Joachim Niang Artificial intelligence and hedge fund performance An analysis of hedge fund trading styles Vaasa 2021 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: Joachim Niang Title of the thesis: Artificial intelligence and hedge fund performance : An analysis of hedge fund trading styles Degree: Master of Science in Economics and Business Administration Programme: Master’s Degree Programme in Finance Supervisor: Klaus Grobys Year of completing the thesis: 2021 No. of pages: 132 ABSTRACT: This study focuses on understanding the relationship between the level of automation employed by hedge funds on the level of performance that these funds are able to obtain. As technologies are constantly evolving and being used to further different fields, one could ask if the adaptation of the latest technological advancements in term of artificial intelligence could be used to fur- ther the trading performance of hedge funds. As hedge funds enjoy less restrictions for their trading processes, they are at a prime position to take advantage of every edge that can be obtained. Using data from the Preqin hedge fund database we can to uncover this level of automation by sorting funds based on their trading styles. The term AIML hedge funds refers to hedge funds using both artificial intelligence and machine learning. These AIML funds are taken as their own trading style and their performance is compared against systematic, discretionary and combined funds which utilize both the systematic and the discretionary methodologies in their trading processes. Using both the efficient market hypothesis and the behavioral finance frameworks, we are able to conduct a detailed analysis of both the motivation for the need of automation and for the existence of hedge funds. Past literature relating to hedge fund performance, artifi- cial intelligence and algorithmic trading, and hedge fund comparisons are also reviewed in de- tail. By only focusing on funds that trade U.S equities we are able to utilize common factor mod- els used for pricing U.S. equities. Performance is analyzed both in terms of the full sample period and by employing subsample analysis to uncover underlying performance persistence. Based on the results of our factor models we are able to see the statistically significant overper- formance shown by AIML funds. Moreover, our subsample analysis supports these findings and shows that the performance obtained by AIML funds is persistent. When the effects of serial correlation between the fund types is taken into account the outperformance of AIML is further established. Lastly, when comparing the alphas of AIML funds against the other hedge fund trad- ing style portfolios, AIML funds exhibit statistically significant outperformance even at a one percent level of significance. Thus, our results indicate that by using artificial intelligence hedge funds can improve their performance on a persistent basis and to stand out from their peers. Our results are not in breach of the efficient market hypothesis as the underlying reasons for AIML fund performance can be noted as their ability to adapt and their ability to take advantage of small market dislocations. Behavioral finance also shows how adaptability combined with an emotionless ability to execute strategies are key for AIML outperformance Our findings present interesting directions for future research and showcase the likely future trend of increased AI usage within the hedge fund industry. KEYWORDS: Hedge fund, Artificial intelligence, AIML, Systematic, Discretionary 3 VAASAN YLIOPISTO Laskentatoimen ja rahoituksen akateeminen yksikkö Tekijä: Joachim Niang Tutkielman nimi: Artificial intelligence and hedge fund performance : An analysis of hedge fund trading styles Tutkinto: Kauppatieteiden maisteri Oppiaine: Rahoitus Työn ohjaaja: Klaus Grobys Valmistumisvuosi: 2021 Sivumäärä: 132 TIIVISTELMÄ: Tämä tutkimus keskittyy ymmärtämään suhdetta hedgerahastojen käyttämän automatisaation ja niiden saavuttaman suorituskyvyn välillä. Koska tekniikoita kehitetään jatkuvasti ja niitä käy- tetään eri aloilla, voidaan teorisoida, josko uusimpien teknisten kehitysten ottaminen osaksi hedgerahastojen strategioita johtaa tulosten paranemiseen. Koska hedgerahastoilla on vähem- män rajoituksia kaupankäyntiin käytettäville strategioille, ne ovat ensisijaisessa asemassa hyö- dyntämään kaikkia saatavia etuja. Preqin-hedgerahastotietokannan avulla pystymme löytämään vastauksen tähän tutkimuskysy- mykseen, lajittelemalla rahastot niiden kaupankäyntityylien perusteella. Termillä AIML-hedge- rahastot viitataan hedgerahastoihin, jotka hyödyntävät sekä tekoälyä että koneoppimista. Nämä AIML-hedgerahastot otetaan omaksi kaupankäyntityylikseen ja niiden tuottoa verrataan syste- maattisiin, harkinnanvaraisiin ja yhdistettyihin hedgerahastoihin, jotka käyttävät kaupankäyn- nissä sekä systemaattista että harkinnanvaraista menetelmää. Käyttämällä sekä tehokkaiden markkinoiden hypoteesia että käyttäytymistaloustiedettä teoreettisina viitekehyksinä, voimme suorittaa yksityiskohtaisen analyysin sekä automatisaation tarpeen, että hedgerahastojen ole- massaolon perusteista. Hedgerahastojen suorituskykyyn, tekoälyyn ja algoritmeihin sekä hed- gerahastojen keskinäiseen vertailuun liittyvää aiempaa kirjallisuutta tarkastellaan myös yksityis- kohtaisesti. Keskittymällä vain Yhdysvaltojen osakkeilla kauppaa käyviin hedgerahastoihin voimme hyödyntää Yhdysvaltojen osakkeiden hinnoittelussa käytettyjä yleisten riskifaktoreiden malleja. Suorituskykyä analysoidaan sekä koko otosjakson perusteella että käyttämällä osaotos analyysiä paljastamaan taustalla olevan suorituskyvyn jatkuvuus. Riskifaktorimallien tulosten perusteella voimme nähdä AIML-rahastojen osoittaman tilastolli- sesti merkitsevän ylituoton. Lisäksi osaotos-analyysimme tukee näitä havaintoja ja osoittaa, että AIML-rahastojen näyttämä suorituskyky on jatkuvaa. Kun kaupankäyntityylien välisen sarjakor- relaation vaikutukset otetaan huomioon, AIML-rahastojen ylivertaisuus todetaan edelleen. Lo- puksi, verrattaessa AIML-rahastojen alfa-arvoja muiden hedgerahastojen kaupankäyntityylien arvoihin, AIML-rahastot saavat tilastollisesti merkitsevän ylivertaisen tuloksen jopa prosentin merkitsevyystasolla. Tuloksemme osoittavat, että tekoälyn avulla hedgerahastot pystyvät pa- rantamaan suorituskykyään jatkuvasti ja erottumaan muiden kaupankäyntityylien joukosta. Tu- loksemme eivät ole ristiriidassa tehokkaiden markkinoiden hypoteesin kanssa, sillä AIML-rahas- tojen ylituoton taustalla olevina syinä voidaan mainita niiden kyky sopeutua ja kyky hyödyntää pieniä markkinahäiriöitä. Käyttäytymistaloustiede osoittaa myös, kuinka sopeutumiskyky yhdis- tettynä automatisoituun kykyyn toteuttaa strategioita ovat avainasemassa AIML- rahastojen tu- losten saavuttamisessa. Tuloksemme esittävät mielenkiintoisia suuntaviivoja tulevia tutkimuksia varten ja näyttävät tekoälyn todennäköisen enenevissä määrin kasvavan merkityksen hedgera- hastojen keskuudessa. AVAINSANAT: Hedge fund, Artificial intelligence, AIML, Systematic, Discretionary 4 Contents 1 Introduction 8 1.1 Purpose of the study 10 1.2 Research hypothesis 11 1.3 Contribution 12 1.4 Structure of the thesis 13 2 Theoretical background 15 2.1 Efficient-market hypothesis 16 2.1.1 Active investing 21 2.1.2 Passive investing 24 2.2 Behavioral finance 26 2.2.1 Discretionary trading 31 2.2.2 Systematic trading 33 3 Literature review 35 3.1 Hedge fund performance 36 3.2 Algorithmic trading and AI 42 3.3 Discretionary versus systematic approach 50 4 Hedge fund characteristics 55 4.1 Main characteristics 56 4.2 Discretionary funds 64 4.3 Systematic funds 67 4.4 Combined funds 70 4.5 AIML funds 71 5 Data and methodology 75 5.1 Data description 77 5.2 Methodology 86 5.2.1 Capital asset pricing model 87 5.2.2 Fama-French three-factor model 89 5.2.3 Carhart four-factor model 90 5 5.2.4 Fama-French five-factor model 91 5.2.5 Summary 92 6 Empirical results 94 6.1 Hedge fund performance 96 6.2 Hedge fund performance persistence 106 6.3 Multiple equation models 110 6.4 Summary of the results 112 7 Conclusions 114 References 121 Appendices 132 Appendix 1. List of funds 132 6 Figures Figure 1. Hedge Fund Management Fee Distribution and Mean Management Fee by Year of Inception (Preqin 2021, p. 30). 59 Figure 2. Distribution of Hedge Fund Managers and Industry AUM by Manager Location (Preqin 2021, p. 26). 59 Figure 3. Proportion of Number and AUM of Hedge Funds by Top-Level Strategy (Preqin 2021, p. 27). 61 Figure 4. Fund Manager Expectations for Hedge Fund Industry AUM in 2021 (Preqin 2021, p. 124). 62 Figure 5. Fund Manager Expectations for the Performance of the Preqin All-Strategies Hedge Fund Benchmark in 2021 (Preqin 2021, p. 123). 63 Figure 6. Number of funds per portfolio through the sample period. 82 Figure 7. AUM per portfolio through the sample period. 83 Figure 8. Annualized excess returns per portfolio through the sample period. 84 Figure 9. Annualized Sharpe-ratios per portfolio through the sample period. 85 Figure 10. Cumulative excess returns per portfolio through the sample period. 86 Figure 11. Time-varying MKT-factor. 100 Figure 12. Time-varying SMB-factor. 101 Figure 13. Time-varying HML-factor. 102 Figure 14. Time-varying RMW-factor. 103 Figure 15. Time-varying CMA-factor. 104 Figure 16. Time-varying UMD-factor. 105 Tables Table 1. Number of funds per trading style and most common equity strategies. 79 Table 2. Descriptive statistics of individual funds in sample per trading style. 80 Table 3. Descriptive statistics of trading style portfolios. 81 Table 4. Performance measurement CAPM. 96 Table 5. Performance measurement Fama and French three-factor model. 97 7 Table 6. Performance measurement with Carhart four-factor model. 98 Table 7. Performance measurement with Fama and French five-factor model. 99 Table 8. Performance persistence CAPM. 107 Table 9. Performance persistence Fama and French three-factor model. 108 Table 10. Performance persistence Carhart four-factor model. 108 Table 11. Performance persistence Fama and French five-factor model. 109 Table 12. Correlation between different trading style portfolios. 110 Table 13. Seemingly unrelated regression. 111 Table 14. Wald coefficient test. 112 Abbreviations AI Artificial intelligence AIML Artificial intelligence machine learning AUM Assets under management BF Behavioral finance CAPM Capital asset pricing model CTA Commodity trading advisor EMH Efficient market hypothesis ETF Exchange-traded fund FSA Financial supervisory authority HFT High-frequency trading ML Machine learning SEC Securities and Exchange Commission 8 1 Introduction Hedge funds are often characterized by the different strategies they employ. Agarwal et al. (2018) for one note that with the reduced regulation that they face, they are often both more willing and capable to pursue alternative methods of investing and utilize strategies that are not available to most other major market participants. At the same time, continuous technological advances have created a lot of new possibilities when it comes to the implementation and creation of different trading strategies. Kooli and Stet- syuk (2020) detail the extremely competitive environment that hedge funds operate in and the growing pressure that this creates towards hedge fund fee structures in relation to recent performance. Thus, there is a clear incentive in exploring the viability of the latest technologies. According to the Investment Company Fact Book (2021), the total net assets for passive exchange-traded funds (ETFs) have grown by a multiplier of four in just the U.S. in the past decade, surpassing 4 trillion dollars. Grégoire (2020) outlines that passive manage- ment represented 42% of all US equity mutual funds in the year 2016 and noted a similar compounding growth. As the importance of passive investing is continuing to grow and receive more capital inflows, this results in there also being capital outflows from some other types of investment vehicles. The Preqin (2020, p. 12) report details