it - a bunch of emotions, prejudices, and twitches, and this is all very 
well as long as you know it. Successful speculators do not necessarily have a complete 
portrait of themselves, warts and all, in their own minds, but they do have the ability 
to stop abruptly when their own intuition and what is happening Out There are sud-
denly out of kilter. A couple of mistakes crop up, and they say, simply, “This is not my 
kind or market,” or “I don’t know what the hell’s going on, do you?” and return to 
established lines of defense. A series of market decisions does add up, believe it or not, 
to a kind of personality portrait. It is, in one small way, a method of finding out who 
you are, but it can be very expensive. That is one of the cryptogram which are my own, 
and this is the first Irregular Rule: If you don’t know who you are, this is an expensive 
place to find out. 
 

To make sensible decisions and judgements about holding, selling, or acquiring assets, port-

folio managers must predict the ups and downs of the financial market dynamics. Profita-

bility is dependent on their capacity to identify securities positioned for future price 



32 

increases and decreases. Accurate forecasts can result in substantial gains, highlighting the 

critical necessity of estimating future market movements in portfolio managers' decision-

making process.  

 

The area of financial decision-making has already been examined, with one example being 

the study by Cesarini et al. (2010), which investigates whether genetic variation can explain 

some of the individual differences in investment decisions, which is studied through ob-

serving how individuals differ in building their investment portfolios. The study examines 

the heritability of risk-taking in financial markets and real-life situations, highlighting the 

significant role of genetic variation in explaining individual differences. According to the 

study, genetic variation determines approximately 25 percent of individual variation in port-

folio risk. The study states that genetic factors have a far greater impact on risk-taking be-

haviour than what is previously observed in studies on portfolio selection. Also, the results 

indicate that specific genetic factors could potentially offer insights into why people have 

varying levels of willingness to take risks. (Cesarini et al., 2010). 

 

Personality psychology is an area of psychology that studies the impact of individual per-

sonality characteristics on behaviour. Personality traits have a huge impact on the decision-

making process. The paper by Gambetti and Giusberti (2019) discusses the complex rela-

tionships between personality traits, decision-making styles, and investments. The research 

observes into control variables like gender, income, and experience, finding that these fac-

tors consistently predict investment perceptions and decisions. Additionally, the study re-

veals that men tend to select riskier investment strategies compared to women, and indi-

viduals with more investment experience tend to embrace higher-risk portfolios, underlin-

ing the role of experience in investment decisions (Gambetti & Giusberti, 2019). 

 

The previous studies have indicated that anxious individuals avoid investing or saving 

money and prefer low-risk options like interest-bearing accounts, while those with high self-

control and a solution-oriented mindset are more open to various asset classes (Gambetti 

& Giusberti, 2012; Oehler et al., 2017; Van Winden et al., 2011). These studies highlight a 

positive link between high self-control and long-term asset investments. Additionally, 



33 

extroversion and independence are associated with a willingness to invest in stocks, espe-

cially the aspect of extroversion known as liveliness, which motivates individuals to take on 

financial risks (Dewberry et al., 2013; Gambetti & Giusberti, 2019). 

 

Furthermore, persons with practical, solution-oriented thinking tend to better manage 

stock trend fluctuations through self-management, and those with high self-control have 

the highest skill on predicting stock trends. Conversely, individuals exhibiting traits such as 

impatience, distrust, introversion, unsociability, and traditionalism often perceive higher 

risks in investment decisions. In opposition, calm, and relaxed individuals with competitive, 

strategic thinking and low tendencies for guilt or self-doubt tend to earn higher returns. 

Additionally, individuals with high levels of extroversion, independence, and self-control 

typically adopt a rational, careful approach when evaluating investment options and are 

motivated to engage in investment activities, while highly anxious individuals tend to save 

money and refrain from making investments due to their perception of high risks, low con-

trol and returns (Bensi & Giusberti, 2007; Dewberry et al., 2013; Gambetti & Giusberti, 2019; 

Maner et al., 2007).  