this effect as record capital outflows as of late are seen for the wider hedge fund industry in 2019. Kooli and Stetsyuk (2020) show that there has been a full percentage point reduction in the capital managed by the industry in 2018, which in term makes it more apparent that the performance-to-fee relationship of many funds is not adequate for many of their investors. Kooli and Stetsyuk (2020) uncover this relationship between fees and performance fur- ther by trying to uncover the real value added by hedge fund managers. In short, they note the same worrying trend where hedge funds do actually perform better than their passive counterparts, but the same cannot be said after taking into account the notably 9 higher fees. Inversely, the Investment Company Fact Book (2021) data shows the clearly declining expense ratios of both index mutual funds and their ETF counterparts, noting that while increased competition plays a role for these continuously reduced fees, also the inherent nature of passive management is essential to take into account. As there is no need for costly active management and active analysis of the traded instruments, the expenses are also significantly reduced. Hence, we see that the hedge fund industry is currently facing some very meaningful challenges. The competition is increasing, and it is also coming from players from outside the industry, mostly represented by passive management. Preqin (2020, p. 24-25) shows that this results in pressure to reduce fees especially during times where the perfor- mance figures are lackluster. As the complex active analysis of securities along with the development of related trading strategies represents both the differentiating factor for hedge funds compared to other fund types and the reasoning for the added fees, it begs the question whether this process of being active can benefit from the various techno- logical advances that have been introduced, from trading algorithms all the way to arti- ficial intelligence (AI). The usage of computerized trading is nothing new as it dates back to the early 1970s and the creation of different types direct market access programs that enable the connection between an algorithm and a stock exchange. Kim (2010, p. 1-4) details that these auto- mated trading systems have then gotten more complex over time, taking on an ever- bigger role in the trading process. Gerlein et al. (2016) show that AI on the other hand can be understood as an evolutionary step from simple automations to tasks where the responsibilities of the algorithm come close to, or even replace, the end-user. Harvey et al. (2017) additionally note that while algorithms are already commonplace for many of the worlds hedge funds, their roles differ to a great extent. In their research paper hedge funds are divided into discretionary hedge funds which make the trading decisions manually and to systematic hedge funds in which the trading decisions are 10 made almost or fully by using trading algorithms. Even with this clear split, the authors note that as technology advances, it may become ever harder to distinguish between the two. The technology is here to stay, but its advantages remain to be seen. 1.1 Purpose of the study The purpose of the study is to research whether using the latest advances in terms of AI leads to meaningful advantages when it comes to the performance of hedge funds. In addition to the performance figures, some descriptive statistics on these funds will also be uncovered as they will reveal important information relating to these funds. One could for example hypothesize that using AI will not only lead to levels of increased re- turns, but also to possible reductions in costs. On the contrary one could also theorize that by using AI these hedge funds are able to provide higher returns that better justify these high costs. As such the general analysis of the data is also of importance. Capocci and Hübner (2004) detail that there are almost 6000 funds managing around $400 billion in capital and as such hedge funds justify an increased attention in financial press as well as in the academic world. The Preqin (2021, p. 5) Global Hedge Fund Report allows us to glimpse the most current figures for the industry and the global assets under management (AUM) now stand at around 3,87 trillion dollars. Additionally, the report notes that there are 18 303 active hedge fund managers, meaning that the number of individual hedge funds is even greater. Also, of interest is the forecasted growth that these AUM figures are expected to reach. With a forecasted compounded annual growth rate of 3,6%, the total AUM is predicted to go as high as 4,28 trillion dollars by 2025. That is why if anything, the importance of researching hedge funds is even more justified than before. The main purpose of this thesis can therefore be simplified as an analysis on the effect of the degree of automation on the degree of performance. In other words, whether having less direct interaction with the trading decisions and handing more control into 11 the hands of differently advanced trading algorithms would yield better returns. This control would then revolve from automated and predictive analysis to fully automated decision making, where to role of the human manager would shift more into that of an observer, with continuously lesser involvement in the day-to-day trading decisions. Therefore, this thesis aims to uncover whether advances in computing, spanning from simple trading algorithms all the way to extremely complex and completely self-suffi- cient systems are truly applicable when it comes to the quest of hedge funds aiming to outperform. The topic is especially current as AI is starting to impact a lot of different fields, from medicine to self-driving cars, and one could then make inferences that simi- lar developments are bound to take place in the financial markets as well. 1.2 Research hypothesis The research question for this thesis is whether the usage of AI is able to improve hedge fund performance. For this purpose, the following hypothesis pair will be used: 𝐻0: 𝐴𝐼𝑀𝐿 ℎ𝑒𝑑𝑔𝑒 𝑓𝑢𝑛𝑑𝑠 𝑑𝑜 𝑛𝑜𝑡 𝑜𝑢𝑡𝑝𝑒𝑟𝑓𝑜𝑟𝑚 𝑓𝑢𝑛𝑑𝑠 𝑜𝑓 𝑐𝑜𝑛𝑣𝑒𝑛𝑡𝑖𝑜𝑛𝑎𝑙 𝑡𝑟𝑎𝑑𝑖𝑛𝑔 𝑠𝑡𝑦𝑙𝑒𝑠 𝐻1: 𝐴𝐼𝑀𝐿 ℎ𝑒𝑑𝑔𝑒 𝑓𝑢𝑛𝑑𝑠 𝑜𝑢𝑡𝑝𝑒𝑟𝑓𝑜𝑟𝑚 𝑓𝑢𝑛𝑑𝑠 𝑜𝑓 𝑐𝑜𝑛𝑣𝑒𝑛𝑡𝑖𝑜𝑛𝑎𝑙 𝑡𝑟𝑎𝑑𝑖𝑛𝑔 𝑠𝑡𝑦𝑙𝑒𝑠 The funds of conventional trading styles are defined as funds that do not use AI, even if standard trading algorithms are in place. The outperformance is then measured by com- paring the possible alphas obtained by these AI hedge funds against the alphas exhibited by their conventional counterparts. In addition to this, the performance in terms of ex- cess returns by using factor models is evaluated and the persistence of this performance is also uncovered to make a strong case for both the outperformance against conven- tional style funds along with the persistence of said performance. 12 1.3 Contribution This thesis aims to contribute to the growing literature on hedge funds and hedge fund performance in various ways. Firstly, it helps to understand a lot of general information on hedge funds and brings forth various up-to-date figures on the field. This in itself is already valuable as is noted by Capocci and Hübner (2004) as they detail the general difficulty in obtaining data on individual funds. They note both the prevailing secrecy within the field along with the fact that hedge funds are not legally required to reveal almost anything in regard to their trading, allowing them to operate out of the eye of the general public. Fung and Hsieh (1999) detail this further by noting the sophisticated investors that hedge funds are only allowed to attract, meaning that regulators and financial supervisory au- thorities (FSAs) do not impose strict restrictions for the types of investments that these funds are able to pursue, or the types of disclosures that they are mandated to give. Treleaven et al. (2013) add to the theme but from the viewpoint of using trading algo- rithms of varying sophistication, noting the difficulty in finding the details on their usage. Therefore, being able to bring to light a very comprehensive and current dataset on hedge funds which shows factors such as AI usage and additional specifics such as the average size of costs for investing in such funds can be thought as being a strong contri- bution for the research within the field. As for any research on hedge funds, data is the most valuable and precious asset. The main contribution of this thesis will still naturally be on uncovering whether AI usage can improve the performance of hedge funds and help them obtain performance that sets them apart from their conventional peers. Certain publicized research papers have already researched the performance differences between systematic and discretionary funds and Harvey et al. (2017) for one come to the conclusion that combining both ap- proaches is the best course of action. Our research paper categorizes funds into system- atic, discretionary and combined funds which use both the before mentioned 13 approaches which enables us to revisit the findings found in their study. The main con- tribution being of course the addition of a fourth category, AI funds which combine ele- ments from all the other categories but using an advanced technological framework that sets them apart from the rest. Therefore, research into hedge fund performance will be especially furthered as to the best of my knowledge no publicized studies carrying similar performance comparisons for hedge funds are available. The aim will be to give both a detailed outlook into the industry and a detailed analysis on the performance that can be obtained with the latest tools available to these funds. 1.4 Structure of the thesis The structure of the thesis will be the following. Firstly, relevant theoretical frameworks for the topic at hand will be researched, with both the efficient market hypothesis (EMH) and the behavioral finance (BF) being the theories that best cater to our needs. This can be justified in short as the EMH allows us to distinguish the reasoning for the need for active trading hedge funds in the first place if the markets are fully efficient and BF helps us to understand some of the underlying reasons why one would want to automate their trading activities and remove behavioral factors from the investment process in the first place. Secondly, we will focus on the relevant literature within the field, starting from research papers focused on comparing hedge fund performance. These studies will then be ana- lyzed and compared in a detailed manner in terms of their main findings and conclusions. The second type of research papers will be ones dedicated to algorithmic trading and AI being used to improve trading performance in general, as this will help give an outlook into what the potential benefits of using such systems are and as such it will also help detail the motivations that hedge fund managers may generate towards employing them. Lastly, we will review literature on research papers comparing systematic and 14 discretionary funds as this is the topic most related to our theme. This choice of litera- ture review topics should help give a general overview of hedge fund performance, al- gorithm and AI performance, and the comparative performance studies done for the hedge funds so far. The third step in our structure will be to uncover as much information as possible on hedge funds themselves and to help obtain a deeper level of understanding of both the individual funds and the industry that they operate in. Along with main characteristics, focus will naturally be on each of the categories of hedge fund trading styles that we have already outlined so that also the non-numerical side of these funds gets uncovered. The fourth step will be the data and methodology stage of the thesis, where very recent figures going up to January 2021 will be shown relating to our sample of funds and a lot of additional metrics such as AUM and fees will be seen for each hedge fund type. Addi- tionally, our methodology for the performance comparison will be discussed. The fifth step is the actual empirical analysis of our data of hedge funds using various factor models. As was mentioned before in the hypothesis section, raw performance, performance persistence and performance comparison to peers will all be carried out to obtain robust results to either accept or reject our null hypothesis. The sixth and final step will naturally be the conclusion where our results are reflected against the results based on the analysis of theory, the results found in the literature review and on the findings based on our general analysis of funds. Thus, the meaning- fulness of our discoveries gets evaluated from a wider and more profound perspective. 15 2 Theoretical background There are a lot of different theoretical frameworks that can be considered for studying both the hedge fund industry and the implications of using AI technologies for asset management. For the purpose of this thesis, the relation between AI usage and hedge fund performance will be explored through the viewpoints presented by the EMH and its somewhat competing counterpart, BF. Fung and Hsieh (1999) note that hedge funds are in general characterized by their usage of dynamic and non-passive strategies. The authors also detail that hedge funds are ac- tive participants within the markets while on the other hand they also display very little correlation with the markets that they themselves operate in. Sun et al. (2012) remark the extreme secrecy among the entire hedge fund industry and the need for these funds to keep trading strategies secret, along with the fees they impose upon their clients that are very high when compared to other investment vehicles. Thus, it can be theorized that if investors are willing to pay a notable premium for having a hedge fund actively manage a portion of their wealth, some market-beating returns are likely to be obtained. Therefore, one could pose a question on whether being active could lead to better returns as opposed to plain buy and hold passive investing. Barber and Odean (2000) on the other hand find an opposing view and note that increased ac- tive trading leads to lower returns. Hence, using efficient-market hypothesis as the first main frameworks for this thesis is more than justified. For BF aspects, the most immediate reasoning for its inclusion can be deducted from the man versus machine setting that is ultimately being reviewed in this thesis. Humans are emotional beings and are affected by a wide variety of different behavioral biases, from overconfidence to loss aversion. Algorithms and machines on the other hand can be typ- ically thought of as emotionless machines that follow their specific objectives no matter what. Ritter (2003) especially writes that most people suffer from overconfidence in 16 their abilities and Statman et al. (2006) and Barber et al. (2005) are able to link overcon- fidence with a very active level of trading. Again, it can be conceptualized that if we all suffer from behavioral biases that are det- rimental to our performance as investors, using AI and algorithmic trading would in gen- eral lead to better returns. Here the question we pose is whether varying the degree of human involvement and therefore reducing the degree to which behavioral factors can interfere in the decision-making process is beneficial. Still it needs to be noted that algo- rithms and AI methodologies are still written and programmed by humans, making them at least theoretically susceptible to some low level of biases. Dawes (1979) indicates that when it comes to the process of forecasting, algorithms dis- play a clear edge over their human counterparts. Continuing on that notion Promberger and Baron (2006) note that humans favor the input of other humans more strongly than that of an algorithm. Lastly Kirilenko and Lo (2013) write that the effects of behavioral factors displayed by humans affect the world of finance to an ever-greater extent than before. Therefore, using the BF framework as the second main theoretical framework for this study is more than relevant. 2.1 Efficient-market hypothesis The EMH is one of the foundational theories of finance that aims to describe how the financial markets operate. Fama (1970) explains market efficiency as simply a state in which the prices of securities fully reflect all of the information that is available for mar- ket participants. If this relation between market prices and information is constant, then they determine the market to display efficiency. The author also notes that there can be no transaction costs, the flow of information must be free and publicly available, and the market participants themselves must be in agreement of the significance of this infor- mation in relation to the market prices for such an efficient market to exist. 17 The EMH has its fair share of supporters and critics and one could easily argue that the necessary conditions for such efficiency described above are far from realism. Though it needs to be noted that even Fama (1970) states that it is enough for these conditions to be sufficiently met for one to be able to observe market efficiency. Wolff and Neugebauer (2019) are one of the many researchers to note the controversy surround- ing the EMH and in their research paper especially the predictability of stock returns is seen as proof for the lack of market efficiency. In addition, in their analysis they are able to find some evidence of this predictability and as such cite a multitude of other studies that come to the same conclusion. Timmermann and Granger (2004) take a different approach when studying the EMH. They note that while there may be predictability in stock returns this is not something that can be used as evidence against the EMH. They reason that EMH in its essence is only concerned with the absence of arbitrage opportunities and these stock price pre- dictabilities can very rarely be profitably exploited. Thus, taking a viewpoint where the markets are actually inherently efficient as presumed inefficiencies are impossible to take advantage of. As the debate for and against market efficiency can be observed from many different viewpoints, Fama (1970) devised different tests of market efficiency to provide evidence for his views. The first one being weak form tests that consider testing whether the his- torical and lagged prices of securities are able to provide an edge when it comes to the forecasting of future prices. He notes that these tests are also the ones most mentioned in the random walk literature as it would naturally imply that the stock prices follow a random and unpredictable path. Fama (1991) later notes that some predictability is still visible within the past prices of securities, but Sullivan et al. (1999) and Bossaerts and Hillion (1999) on the other hand display the inability to profit from using trading rules based on this finding. 18 Semi-form tests are concerned with testing whether the prices of securities adjust to publicly available information at a rapid rate. This on the other hand is something where one would assume great efficiencies due to the constant advances in technology. Gerlein et al. (2016) note the trading timeframes of high-frequency traders (HTFs) that go down to as far as nanoseconds. As such, one could assume that in today’s markets the trans- mission of information to prices is highly efficient. The strong-form tests are the strongest possible tests outlined by Fama (1970) to prove market efficiency and in these tests the levels of monopolistic and insider information that have affected past price changes are attempted to be measured. These tests are naturally the hardest to carry out due to the specification of insider information, but the author was still able to only observe limited evidence towards the rejection of the EMH. The weak- and semi-strong form tests against the EMH hypothesis are outright rejected in his study and he notes that prices adjust in an efficient manner when considering past price history and public fundamental information. For the purpose of this thesis these specifications and tests of market efficiency are es- pecially important as a lot of the trading strategies employed both by hedge funds and different algorithmic methods focus on these to generate returns. Caldwell (1995, p. 1- 5) notes that the primary strategy used by the first ever hedge fund was the long-short equity position along with additional leverage and Kooli and Stetsyuk (2020) uncover that this is also the most common investment style that is being used by hedge funds today. Fung and Hsieh (1999) broaden this information by detailing the extensive use of mechanical trading rules by hedge funds and Treleaven et al. (2013) show how these same rulesets are mostly dependent on financial and economic data. Returning to the thoughts presented at the beginning of the chapter, if hedge funds are able to charge fees that are substantially higher than for other investment types then their returns must be on par or otherwise as per logic no one would invest in such funds. Therefore, if these funds mainly function based on technical price data and fundamental 19 economic data one could assume that their sheer existence is against the weak- and semi-strong forms of market efficiency. Again, the EMH can theorize this further to maintain its relevance and it is especially well explained by Timmermann and Granger (2004). They note that while EMH can be quickly understood as a way to render all attempts at forecasting future prices to be a useless activity, in reality predictability can exist for short periods of time. They detail that this is due to the uniqueness of the investment ideas and as soon as they are discovered by a wider number of investors, the ability of these strategies to generate abnormal returns disappears. Sun et al. (2012) detail this same effect of the inverse economies of scale where they remark that only unique investment ideas can produce performance that can beat the market. Therefore, it can be seen that actually the existence of hedge funds, the ability to make profits by active trading and the ability to forecast prices, all factors used to undermine the EMH in various research papers, are not against the EMH. It is merely the longer- term persistence of being able to do these actions that renders them to violate the EMH. Timmermann and Granger (2004) outline that even asset price bubbles are not against the EMH as long as the risk premiums are indicative of the inherent dangers. As Hwang et al. (2017) note, hedge funds are absolute return vehicles designed to pro- vide returns irrespective of market cycles and conditions, hence the term hedge. This inherent market neutrality means that strategies play an even more important role for the return characteristics of the funds. Sun et al. (2012) detail further that strategies known to market participants stop working due to increased competition for the same pool of returns. The Preqin (2020, p. 20-21) Global Hedge Fund Report shows that asset inflows for the hedge fund industry are moving towards emerging managers and this strongly reflects the findings of Sun et al. (2012). They uncover that especially young and upcoming hedge 20 fund managers introduce new and innovative ideas and this uniqueness also leads to higher returns which is in term shown by the Preqin (2020, p. 20-21) report where emerging managers have been able to beat their more established counterparts from at least 2012 to 2019. All the findings discussed above lead us to the following reasoning. Price predictability is something that disappears over time and this is in line with the EMH. Hedge fund returns are not against the EMH weak- and semi-form tests but they cannot be persistent for this to hold true. Naturally one could argue that performance persistence cannot be achieved by continuously using the same trading strategy, but a fund manager that is always inno- vating and using different methods to find returns could arguably obtain persistent re- turns and still not violate the basic principles of EMH detailed by Timmermann and Granger (2004). Hence, it could be further argued that forecasting is not a meaningless activity and on the contrary a lot more forecasts would be required if each truly valuable forecast is only valid at a certain point in time. The findings of Sun et al. (2012) mirror this thought pro- cess as they note that continuous success requires continuously developing ideas that are both unique on their own, but also different enough compared to the ones employed by the other market participants. The main point of this thesis from the view of the EMH is therefore the following. Sun et al. (2012) point out that the development of new trading strategies is a very expensive process. Additionally, Timmermann and Granger (2004) note that most forecasters go through a multitude of different models to come up with their forecasts. If new and dy- namic, seemingly adaptive strategies that would work for the specific period where mar- ket efficiency has yet to diminish returns are being required, it wouldn’t be unreasonable to assume that AI can provide the answer. 