 

Financial decisions are often made under uncertain and complex settings, causing the de-

cision maker to rely on intuition, which plays a major role in most judgments with diverse 

psychological biases. The intuitive decision-making procedure is called heuristics. Using 

heuristics, as demonstrated by Tversky and Kahneman (1974), can lead to numerous cogni-

tive biases and particular fallacies. People are subjected to various "irrationalities" when 

making decisions, and these irrationalities can be categorized into two general group:  First, 

Information processing - investors often fail to process information accurately, resulting in 

incorrect estimations of future probabilities of potential events and related rates of return 

(Bodie et al., 2023). Second, behavioural biases – people frequently make decisions that are 

inconsistent or systematically inefficient, even when they have information about a proba-

bility distribution of returns (Bodie et al., 2023; Slovic, 1972). The potential biases, as well 

as the two stages of decision-making (system 1 and 2) and heuristics, emphasise the degree 

of complexity and nuance involved in making decisions (Tversky & Kahneman, 1974). 

 



34 

Moreover, there are five main types of errors in information processing. For example, Lim-

ited attention, under- and overreaction results from the limited capacity of human attention 

and time, which prevents individuals from effectively processing all available information 

during decision-making, leading to reliance on intuition (Hirshleifer & Teoh, 2005). This can 

cause overreactions to important news and underreactions to less notable information. Ac-

cording to Daniel et al. (1998) and De Bondt & Thaler (1995), Overconfidence occurs when 

investors overestimate their own abilities and beliefs about forecasts, whereas Confirma-

tion bias is the tendency to interpret new information in a way that reinforces or endorses 

our previous beliefs (Wason, 1960). Conservatism bias, on the other hand, results in slow 

adjustments to new information, resulting in underreaction to new information (Kahneman 

et al., 1982). Lastly, representativeness bias occurs when investors draw too quick conclu-

sions about trends or patterns (Barberis et al., 1998). 

 

Even with flawless information processing, people will make decisions that aren´t entirely 

rational. The behavioural biases significantly impact how investors approach the balance 

between risk and return. Framing, the concept by the pioneers in psychological literature, 

for instance, shows in which way the decision is presented, can influence choices (Tversky 

& Kahneman, 1974). For example, framing a decision as an obligation rather than an option 

can result in different outcomes (Tversky and Kahneman, 1979). Mental accounting, as de-

scribed by Thaler (1985), reveals how we give different values to things, like money, based 

on mental categories. This means that decisions can never be completely neutral (Thaler, 

1985). Regret avoidance, in contrast, explains why investors avoid admitting bad investment 

choices and often make emotional decisions instead of logical ones to prevent regret. This 

behaviour is driven by a desire to prevent regret from buying the investment in the first 

place (Bell, 1982; Loomes & Sugden, 1982).  

 

Relatedly, Affect and feelings refer to the personal feelings that an investor may have about 

a particular instrument or company, which may impact investment decisions (Gilovich et al., 

2004; Mellers et al., 1997; Schwarz & Clore, 1988). Lastly, loss aversion is a vital concept in 

prospect theory, developed by Kahneman & Tversky (1979) and is a descriptive model of 

risky decision-making. According to the theory, investors value gains and losses differently, 



35 

preferring perceived gains over perceived losses. The psychological pain of losing is almost 

twice as intense as the pleasure of gaining. And when given two equal options, an investor 

will choose the one in term of potential gains, so whether the uncertainty of returns is 

framed as risky losses or risky gains, matters (Kahneman and Tversky 1979). 

 

Also, a study conducted by Lerner et al. (2015) studies the impact of emotions on decision-

making. According to the findings, emotions have a powerful and constant influence on 

decision-making, and interestingly, certain emotions, such as sadness, can even lead to 

more systematic and deliberate ways of thinking. In addition, Love (2010) found that major 

life events such as divorce, widowhood, and changes in family composition can significantly 

impact optimal portfolio allocations. Divorce and widowhood have a high impact on alloca-

tion, with variations based on gender, number of children, and age. Widowhood particularly 

reduces stock holdings, while divorce leads to divergent portfolio adjustments, with men 

favouring riskier investments and women opting for safer ones (Love, 2010). 