21 Gerlein et al. (2016) especially go through the functionalities of machine learning (ML) which is marked, as the name suggests, by its ability to learn. They note the capabilities of such models for finding so called hidden forecasts that humans are unable to easily uncover and exploit. Therefore, one could imagine that a forecasting machine that is continuously adapting and changing similarly to the way efficient markets adapt and change as described by Timmermann and Granger (2004), it could theoretically be pos- sible to always be on the on the bleeding edge of forecasting models that work for the time being and provide abnormal returns. Algorithms and AI models themselves function mainly using financial and economic data as is shown by Treleaven et al. (2013). Therefore, a similar type of analysis can be carried out for these systems, especially as the above authors detail that algorithms mainly func- tion based on technical analysis. Technical analysis on the other hand is in its basic forms fully reliant on historical price data, meaning that the weak-form tests of EMH risk being violated if such performance would remain persistent. Wolff and Neugebauer (2019) detail further the difference between AI and ML as op- posed to plain rule-based algorithms mainly by their ability to learn. They show that these new approaches are able to learn without being given an explicit model and the authors estimate that such a flexible and adaptive approach might prove superior to simple rule-based systems. Similarly, to what has been discussed before, AI and ML have at least inherent potential to adapt and vary between multiple different forecasts for multiple different periods. When taking the EMH into consideration, these possibilities become very interest when considering performance persistence. 2.1.1 Active investing Active investing can be simply understood as the act of being an active participant within the markets and not simply following a passive buy and hold strategy where the under- lying market index or ETF is bought. Sharpe (1991) defines an active investor as one 22 holding a portfolio of stocks that differs from the market portfolio. He elaborates further by noting that an active investor is fundamentally acting based on presumed mispricing that they observe within the markets. As the thoughts and opinions on the true intrinsic value of securities might differ from day to day, active investors adjust their positions similarly by trading and hence being active. Hedge funds are inherently active as they function as absolute return investments. As a hedge fund aims to produce returns irrespective of the current state of the market it can already been seen that the definitions of being active by Sharpe (1991) are quickly met. Ammann and Moerth (2005) point out that the low correlations between hedge funds and other asset classes caused by this underlying investment philosophy is one notable reason why investors choose to invest in hedge funds in the first place. They function as diversifiers of risk when taken as part of a wider portfolio. Jensen (1978) on the other hand looks at active investing through the viewpoint of the EMH. He notes that if the markets are efficient as described by Fama (1970) then there are no possibilities for obtaining returns that are greater than the returns of the market. Rubinstein (2001) accordingly notes the inability of most fund managers in beating the market. The more recent findings by Timmermann and Granger (2004) that were dis- cussed in the previous section show that these views by Jensen (1978) are often not the case, but once again especially then longer-term persistence of this performance is the deciding factor. Sharpe (1991) continues by reasoning that an active investor cannot beat a passive man- ager after taking transaction costs into account. He argues that this is due to the many components that are needed for truly active investing, which involve expensive research and the development of costly trading strategies as was mentioned by Sun et al. (2012). He additionally details that a small and rare minority of outperforming managers does truly exist but to uncover the true advantage that active investing can give, the returns 23 of these funds need to be benchmarked against a comparable passive alternative. Hence, active investing is meaningless unless its passive counterpart is beaten. Timmermann and Granger (2004) add to the debate the potential short-term forecasta- bility of asset prices that can be seen as favoring the approach of passive investing. Sim- ilar to what Sharpe (1991) discussed, only brief advantages can be obtained as on the whole, overperformance in one period will turn into underperformance on the next when comparing active strategies against their passive correspondents. The authors also propose an interesting viewpoint for the debate between active versus passive investing as they note that if truly profitable active strategies are discovered by researchers, they likely wouldn’t be published in scientific journals. This in term leads to interesting implications where one could theorize that only unsuc- cessful active strategies get shared to the wider public, causing a larger than actual skew in results towards favoring passive investing. As was previously shown by Sun et al. (2012) hedge funds are very secretive and Treleaven et al. (2013) documented the same for the usage of trading algorithms. Additionally, Sun et al. (2012) analyzed the strong effect of competition towards the expected returns of different strategies and when taking into account the limited time window during which these strategies are able to provide ab- normal returns as was shown by Timmermann and Granger (2004), withholding profita- ble active strategies seems to be highly motivated. Ammann and Moerth (2005) detail this effect of overcrowding on a particular trading setup further by analyzing the size limits in terms of AUM set forth by some funds. Even on a fund level, certain trading setups experience diminishing returns if they a scaled up, an event described by the researchers to show the effects of limited capacity. Timmer- mann and Granger (2004) come to the same conclusion, noting how increasing position sizes from the viewpoint of one fund would increase both the transaction costs along with the actual market impact of the trade, rendering the actual opportunity impossible to take advantage of. Thus, it can be seen that for active strategies there are inherent 24 size limits and a common consensus amongst the researchers is this effect of diminishing returns of scale. When it comes to algorithmic trading, similar findings that were uncovered for hedge funds can be put forth. Algorithms and AI methods rely largely on technical analysis as was shown by Treleaven et al. (2013). Dash and Dash (2016) confirm this reliance and detail the constant need for historical data required by these algorithms. Treleaven et al. (2013) also note that acquiring the input data for these algorithms is highly expensive and Sun et al. (2012) mention the expensiveness of developing trading strategies. Lastly, Gerlein et al. (2016) uncover in more detail the computational resources needed for de- ploying these trading systems. If such a complex and costly system is put into place one can without a doubt assume that an asset manager would expect a return for this investment. An abnormal return to be more precise as the whole reason for carrying out costly research is to obtain market beating returns as was noted by Sharpe (1991). As such, algorithmic trading and various AI systems can be assumed to always represent active trading, and this can also be in- ferred from the literature surrounding these automated trading systems which revolve around testing the weak-form and semi-form hypotheses of the EMH. An observation that is also noted by Timmermann and Granger (2004). 2.1.2 Passive investing Passive investing is naturally an essential part of the EMH. As is noted by Jensen (1978) if the market is fully efficient, it is impossible to obtain abnormal returns as all the avail- able information is already incorporated within the prices of individual securities. Tim- mermann and Granger (2004) identify that in the strictest form of the EMH outlined by Fama (1970) all forecastability in asset prices would disappear and it would be impossi- ble to beat the market. 25 This is easily something highly disputed as we have seen in the previous sections. The active versus passive debate has therefore become an essential part of the EMH related literature. Sharpe (1991) details that fund managers are unable to beat the market on average and investing through the means of active methods is counterproductive. For an individual investor such arguments would naturally sound concerning and for fore- casters such inability to profitably predict asset prices becomes very concerning. French (2008) argues that a typical investor would obtain notably higher returns if he or she would switch to passive investing. This mirrors the views of Sharpe (1991) and of Fama (1970). Indeed, it can easily be theorized that the average investor is unable to beat the market, but questions remain as to what the effect becomes when observing the not so average investors, out of which hedge funds and their highly liberated arsenal of trading tools is a great example. French (2008) continues by aiming to uncover this relationship for hedge funds, coming to a conclusion that after taking into account the higher fees involved, on average these funds are unable to beat passive investing. He also notes that a passive investor enjoys the benefits of greater diversification by the sheer nature of holding the market portfolio. Lastly, he remarks that investors preference towards these active investment opportuni- ties is likely due to behavioral factors and lack of knowledge of better alternatives. The views expressed above show an interesting pattern where active investing is seen on the whole as driven by behavioral factors. Passive investing on the other hand is seen as the logical more profitable course of action, as after all the markets are presumed to follow the EMH. Kooli and Stetsyuk (2020) argue this view by uncovering that on average a hedge fund manager is actually able to beat the market, and this can be attributed to skill as opposed to luck. In their research they are also able to showcase that while some funds destroy investor value, the value added by the most profitable funds is more than enough to offset this balance. Most importantly the authors note that they find no evi- dence that this value that has been created is being shared to the investors. 