 

Generally speaking, the broad spectrum of behavioural, psychological, and personality-re-

lated biases inherent in human decision-making has a significant impact on humans' invest-

ment decisions. It is fairly probable that these biases also influence the decisions made by 

portfolio managers, thereby affecting their performance. The complicated quality of human 

biases emphasises the difficulties in attaining optimal investment performance from the 

viewpoint of a human portfolio manager. 



36 

4 AI-powered funds 

4.1 Overview of AI-powered funds 

In todays around the clock global market environment, with an extensive range of unique 

and exotic financial instruments, artificial intelligence offers abilities that are rapidly sur-

passing traditional algorithms in finance and trading. In addition, trading systems powered 

by AI can play a significant role in helping traders to make sensible investment decisions 

based on huge amount of available real-time data. 

 

In financial industry, AI plays a significant role by enabling advanced form of algorithmic 

trading, which involves the use of automated algorithms to manage various features of the 

trading process (ESMA, 2023). Advances in quantitative finance and machine learning have 

allowed computers to undertake financial analysis with greater speed and effectiveness 

than humans. In contrast, the complicated nature of financial markets, combined with the 

emergence of new financial products, has made real-time trading decisions difficult for hu-

mans (ESMA, 2023). While algorithmic trading is often used to enhance and automate order 

submissions and executions, it is typically applied only after a portfolio selection has been 

made (Hendershott et al.,2011; Lo et al., 2000). AI, on the other hand, takes a different 

strategy, making decisions early in the portfolio selection process, from the pre-trade to the 

post-trade stage (Abis, 2020). 

 

Several AI-powered funds are in the form of ETFs. As artificial intelligence continues to 

demonstrate its growing capabilities, ETFs have started to harness the power of machine 

learning (ML) and natural language processing (NLP), states Zhang et al. (2023). The utiliza-

tion of AI technologies enables these ETFs to create investment portfolios with superior 

features based on AI technology. AI-powered ETFs are intended to use ML algorithms to 

recognize market patterns and trends in order to make investment decisions, Zhang et al. 

(2023) explain. These algorithms are often trained on massive amounts of historical finan-

cial data, allowing for faster and more accurate data processing than human capabilities, 



37 

they summarize. Next, this research discusses AI-based ETFs that are relevant to the study 

and are used in the comparison together with traditional funds.  

 

The origins of the first AI-based fund can be traced back to a discussion of three experienced 

professionals in a business school class, each of them looking for a means to turn their re-

sumé accomplishments into a thriving business (Field, 2022). Among them were Fidelity 

Investments vice presidents Art Amador, Intel's director of engineering Chida Khatua, and 

Apple's investment portfolio manager Chris Natividad, who came up with the idea for the 

world's first ETF managed entirely by artificial intelligence (Field, 2022). 

 

Afterwards, AIEQ debuted in October 2017, and it was the world's first AI-managed public 

equity ETF, to fully utilize artificial intelligence machine learning techniques as a method for 

stock selection. AIEQ became soon one of the most popular funds in 2017, raising over 70 

million dollars in just a few weeks (ETF Managers Trust, 2023a). 

 

The AIEQ is actively managed ETF which is listed on NYSE Arca and it employs an investment 

strategy that focuses on equity securities listed on U.S. exchanges (ETF Managers Trust, 

2023a). The strategy relies on the EquBot Model, developed by EquBot Inc, and the EquBot 

utilizes IBM's Watson AI platform to conduct a comprehensive analysis of U.S. common 

stocks, including SPACs (Special Purpose Acquisitions Corporations) and REITs (Real Estate 

Investment Trusts), using up to ten years of historical data in combination with recent eco-

nomic and news data (ETF Managers Trust, 2023b). The Fund's investment advisor and sub-

adviser rely on EquBot Model recommendations to determine which securities to buy and 

sell. The AIEQ ETF's primary investment strategy revolves around AI-driven analysis to opti-

mize its portfolio composition and performance (ETF Managers Trust, 2023a; 2023b). 