26 Which leads us to the following analysis. For the context of the EMH, algorithmic trading systems and hedge funds can be grouped as one. They both represent active participants within the market and they both attempt to forecast and take advantage of different quickly disappearing opportunities, using a diverse set of strategies. Their performance is also against the EMH if it is persistent, but the same cannot be said if strategies are regularly changed. As we have shown forecastability does exist, but it is something that can evaporate quickly. Therefore, whilst passive investing is a natural favorite of EMH literature, it can be seen that the success of hedge funds or algorithmic systems is nothing that the theory cannot cope with. French (2008) outlines the challenges of this active approach by noting the zero sum or even negative sum game, where the profits of one investor mean losses for someone else. Sun et al. (2012) expose this relation further by detailing the need of unique investment ideas that are needed to beat both the other market participants along with the passive market-indices. Ultimately, it can be hypothesized that a limited number of hedge funds could be able to produce abnormal returns that favor active investing and their performance can be persistent as long as they remain continuously innovative. Which is a finding that brings us back to the topic of this thesis. AI can easily be seen as one of the most innovative and revolutionizing forces for a multitude of fields as is shown by Mullainathan and Spiess (2017). Employing it to remain innovative in regard to trading strategies for the markets seems like an obvious solution to the many issues related to active investing that are highlighted by the EMH. 2.2 Behavioral finance The analysis of the EMH showcased the debate between active and passive investing that is naturally very relevant for a research paper analyzing active trading. If algorithmic trading systems and hedge funds could be thought of as representing the same type of 27 investor in the viewpoint of the EMH, the BF framework provides the clear distinction between the two. This man versus machine setting is something where behavior un- doubtedly plays a role and the difference between a human trader and its algorithmic counterpart are much more diverse than pure EMH literature would lead one to believe. BF is seen by some researchers as an opposite view to the market efficiency hypothesis proposed by the EMH, whereas some other studies site it as an extension for the frame- works that are already in place. Ritter (2003) highlights the most notable differences be- tween BF and EMH by noting the rejection of the rational investor as proposed in the EMH. He details the bounded rationality that influences the decision-making process of investors as one in which different patterns of behavior and characteristics are too mean- ingful to be ignored in the way of the EMH. Markowitz (1952) for one notes, that the perceived utility is often defined over current gains and losses instead of focusing on the cumulative gains, hence showcasing the process of bounded rationality. Ritter (2003) especially highlights overconfidence which he sees as causing investors to weigh recent events to an exceeding extent. Gervais and Odean (2001) note that such traits can also be developed as an investor with a lot of recent success might feel very overconfident in their own abilities. Odean (1998) saw the link between overconfidence and excessive trading as was discussed before and Barber and Odean (2000) point out the reduced returns caused by this additional trading. Thus, the above serves as an obvious and easily understandable train of events where behavioral factors lead to actual and quantifiably reduced returns for an investor. Natu- rally overconfidence serves as only one example of psychological factors affecting inves- tors. Lord et al. (1979) for one note belief perseverance that leads to the inability of an investor to change his opinion once it is set. Buehler et al. (1994) and Weinstein (1980) add a systemic planning fallacy that showcases the over-optimism and wishful thinking of investors. 28 These various behavioral factors contribute to Barberis and Thaler (2003) remarking the need for change in the standard financial paradigm based on the EMH. They note that BF itself can be understood as a study on the limits to arbitrage and human psychology and Ritter (2003) comes to the same conclusion. Similarly, to the themes discussed for the EMH by Timmermann and Granger (2004), even if arbitrage opportunities would present themselves it would often be both risky and bring meaningless rewards if trans- action costs are taken into account. Therefore, pricing inefficiencies might persist but Timmermann and Granger (2004) on the other hand do not consider this as violation EMH if no profits can be obtained. De Long et al. (1990) detail also the risks involved with arbitrage as noise trader risk, where the perceived pricing inefficiencies first become worse, creating notable risks for arbitrageurs. As we are exploring the man versus machine aspect in our thesis, our attention turns solely to the human psychology aspects of BF. Mainly finding where both machines and humans prevail will help to uncover the primary motivation for the development for such trading systems. Barberis and Thaler (2003) are able to uncover interesting findings in their research paper that suit the analysis of hedge funds particularly well. They note that while there is a strong belief amongst people that experts make less mistakes, this added experience is something that might easily cause overconfidence for said individ- uals. Continuing on the topic, the researchers also note that even if advanced quantitative models are being used, overconfidence might present itself if there aren’t enough means to evaluate the accuracy of these models. In other words, especially the testing and feed- back environment for different types of trading algorithms is especially important. In re- gard to this, the authors also note that on their own people are in general not good at estimating probabilities and this on the other hand would put human managers at a dis- advantage as an algorithm would naturally be able to give a figure value, instead of a ballpark estimate. Interestingly, while they note that human traders exhibit all of the before mentioned characteristics, the authors also detail that hedge funds are actually 29 one of the main market participants trying to take advantage of these biases that other investors might show. Ritter (2003) gives a good outline of the main biases that humans exhibit. Heuristics are of particular importance and these can be understood as easily available rules of thumb, but as factors which easily lead to erroneous assumptions. Conservatism on the other hand can be especially harmful in trading as this makes individuals anchor to their beliefs even when the fundamentals around which their original thoughts were based on change. Similarly, the disposition effect makes investors vary of realizing losses, hence letting losses grow to a disproportionate level. The author also notes that especially hedge funds aim to profit from these behavioral traits. As hedge funds appear to seek returns by capitalizing on these psychological biases, one could also make the logical assumption that these funds themselves end up displaying some of the same factors. If for one a manager would be overconfident in their forecasts to take advantage of these types of investors and have a conservative stance towards changing opinions, a fund might rack up large losses in the process. From the perspective of trading algorithms, Treleaven et al. (2013) note the rule-based approach utilized. Similarly, to the rule-based trading strategies employed by human traders, algorithms use a similar if-else system where proven strategies are programmed into step-by-step actions that the trading algorithm can then execute. Wolff and Neugebauer (2019) further this to the usage of AI and ML, noting the lesser need for distinct rules and models, and instead emphasizing the more free approach where the machines are themselves able to learn and improve based on a certain feedback loop where good actions are rewarded and negatives ones discouraged. The authors think that especially this flexibility to adapt will lead to the great potential of these models both now and in the future. 30 Therefore, in terms of BF, interesting thoughts can be made. Static rule-based algorithms are emotionless execution machines of trading strategies but still the effects of behavior cannot be fully ruled out as the human programmer might still have used erroneous assumptions or similar factors that make them impacted by human psychology. Still from the most part trading algorithms can be thought of as rather immune in terms of the effects of behavior. AI on the other hand aims to mimic the human brain and the ability to learn will likely also make the machine learn different heuristics which are counterproductive. As op- posed to this, an AI model would also learn from this experience and no longer repeat the errors of the past which is something that cannot be said for humans as shown by Ritter (2003). Chincarini (2014) argues that trading algorithms and therefore additionally AI are able to fully eliminate behavioral errors and note that using these methods also enables funds to lower their trading costs. Dawes (1979) additionally writes that when it comes to the process of forecasting, algorithms prevail over their human counterparts. Ritter (2003) notes the hunt for misvaluations carried out by hedge funds, which implicitly details their use of forecasting models to find the correct asset prices. Promberger and Baron (2006) on the other hand note that people regard the opinions and input given by a human more strongly than that of an algorithm. Hence, the following course of action can be seen. Algorithms and AI are to be consid- ered practically immune from behavioral biases, with AI held a bit more highly in this regard as it doesn’t have to follow any specific rules programmed by a human. These systems make better and less erroneous predictions in terms of the BF framework and they are therefore able to prevail over their human counterparts. The recipients of these forecasts are still humans and they evaluate these forecasts through their own emotional processes and hold it at a lower value. 31 As humans are skeptical and often resistant to change, superior systems might still not get taken into use even if their performance is proven. While from a BF point of view trading algorithms are naturally perfect, especially Dietvorst et al. (2015) describe this phenomenon as algorithm aversion where these algorithms and AI are mistrusted no matter the proof. As we have seen, when observing the two opposite sides of the active trading spectrum, trading algorithms and human traders, it is especially the behavioral aspects that set them apart. Additionally, as advantages for algorithms one can also note the speed of execution, the capability to process information at a scale unimaginable for a human and the ability to work tirelessly day and night. Behavior sets us apart from machines and when it comes to trading this as we have seen can be considered a negative aspect. 2.2.1 Discretionary trading Discretionary trading is a trading style, which mainly involves the use of mechanical trad- ing rulesets as is shown by Fung and Hsieh (1999), but by the means of a human trader. In other words, a detailed trading strategy is constructed, and it is left up to the fund manager to ensure that this strategy gets executed correctly. Therefore, the discretion- ary approach to trading can be thought of as the early beginnings of hedge funds, where the possible assistance provided by computers was practically non-existent. Currently, discretionary trading involves the usage of technology to a great extent as is shown by Harvey et al. (2017) but the actual decision-making process is still carried out by humans. Therefore, from the viewpoint of BF, discretionary trading represents the human side of the man versus machine comparison. While the usage of the discretionary trading style is fairly similar to the rule-based methodologies employed by their mostly fully automated counterparts, systematic traders, it is in the analytical process where differences can firstly be observed. 