 

Alongside AIEQ, this research investigates another AI-powered fund listed on Nyse Arca. 

The fund is commonly known by its ticker QRFT, but its full name is QRAFT AI Enhanced U.S. 

Large Cap ETF. It operates as an actively managed exchange-traded fund with a primary 

focus on large-cap stocks traded on U.S. exchanges (QRAFT Technologies, 2023b). QRFT 

aims to achieve long-term capital appreciation by shifting its investments across five factors: 



38 

quality, size, value, momentum, and low volatility. The fund allocates its assets into equity 

securities, including common stock, American Depositary Receipts (ADR), and Global De-

pository Receipts (GDR) (QRAFT Technologies, 2023a). The ETF applies an artificial intelli-

gence system called the QRAFT AI Quantitative Investment System (QRAFT AI) to select 

which stocks to include in the portfolio. While the primary stock selection process heavily 

relies on AI, the fund's automated framework incorporates human intuition and oversight 

in combination with the capabilities of AI (QRAFT Technologies, 2023a). The investment de-

cisions for QRFT are ultimately entrusted to its advisor company which has full discretion 

over investment decisions for the fund (QRAFT Technologies, 2023a; 2023b.) 

 

4.2 Investment decision-making process of AI 

Traditional investment strategies rely on predefined rules or criteria, such as sector, size, or 

quality, to manage a portfolio. This approach can be limiting because it doesn't fully tackle 

the diverse elements of the global market landscape. The critical advantage of AI-powered 

ETFs is their ability to adapt their investment strategies and make decisions based on real-

time market data. For instance, according to ESMA (2023) and Funds Europe (n.d.)., in times 

of increased market volatility, an AI-powered ETF can adjust by allocating more resources 

to assets expected to perform well under volatile market conditions. Similarly, when a new 

investment opportunity emerges, an AI-powered ETF can swiftly analyse relevant data to 

determine its potential as an investment target, they state. 

 

Next, the research will observe how the processes of the AI-powered fund really operate. 

The section begins by introducing the AIEQ operating model, which Chris Natividad and 

Chida Khatua, the founders of AIEQ, present in an article written by Field (2022).  

 

AIEQ is managed by EquBot. The article explains that EquBot is the primary operator of 

AIEQ, and the fund is dependent on EquBot's tens of thousands proprietary models. Every 

day, the EquBot platform collects and analyses data on the around 6,000 US companies that 

AIEQ tracks, Field (2022) writes. This data contains millions of data points from news, social 

media, industry and analyst reports, financial statements, technical, macro, and market 



39 

data, among other things, as well as structured data from third-party data suppliers (Field, 

2022). The EquBot Model also ranks companies, based on their potential to benefit from 

current economic conditions, trends, and world events, selecting approximately 30 to 200 

companies with the greatest potential for appreciation over the next twelve months (ETF 

Managers Trust, 2023a). These selected companies are assigned corresponding weights, 

with the aim of achieving maximum risk-adjusted returns compared to the broader U.S. 

equity market (ETF Managers Trust, 2023a). 

 

In addition, EquBot applies IBM Watson to support monitoring AI models, assist in its in-

vestment decision-making procedures for selecting securities, and extract insights from 

data, Field (2022) asserts. IBM's Watson AI is a powerful computing platform that can pro-

vide responses to presented questions, and it achieves this by connecting extensive data 

sets, encompassing both structured data and unstructured data, states Field (2022). It 

learns via structured data, such as financial statements, growth, expenditure on R&D, and 

market movements, he adds. However, he notes that it also obtains insights from unstruc-

tured data, such as news stories, blogs, social media, and company announcements. The 

IBM Watson platform utilizes machine learning, sentiment analysis and natural language 

processing in its processes, according to Rothney (2021). Watson AI constantly learns and 

improves from each analysis it performs, such as recognizing patterns, thereby enhancing 

the accuracy of its responses with each subsequent inquiry, she continues. IBM Watson 

monitors over 80,000 AI models, which is far too numerous group for humans to monitor 

(Field, 2022). After all, the platform operates like an equal army of 1,000 research analysts, 

traders, and quantitative analysts operating around the clock (ETF Managers Trust, 2023a). 

 

Moreover, as specified by Field (2022), the AIEQ’s AI models are trained on five to 30 years 

of historical data, with more emphasis placed on recent data. The models are trained on a 

cost function, which implies that the model forecasts the expected return for each historical 

data point, such as an older news story from 2001, he explains. Likewise, he says that the 

models incorporate trust points to differentiate data sources, e.g., the models assign differ-

ent weights to a New York Times article compared to a blog post. All this data is then 



40 

integrated into knowledge graphs by EquBot, whose serve as important educational tools 

for AIEQ, he summarizes. 

 

Whereas the system depends on a substantial group of 80,000 models, three of these mod-

els play a particularly significant role in shaping its decisions. Field (2022) talks about these 

models in his article. The key AI models are: Financial Model, which evaluates a company's 

financial state and performance over different time horizons, primarily using earnings and 

spending data. Quality Model, which utilizes around 170-line items, such as innovation rank-

ing, to assess a company's current quality. And lastly, Sentiment Analysis Model, which ap-

plies IBM Watson's natural language processing tools to extract metadata and analyse the 

sentiments of over a million content pieces daily. In addition, EquBot employs a combina-

tion of internal tools and IBM Watson's OpenScale tool to continuously monitor 10 key met-

rics for each model (Field, 2022). These metrics help to flag any potential bias or deviations 

in the models' behaviour, while they track the decision-making processes of each model 

using decision trees, Field continues. Besides, two people monitor the actions of potentially 

biased metrics full-time, while the individual owners of each AI model check for any red 

flags and warnings daily to ensure responsible AI-driven decision-making (Field, 2022).  

 

The figure 5. shows the AIEQ’s investment process. AIEQ builds four DL prediction models 

for each analysed company, which are: finance, news and information, management, and 

macro. These all have multiple underlying signals (ETF Managers Trust, 2023b). 

 

 
Picture 1. AIEQ’s investment process (ETF Managers Trust, 2023b). 
 



41 

In addition to AIEQ, QRFT ETF uses an AI-based decision-making platform called QRAFT AI 

in its investment processes. The QRFT ETF selects assets using a unique AI algorithm that 

discovers patterns, signals, and connections through analysing data (QRAFT Technologies 

Inc., 2023a). QRAFT AI employs machine learning and deep learning technologies in its op-

erations (QRAFT Technologies Inc., 2023b). 

 

According to the fund prospectus sheet by QRAFT Technologies (2023a) the QRAFT AI eval-

uates, and filters information from the database based on defined criteria, to support the 

fund's defined investment thesis. The prospectus defines also that QRAFT AI selects and 

weights NYSE and NASDAQ-listed US companies by defined factors in order to provide broad 

exposure to a range of market factors affecting the US market. These factors are Quality 

(company's profitability), Size (market capitalization), Value (the company's market value 

compared to its book value), Momentum (the security's recent price return versus to the 

overall market over time), and Volatility (security's systemic risk versus the overall market 

as a whole). This collection of data is called the "Database of Large US Companies" (QRAFT 

Technologies Inc., 2023a; 2023b). 

 

Moreover, the fund prospectus describes the investment process of QRAFT AI (QRAFT Tech-

nologies Inc., 2023a). At first, QRAFT AI evaluates each stock's relative price appreciation 

potential in comparison to other companies over the next four weeks, and this evaluation 

involves utilizing deep learning methods, which include handling massive amounts of data. 

Then, the system examines the distribution of each stock's relative potential for price ap-

preciation during this period, using complex deep learning structures like Bayesian neural 

networks to estimate the level of uncertainty in its forecasts. Next, based on this analysis, 

QRAFT AI selects the top 300 to 350 stocks from the database by averaging the distribution 

of their relative potential for price appreciation. QRAFT AI also compresses this data and 

assesses how each individual factor may evolve and impact a company over time. This pro-

cess identifies companies with the highest potential to outperform their U.S. large-cap 

peers in the upcoming four-week period. The equities in the database are then weighted 

according to a methodology designed to optimize risk-adjusted returns when compared to 

other companies. Afterwards, the final portfolios are provided to the U.S. Large Cap 



42 

Database for use by the Adviser’s financial experts. QRAFT AI repeats these procedures 

every four weeks, and the financial experts at the Fund's Adviser make or adjust invest-

ments in the fund based on the newly generated information (QRAFT Technologies Inc., 

2023a). 