32 Preqin (2021, p. 106-109) notes especially the reduced usage of models, as discretionary trading is more focused on the individual skillset of the trader. Treleaven et al. (2013) also detail that sometimes different analytical methods are used in terms of fundamen- tal analysis to forecast security prices, which involves using factors such as a firms’ bal- ance sheet data and macroeconomic variables to gain an understanding of underlying value. The authors also note the possible use of economic data and figures reported by central banks and government institutions with releases such as general unemployment and current interest rate, which can be considered natural as humans are more flexibly able to take advantage of a more various set of data. While the data used by discretionary traders can be seen as sometimes being different to the one commonly used by algorithms and AI, the main difference when being com- pared against systematic traders is the before mentioned execution process of trading strategies. Discretionary traders are therefore subject to all of the potential behavioral biases we have seen in this chapter so far and this is naturally something that would render them at a disadvantage. Still compared to plain trading algorithms discretionary managers would especially benefit from their ability to adapt, but when being compared against AI the advantages are less clear. Sun et al. (2012) note that funds using the discretionary trading style might benefit from the above flexibility as going after innovative ideas is easier. This is especially true in the case of small funds but something that can be seen as having some general implications for discretionary funds as a whole. Unique ideas for investing depend on the analysis process that has been done and the authors also note how time consuming this process is in terms of the potential profits. This is due to the findings discussed by Timmermann and Granger (2004) where the uniqueness of these ideas quickly disappears. Therefore, discretionary trading can be seen as somewhat less capital intensive to begin with as less is needed in terms of the technological infrastructure and no capital needs to be spent on developing complex trading algorithms. Still in the long run discretionary 33 traders continuously need to innovate and to do so with a much-reduced reliance on said technology. As we have seen, the performance of discretionary trading is heavily focused on the skills of the actual trader. As the dependence on the individual is great, so are the risks that the trader is exposed to in terms of behavioral biases. A human manager tends to be highly affected by a number of different biases and these might make the following of a well thought out trading strategy different when being implemented in the real world. The physiological limits on humans would also set their own limits on the execution of these strategies as it would likely be impossible to always be present and to take ad- vantage of every opportunity that would present itself. 2.2.2 Systematic trading As opposed to discretionary trading, systematic trading involves the extensive use of technologies and different types of trading algorithms to execute a trading strategy. Tre- leaven et al. (2013) note that most systematic traders aim to replicate and copy the step- by-step processes of successful traders and then obtain rewards through the perfect ex- ecution of these rulesets. While the difference between discretionary and systematic trading can be noted espe- cially in the execution process of trading strategies, also the process of generating these strategies is different. As is detailed by Treleaven et al. (2013) systematic trading mainly comprises of utilizing technical analysis to obtain trading signals and this involves the use of price data to uncover patterns and different trends to help forecast future directions of this price. While technical analysis plays an important role also for discretionary trading, systematic trading is additionally marked by the quantitative side of their investment processes, which involves the usage of different types of mathematical models to forecast and 34 predict future prices. Treleaven et al. (2013) detail this as involving the usage of similar financial and economic data used by discretionary traders, but by the means of models and not individual discretion. As such systematic trading involves potentially different types of data, different methods used to extract information from the data and different methods for the usage of this information to make actual trading decisions. Still it can be seen that the main difference between these two types of trading styles is the role of the human trader. In discretion- ary trading the human trader is very involved in the day-to-day processes, whereas in systematic trading the traders take more the role of an observer while the algorithms carry out the daily operations. Therefore, one can think of systematic trading as requiring more planning of long-term perspectives and less focus on the short-term fluctuations. Consequently, discretionary trading can be seen as representing the side of trading styles subject to behavioral biases and systematic trading showcasing the more methodical model focused automated approach. Discretionary traders are thus more easily at risk of different behavioral biases while systematic traders are by the nature of their trading style almost fully immune to the effects detailed by BF. Additionally, we have been able to observe some initial findings in terms of the usage of AI which seem to enable the best practices of both the different styles of trading. The emotionless of the systematic side and the ability to adapt of the discretionary style. In this chapter we have been able to divide the active traders outlined by EMH into two distinct categories separated using BF. While this framework allows us to maintain a clear distinction, still further analysis is needed into the more defined categories that exist between both the discretionary and systematic trading style. The utilization of both methodologies is completely possible and as we have seen AI is something that can from a behavioral point of view be seen as showcasing more human like traits, without human like biases. 35 3 Literature review In the previous chapter we have seen how both the EMH and the BF frameworks help us in understanding the key categorical differences that serve as the base for our further performance comparison. The analysis of the EMH helps us to see how active trading hedge funds stand out from the rest on a top level and the themes of BF help in creating the categorical differences between the distinct trading styles employed by hedge funds. Now our attention shifts to analyzing relevant literature within the field to further un- derstand what has been done and uncovered relating to our topics so far. This analysis of literature is split into first reviewing research papers that analyze hedge fund perfor- mance using different methods and methodologies to give a more general overview of what is the consensus on the performance of these funds. Additionally, the focus will be to present some initial motivation into the choices made further along during the actual performance comparison between our sample of funds. The second part of the literature review will be focused on reviewing studies relating to algorithmic trading and AI as this is a theme of special importance for the topic of this thesis and needs to be further reviewed in more detail. Here the focus will be especially on the types of advantages and disadvantages that these models are able to bring along with the type of performance that can be expected when they are being used in real live- market environments. The final part of our literature review will focus on the research papers closest to the topic of this thesis, the past analyses between discretionary and systematic funds. While systematic funds can be seen as less advanced than pure AI funds in some respects, these studies help in showing what findings have been made so far when similar man versus machine setups have been utilized by other researchers. 36 3.1 Hedge fund performance There are several studies relating to hedge fund performance and the research carried out by Capocci and Hübner (2004) serves as a good starting point for this review. In their research paper, the authors first detail several findings for the hedge fund industry as a whole noting the concentration of hedge funds within the U.S. along with the greatly varying fund sizes measured in terms of AUM. The key figures being that 90% of manag- ers operate from the U.S. and that over 80% of hedge funds have AUM figures of under 100 million. The industry is marked by high fees and high minimum investment amounts and the access to funds is limited to only accredited investors. In their analysis of hedge fund performance, the authors especially note that based on their factor models hedge fund returns show a positive coefficient towards the Fama and French three-factor model size factor, meaning that funds generally invest in small stocks. They also note that while performance persistence might be disputed, when measuring sheer performance, 27% of the funds in their sample display statistically significant ex- cess returns. The authors also detail the adjusted coefficient of determinations that they are able to obtain by using their factor models, noting values of 0,44 for the single factor capital asset pricing model (CAPM) and 0,60 for the Carhart four-factor model. Performance persistence is measured in part by employing a subperiod analysis which shows that hedge funds on the long-term are able to deliver great returns but the same cannot be said for the short term where returns are notably more varied. Ammann and Moerth (2005) on the other hand investigate the impact of fund size to returns and note interesting findings in terms of the negative relation between increas- ing inflows and diminishing returns. This is then further detailed by Lim et al. (2016) who are able to display this effect of investors chasing past returns in terms of their asset allocation decisions. Hence, one could hypothesize that investors chasing returns make 37 funds unable to take full advantage of their preferred strategic opportunities due to the impact that this increased size brings to the markets. The authors are able to discover and prove the same causality by noting the reduced ability of larger funds in being able to take advantage of trading strategies that exhibit fundamental capacity constraints. In their findings they are able to discover that while small funds do not need to take these capacity constraints into account simply due to their size, they struggle as a result of the higher fees and expenses that they have to endure as they cannot take advantage of certain economies of scale that are available to larger funds. Larger funds are also noted as being in a more dominant position as they are more easily able to control the assets that they manage by imposing various withdrawal conditions upon their investors. This in term creates possibilities according to the researchers as having a stable asset base also enables investing in less liquid types of financial assets in search of returns. In their research paper they are also able to discover that while this is the case, smaller funds are able to have more flexibility in terms of their potential trading strategies, they are able to take on additional risks and they are able to focus more on specific ideas and innovations to further their returns. Larger funds are able to attract capital more easily due to their proven track record, but this size might also make these funds take on a more defensive stance towards investing. From a more systemic risk point of view an interesting finding is the fact that smaller funds are more quickly able to react to differ- ent types of events as their portfolios are in general more liquid due simply to the re- duced size of their positions. Still in their final results the researchers are able to find a positive relationship between the size of the hedge fund in terms of AUM and the per- formance that the fund is able to obtain. 38 Lastly, the authors note that larger funds display lower volatilities and higher returns which in term allows them to have higher Sharpe ratios. One interesting dilemma noted is the agency problem related to the size of a fund. As the manager is compensated based on a proportion of the AUM, one might be inclined to grow their asset base un- controllably to earn more for themselves while maintaining the same strategy. Therefore, the need for a balance between manager revenues and fund performance is noted. Contrary to their findings Berk and Green (2004) note that as investor flows chase past returns, these opportunities disappear due to increased competition and fund growth and hence an opposite economies of scale effect is noted. Herzberg and Mozes (2003) are able to discover that small hedge funds obtain better performance in general but that especially their risk adjusted returns are of more relevant significance. Edwards and Caglayan (2001) are able to find that as hedge funds grow, their perfor- mance also increases but this ratio declines rapidly. Gregoriou and Rouah (2002) on the other hand find no meaningful connection between the size of a fund and the returns that it is able to obtain. Sadka (2010) takes a different stand to comparing performance amongst hedge funds as he notes that most of the variation between the returns of in- dividual funds are actually being driven by liquidity risk, where funds holding illiquid se- curities take on more risk but earn a premium over other funds. When it comes to performance especially the persistence of this performance is of im- portance as can be reasoned from both the viewpoint of a fund manager and that of a prospective investor. In regard to this, the research paper by Capocci and Hübner (2004) also details its importance due to the dynamic nature of hedge fund investors. The attri- tion rates for the industry are notably more significant than those seen within mutual funds and as such persistence in performance takes on an even more important role. Agarwal and Naik (2000) for one are able to find such persistence in the performance figures of the hedge funds in their sample. 39 Liang (1999) also makes interesting findings in terms of performance, noting that hedge funds are on average able to outperform mutual funds but the same cannot be said when the performance is compared against the returns of appropriate market indices. Also, the characteristics of this performance are detailed as the author notes the higher volatility that hedge fund returns are subject to when being compared against either mutual funds or market indices. Lastly, the impact of fund characteristics upon the de- gree of performance are also detailed, with fund age and the degree of leverage em- ployed being seen as meaningful. As we saw in the analysis of EMH, overperformance is highly disputed and Carhart (1997) for one attributes most of it down to random factors as far as the average returns of funds are concerned. Opposed to this, Kosowski et al. (2006) find in their research paper that at least mutual funds are able to exhibit alphas net of fees that are both large and too persistent to be caused by luck. Kooli and Stetsyuk (2020) continue on this topic from the view of the hedge funds as they measure the skill shown by hedge fund managers by researching the value that they are able to add. In their research paper they come to the conclusion that hedge fund managers are on average skilled but more interestingly, they note that it appears that the revenues attributable to this skill are not being shared to the investors of these funds due to the high fees involved. They further conclude that after the returns are taken net of fees the amount of funds that are able to deliver abnormal performance is notably reduced. From an industry wide perspective an especially relevant finding is also that the most successful hedge fund managers are clearly able to offset the losses incurred by the worst performing funds, thus making the average of managers show clear skill in terms of overperformance. Lastly, they note that size has an impact on the variation of returns amongst funds and hedge funds in particular seem to benefit from the reduced regulation that they face. 40 Also, in terms of managerial performance, the high fees and therefore high compensa- tions that the managers are able to obtain are noted as important incentives behind this outperformance. Agarwal et al. (2018) analyze performance by splitting returns into parts explained by traditional factors and parts unexplained which they describe as exotic risk. This is done to uncover the uniqueness of trading strategies that hedge funds are able to pursue with the lesser regulatory frameworks that they are under. They note the addition to the lit- erature that they are able to bring by not only interpreting the portion of return unex- plained by traditional factors as alpha but by also uncovering the factors that this excess return is attributable against. In their research they note that while some investors do not pay specific attention to the risk factors a fund is exhibiting, certain investors are actively seeking them as they look for funds employing specific strategies. Conversely to the findings by other research pa- pers noted before, the authors do not find performance persistence in their sample of funds. One main finding they are able to produce is the fact that investors seem to put more emphasis on these before mentioned exotic risk exposures of hedge funds as they note that these serve as the main reasons for an investor choosing to invest in hedge funds in the first place. Exposure to traditional factors is available through mutual funds and the high costs of investing in hedge funds do not justify investing purely based on returns attributable to these factors. The authors are also able to uncover that investors use the alpha value obtained through the CAPM to evaluate and rank funds. Hence, investors seem to exhibit a preference towards market beating returns. As such especially the CAPM is noted as explaining fund asset flows and the authors also note evidence of abnormal returns being eliminated due to increased inflows of capital. 41 Kacperczyk et al. (2014) define skill as either an inherent ability to pick winners or to time the market and in their research the authors are able to show the hedge funds are able to obtain substantial outperformance compared to their mutual fund peers of pas- sive benchmarks. Contrary to some of these findings Ackermann et al. (1999) on the other hand do not find evidence that hedge funds on average would outperform the S&P500 stock index and they also note some of the findings seen before where hedge fund returns are attributable to characteristics of individual funds. Bali et al. (2013) also find that hedge funds are unable to outperform the S&P500 and Stulz (2007) proposes an interesting hypothesis where he notes that the performance of hedge funds will con- verge towards the performance exhibited by mutual funds in the long run. As the final paper on hedge fund performance used for the literature review part of this thesis, the research paper by Hwang et al. (2017) is focused on studying the relationship between systemic risk and hedge fund returns. When researching the risk profile dis- played by hedge funds, the authors were able to find that there is a positive and statis- tically significant relationship between the level of systemic risk that a fund is exposed to and the level of returns that the fund is able to attain. In other words, funds investing in high beta stocks earn better rewards for this added risk-taking. As such, they note negative returns during periods of market downturns, but this is to be expected as the high beta portfolios of these funds amplify the movements of the market. The authors similarly note that the added returns are due to the added exposure that these funds are risking in different negative systemic events. They are also able to show that the positive relation between this level of systemic risk and better perfor- mance also holds after taking into account different firm specific characteristics. Billio et al. (2012) add to this by detailing that when negative developments take place, small funds are more affected by the spillover effects of these systemic risks and Boyson et al. (2010) note the contagion experienced by hedge funds in times of crises. Acharya et al. (2017) interestingly note that large hedge funds can grow to sizes where they themselves serve as a source of systemic risk. 42 Lastly, the authors note that as hedge funds benefit from taking on added systemic risk in terms of risk premiums, these practices are likely to continue but they also detail the effects that various crises have had on hedge funds, both in terms of AUM, returns and number of funds. Thus, coming to a conclusion that while these practices entail clear risks, the profits are also distinctive and as such justify these risks for most funds. As we have now seen, the performance analysis of hedge funds has been done using various different methods and comparisons in past literature. While the returns are com- pared against indices, with the S&P500 being the most popular, the performance of hedge funds is also often compared against that of mutual funds. The evaluation of dif- ferent risk exposures is also present in various studies as is the analysis of performance persistence which is deemed as especially relevant. Finally, the analysis of hedge fund performance using different fund characteristics and styles remains the most common method of performance evaluation in the literature we have selected, and it seems that especially the comparisons amongst hedge funds are deemed relevant in the research within the field. 3.2 Algorithmic trading and AI In the following section our attention turns to the literature analysis of research papers published on both algorithmic trading and the usage of AI for trading purposes. Starting with the research paper by Paiva et al. (2019) we are able to observe some findings re- lating to the forecasting ability of AI. The authors detail the inherent complexity of the process of price and return forecasting for the stock market which is caused by the na- ture of the market itself. It is especially the dynamics of market prices and the large amount of the so-called noise within those prices that makes it difficult to detect what factors are truly meaningful for the process of forecasting. Additionally, the market is impacted by various external factors on a continuous basis, making the stock market an incredibly complex playground for different types of models. 