 

All in all, artificial intelligence finds functions in trading through two key opportunities. First, 

it offers trading strategy recommendations, and secondly, it runs automated trading sys-

tems that not only make predictions but also determine the appropriate actions and even 

execute trades. AI-powered trading systems can autonomously identify and execute trades, 

functioning independently without human involvement (OECD, 2021a). 

 

4.2.1 AI techniques and models for investment decision-making 

Artificial intelligence technologies and models for making investment decisions are based 

on machine learning, which is the fundamental technology of artificial intelligence. Several 

major proprietary trading firms have incorporated ML models into their trading strategies 

and ML is already largely employed in trading activities including liquid assets, such as eq-

uities, futures, and foreign exchange, due to the multitude of real-time data for these in-

struments (BoE & FCA, 2019; ESMA, 2023). Generally, different ML approaches and models 

can accomplish different things, and each one tends to succeed at specific functions, mak-

ing them suitable for different purposes (BoE & FCA, 2019). Often, the best results come 

from combining predictions and opinions from various AI techniques, known as ensemble 

methods, which have been proven to generate more accurate results than any single 

method alone (BoE & FCA, 2019). 

 

For instance, machine learning approaches such as LASSO regressions, elastic nets, and ar-

tificial neural networks (ANNs) have natural mechanisms for selecting the most important 

components from data set, increasing the reliability of predictions. According to research 

like Feng et al. (2020) and Freyberger et al. (2020), LASSO regression can automatically find 

the most relevant parameters for predicting future returns from a large pool of return-pre-

dictive signals. Furthermore, the LASSO technique can be used to find lead-lag correlations 

between various asset classes or markets, allowing for evaluation of which industry or 



43 

market returns act the most crucial role in predicting returns compared to all other markets 

or industries, they confirm. In addition, advanced versions of LASSO regression, known as 

"elastic nets," offer a balanced approach by ensuring that estimated coefficients do not 

become excessively large, which reduces the chance of the model "overfitting" and reduces 

spurious coefficient estimates to zero, considerably improving the model's performance 

(Bartram et al., 2021; BoE & FCA, 2019; Feng et al., 2020; Freyberger et al., 2020). 

 

Then again, ML approaches such as artificial neural networks, support vector machines, and 

tree-based models are successful at detecting non-linear patterns, such as how input vari-

ables interact (BoE & FCA, 2019). This ability increases the creation of single and multi-

factor signals by collecting more complicated correlations and intricate details in the input 

data (Bartram et al., 2021; Bartram et al., 2020). 

 

As mentioned in the paragraph introducing the working models of AI, Artificial Neural Net-

works (ANNs) are computer algorithms that imitate the neural network structure of the 

human brain. They learn by adjusting connection weights to minimize errors between pre-

dicted and desired data labels, making them valuable for tasks such as stock price predic-

tions (Weng, 2022). They are also competent at pattern recognition, capable of identifying 

complex patterns in data (Montesinos et al., 2022). Moreover, Bayesian neural networks, a 

variant of ANNs, employ Bayesian reasoning to better understand the probability distribu-

tion associated with various neural network configurations. Bayesian networks can be used 

to anticipate execution shortfall as a measure of transaction costs (Pan et al., 2021; Wu, 

2021). This method is especially beneficial when data is missing as it may generate the most 

likely result based on the existing data (Ticknor, 2013). The approach prevents overfitting, 

facilitates learning from limited data, and offers a measure of confidence in predictions. 

Essentially, Bayesian neural networks incorporate probabilistic reasoning to enhance the 

robustness and informativeness of their predictions (Ticknor, 2013).  

 

In contrast, tree-based models are an example of one of the most common of ML trading 

strategies. A simplest form is a decision tree, which is used for classification and regression 

tasks (Kumar & Ravi, 2007). Buschjager and Morik (2018) express that it features a 



44 

hierarchical tree structure with a root node that doesn't have any incoming branches, in-

ternal nodes (decision nodes), and leaf nodes. They add that both root and decision nodes 

evaluate and partition the data into more comparable subsets, which are represented by 

the leaf nodes or terminal nodes. The leaf nodes symbolize all the potential outcomes inside 

the data set (Buschjäger & Morik, 2018). Figure 5. Presents the decision tree algorithm.   

 

 

 
 
 
 
 
 
 
 

Figure 5. Example of decision tree algorithm.  
 

Equally, according to Buschjager & Morik (2018) and Ho (1995) random forest is an ad-

vanced model that is made up of several decision trees to produce a single outcome. Unlike 

traditional decision trees, each individual decision tree makes its own predictions, which 

are then combined using an averaging process to form the random forest's predictions, they 

continue. This algorithm is useful for solving regression or classification problems. Both 

methods, decision trees and random forest, are proven to be quite successful in predicting 

outcomes in traditional financial data analysis scenarios, such as forecasting stock prices 

and identify patterns in market data (BoE & FCA, 2019). They can also be utilized to deter-

mine whether to buy or sell a stock, based on factors such as current price, trading volume, 

and market trends (BoE & FCA, 2019). The basic theory is that several decision tree models, 

each with a unique perspective, produce more accurate forecasts than a single decision tree 

(Buschjäger & Morik, 2018) 

 

Likewise, Support Vector Machines (SVM) are supervised algorithms that have applications 

in both classification and regression. According to Hao et al. (2013), they are also known for 

their resistance to overfitting, which ensures reliable learning from training data. SVMs are 

effective in learning boundaries that separate feature spaces into distinct classes, allowing 

new data points to be classified, they state. However, that the computational requirements 

Eagle Bear Dog Guinea Pig 

No No Yes Yes 

Yes No 

Can fly? 

Has feathers? 

Has tail ? 



45 

of SVM make the model unsuitable for large data sets, they add. SVMs are powerful in pat-

tern recognition, data analysis, and finding insights and relationships from data set (Hao et 

al., 2013). Basically, they operate by taking inputs and providing valuable results, allowing 

them to find hidden patterns in the data, Hao et al. (2013) summarises. SVM can also be 

used for portfolio selection, by using a model based on the predictions it generates (Hao et 

al., 2013). 

 

Conversely, Natural language processing (NLP) tools can be applied in ML to build factors 

based on textual input from sources such as corporate annual reports and news articles. in 

the words of Qian et al. (2022) NLP models are useful for sentiment analysis because they 

are effective at analysing unstructured textual data and can derive relevant information 

from it. Equally, sentiment analysis is a technique for analysing the relationship between 

market movements and financial news. For example, news about a company and general 

stock market can have a substantial effect on stock movement: thus, sentiment analysis aids 

in assessing market sentiment and making informed investment decisions, Qian et al. (2022) 

affirm. It is also an excellent tool for examining unstructured content about a specific com-

pany in order to identify inconsistencies and anomalies (Landauer et al., 2023). 

 

Modern Portfolio Theory (MPT), also known as Mean-Variance Optimization Model (MVO), 

presented by Markowitz in 1952, and the Capital Asset Pricing Model (CAPM) an extension 

of MPT, developed by Sharpe in 1964, are foundational theories of the relationship between 

risk and return in investment decisions. Based on Lin & Liu (2008), traditional portfolio op-

timization strategies, such as Markowitz MVO Model (1952), have limitations due to their 

rigid structure and the difficulties in precisely calculating expected returns and variance-

covariance inputs. ML technologies can solve these restrictions by providing more accurate 

estimates of expected returns as well as replacing the variance-covariance matrix with more 

reliable alternatives, they assert. Additionally, genetic algorithms, rooted in Darwin's and 

Matthew's (1859) theory of natural selection, are valuable in portfolio optimizat