43 The authors detail this further by breaking down the process of forecasting as one in- volving only the linkage between the past and the present. Especially interesting is their discussion of the two main methods used both in the literature and on the field for this process. The first one being different econometric models based on statistics and imple- mented using trading algorithms and the second one being the usage of advanced ML models which are then implemented by the means of AI. The trading algorithms are defined as ones using tools such as linear regression and GARCH-modelling. Whereas AI algorithms are noted as using artificial neural networks, random forests, support vector machines and other similar frameworks for their process of generating forecasts. Also, the level of flexibility of AI models is detailed as the authors describe the ability of these models to utilize both quantitative and qualitative sources of data. While trading algorithms need to rely mostly on financial time series data, AI models are able to function with a much more flexible and diverse dataset and work with data that is imperfect. In their research paper, AI models are based on technical analysis, meaning that they are functioning based on return data for individual securities. The authors note that when working on the same data, AI models as opposed to trading algorithms are able to find complex patterns and so-called hidden meaning behind the data, which refers to com- plex relationships and causalities that would otherwise be impossible to detect. Also, the main reasoning behind using technical analysis is detailed as they note that this revolves around the belief that past patterns in prices repeat themselves and hence, that prices do not follow a random walk process. In their final findings, the authors are able to show that their AI model is able to generate meaningful and significant returns, but the authors also note the great impact that trans- action costs have on the profitability that their model is able to obtain. When more re- alistic assumptions are taken into play and transaction costs are accounted for, the model struggles to make a profit. 44 Dash and Dash (2016) on the other hand detail a lot of background information regarding the usage of algorithmic trading and AI. They especially note the increasing relevance of the topic as data is currently more available than ever before and this makes it possible to develop highly advanced models. Similarly, to the paper by Paiva et al. (2019), the authors describe the forecasting based on financial time series data as a very difficult process, noting the different trends, variations and irregularities within the data. While the difficulty of understanding this data is being understood, they also deem AI models as best suited for this purpose due to their high level of automation, speed and flexibility for going through these very large datasets and finding hidden meaning. The process of data mining is therefore also mentioned, and this is described as simply in- volving the extraction of meaningful statistics from big data as detailed by Witten et al. (2011, p. 191-202). AI models are seen as both tools for automating and as advantages for decision making. As more information is available to investors using AI, Dash and Dash (2016) note that this will likely also enable a reduction in the level of risk that the investor needs to take in order to obtain a profit. When it comes to literature in the field, the authors note the common trend of using technical analysis to create different types of indicators, which are then used to develop trading signals and strategies for or by the models to generate returns. The authors also detail the use of supervised AI models, which entail the training of these models using a set of inputs along with a set of desired outputs. The process of trading is trained as a simple classification task, where the buy, sell and hold decisions simply represent a set of outputs based on some set of inputs. In their final conclusions, the authors do not deem the sole use of technical analysis as being sufficient as they note the need for the usage of different types of big data analysis to further the probability of their AI forecasts. 45 White (2000) discusses some pitfalls of the datamining approach as he notes that some perceived results are only caused by luck instead of real forecasting ability. Gerlein et al. (2016) on the other hand note the unbiasedness of AI model creation guaranteed by the splitting of the data into so called training sets and testing sets. With this approach the AI model is first trained on the training set and later the actual forecasting ability of the model is validated by applying it on the testing set that it has not been exposed to before. Continuing on the research paper by Gerlein et al. (2016) focused on the creation of profitable ML algorithms, the authors similarly note that AI is well suited for the process of finding hidden relationships within data and consequently having a strong capability towards forecasting. It is also noted that most research papers and experts in the field train their AI models based on different types of variables, attributes and indicators that have been processed from the financial time series data, instead of using this raw data on its own. The authors note that while the usage of AI models seem to imply better forecasts, this does not always translate to higher profits. In this regard, they note that different models must always be evaluated based on their actual performance and not solely based on the accuracy of their predictions, as this does not always reflect well when being applied to actual live markets. Especially higher volatility situations are seen as troublesome for AI models due to the fact that this renders the generalization of forecasts and finding causalities increasingly difficult. It is noted that traders should incorporate the results of multiple AI models as a weighted average to obtain meaningful results. Matias and Reboredo (2012) further this discussion of AI models by noting their ad- vantages in solving different types of problems by using nonlinear data. Ballings et al. (2015) additionally detail that various factors influence the stock market, and this results in highly nonlinear price data for the market as a whole. Hence, it can be seen that this nonlinearity an important aspect to consider as far as predictions are concerned and again the suitability of AI models for the purpose of price predictions is displayed. 46 Mullainathan and Spiess (2017) uncover more how AI models function and also discuss what types of developments have led to their creation. Firstly, they discuss both the con- tribution created by advances in computing and the findings made in the field of statis- tics. The way these models function is simply explained by means of comparison against standard algorithms which need distinct rulesets to carry out their tasks. AI models on the other hand are given an input and an output and the models are tasked with finding an underlying function that best predicts an output based on a set of inputs. As such, AI and ML have a lot more freedom in finding different solutions and therefore also the results might bring additional findings that were not originally considered. The authors especially note that AI is not to be used only to solve old problems using new ways, but to solve completely new problems, before thought too difficult to tackle. Due to the noted ability of AI in finding hidden patterns, one can also hypothesize that it would be very well suited in the field of technical analysis where the uncovering of dif- ferent patterns within the return data are used as trading signals. The authors note this ability of AI as these models are able to discover patterns without the need for specifying them in advance. In regard to typical regressions, AI models are noted as finding optimal models of best fit especially in nonlinear datasets. Also, the before mentioned flexibility of AI is noted in the research paper and these models are explained simply as tools to extract substance from big data. Finally, they note that AI models are allowed to choose the models and rules that work best for the data and no such rules are specifically programmed. Therefore, these models find meaning based on the data itself and not on the presumptions of the programmer, making them likely less to be biased and better suited for the task at hand. Antweiler and Frank (2004) for one create a ML algorithm to go through online posts as a way to forecast and explain stock market volatility and they are able to obtain a statis- tically significant small positive performance. Hendry and Clements (2004) additionally 47 make interesting findings noting that combining multiple forecasts from multiple differ- ent models creates more accurate results than simply relying on one model, a view which is shared by Bates and Granger (1969). Wolff and Neugebauer (2019) set out to study how well different types of ML models are suited for stock return predictions, noting the wide use and acceptance of these models in other fields such as facial recognition. They also define AI models by their ability to learn as opposed to static rule-based algorithms and similarly to past studies they also note how well the models are suited for nonlinear datasets. Interestingly, the authors are still unable to find significant outperformance by these models as opposed to more advanced types of linear regressions. When it comes to the process of stock return predicting clear outperformance is still noted against a buy-and-hold strategy. In the research paper the training of the model is seen as of particular importance along with the usage of new data to test the model on to obtain unbiased results. It is also seen that the models need a large amount of data during the training phase to obtain decent forecasts in live environments. One key observation is the ability of the program- mer to control certain tuning parameters for an AI model which in term determine the degree of fit that the model will aim for. While a model can be almost perfectly fit to the training data, this in term results in poor performance when the same model is exposed to out-of-sample data due to overfitting. This is also noted to be of concern when dealing with stock market predictions, as there is such a large amount of noise within the data. Also, different complexities of AI models are examined along with their pros and cons, and the problem noted with very complex models is the large amount of training data that they need. Conversely, these complex models are also described to be specially well suited to model complex relationship as they have inherent flexibility. The authors detail that this need for data becomes a problem when using solely financial time series data due to the noise and changing factors that drive returns over time. Therefore, older re- turn data is significantly less relevant. 48 Lastly, it is noted that while ML and AI technologies are beginning to be more widely used within the asset management industry, the low signal-to-noise ratio of stock return data makes advanced linear models the preferred option. Still they note that AI models are better when the number of potential predictors for forecasting is very large within the dataset in question. Treleaven et al. (2013) examine a lot of descriptive information on the usage of algorith- mic trading and similarly to the hedge fund industry, the secretive nature of the field is uncovered. In their study algorithmic trading simply refers to the usage of algorithms to automate either any part or the entirety of the trading process. In terms of hedge funds using algorithmic trading the real-life implementation process of this trading style is also detailed, with pre-trade analysis, signal generation, trade execution, post-trade analysis, risk management and asset allocation noted as key steps by the authors. For the process of both creating and improving models, especially backtesting and dif- ferent simulations based on historical data are seen as relevant. Additionally, the risk of employing these systems is noted, with possible programming errors resulting in unex- pected behaviors and great potential losses. Some of the main challenges for both the implementation and literature within the field are noted as being the lack of understand- ing of the interactions that these algorithms have amongst each other and the widely varying behaviors that these systems exhibit if certain variables are changed. Khandani and Lo (2011) analyze situations where different trading algorithms used by systematic traders are seen as exhibiting a high degree of correlation amongst each other. The authors note that similar factors are use