Académique Documents
Professionnel Documents
Culture Documents
Thesis Advisor
Dr. Steve Phelps
Declaration
I hereby declare that this thesis is my own work, and has not been submitted for
the award of a higher degree elsewhere. Where other sources of information have
been used, they have been acknowledged.
The work in this thesis has contributed to the following publications:
Sbruzzi, E. and S. Phelps: 2011. Optimal Level of Leverage using Numerical
Methods. In, 7th Conference of the Portuguese Finance Network, Aveiro,
Portugal.
Sbruzzi, E. and S. Phelps: 2013. Testing leverage-based trading strategies
under an adaptive-expectations agent-based model. In, 12th International
Conference on Autonomous Agents and Multiagent Systems (AAMAS 2013).
Abstract
Leverage offers the possibility of enhancing financial returns and, consequently, the
profit and the end of period wealth. Leverage is gaining importance and has been
widely adopted in the financial markets for two reasons. Firstly, brokers are interested in offering margins because they can charge higher transaction fees and
make profits from lending margins. Secondly, investors are also interested in taking
leverage because of the ability to enhance their individual returns. The motivation
of this thesis is that, even though leverage is gaining importance in modern investments, existing models in the literature models assume that the series of financial
returns are normally distributed. However, financial returns present high-level of
kurtosis and, hence, are not normally distributed. Thus, existing analytical models
underestimate extreme returns and consequently underestimate the risk of default.
I contribute to this field by proposing a new trading strategy that uses numerical
methods to calculate the optimal level of leverage instead of the existing analytical
models. The use of numerical methods allows me to relax the assumption of normally distributed returns, and hence minimises the risk of underestimate extreme
returns and the risk of default. I investigate whether the use of numerical methods leads to a more accurate optimal level of leverage than analytical models, and
if the use of the optimal level of leverage using numerical methods improves the
investment performance. In order to test the ability of the optimal level of leverage using numerical methods to improve the investment performance, I employ two
different approaches: back-testing and agent-based modelling. Back-testing allows
me to test the optimal level of leverage using numerical methods using empirical
evidence, and agent-based modelling allows me to test the optimal level of leverage
using numerical methods in a totally controlled environment. The conclusions are
that the use of numerical methods leads to a more accurate optimal level of leverage
than analytical models; using daily historical data as an empirical evidence, the optimal level of leverage using numerical methods improves investment performance;
and in a totally controlled environment, the ability of the optimal level of leverage
to improve investment performance depends on the size of the market.
Acknowledgments
Writing a thesis and getting a PhD is a life-changing opportunity. It requires the full
dedication of the candidate and the support of every one who is part of his life. For
this reason, I would like to express my gratitude to a number of people. Without
their help, support and patience, certainly this work could not have been done.
Firstly, I would like to thank the group of the most important people of my life,
my family; my parents, Mr. Arnaldo Felipe Sbruzzi and Mrs. Luciana de Souza
Sbruzzi; my wife, Mrs. Ana Paula Ribeiro Duarte Sbruzzi; and my son, the reason of
my life, who was born during the PhD period, Lus Felipe Duarte Sbruzzi. Without
them, nothing here would make much sense and I would not be able to reach any of
my objectives. They are unique for me and their love and support makes me strong
to reach my dreams.
Secondly, I would like to thank my supervisor Dr. Steve Phelps. His patience
with me was impressive. He provided me with an appropriate and supportive environment that allowed me to go further in my research. He offered me the chance
to express my ideas without fear of reprisal. He helped me to limit the scope of my
research and was receptive to correct the natural mistakes that I have made during
the research.
Thirdly, I would like to thank Mr. Davi Baccan. Mr. Baccan is a closed friend
and he has been pursuing a PhD in Computer Science at the University of Coimbra,
Portugal. Mr. Baccan has provided me with every necessary informal support in
many area of the research.
Fourthly, I would like to thank Cynthia Rocha. With her support, I could review
Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 1: Introduction . . . . . . . . . .
1.1 Motivation . . . . . . . . . . . . . . .
1.2 Research Objectives . . . . . . . . . .
1.3 Summary of Research Contributions
1.4 Overview and Structure . . . . . . .
1.5 Publications . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 3
. 4
. 7
. 8
. 9
. 11
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
13
17
20
23
23
25
28
29
30
32
33
38
Mean on Leverage
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
41
42
42
43
44
46
52
53
.
.
.
.
.
.
.
.
.
.
55
56
58
59
60
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4.4
4.5
4.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Chapter 5: A Back-testing
Numerical Methods . . .
5.1 Introduction . . . . . .
5.2 Experiment . . . . . .
5.2.1 Methodology .
5.2.2 Single Iteration
5.2.3 Results . . . . .
5.3 Discussion . . . . . . .
5.4 Summary . . . . . . .
of
. .
. .
. .
. .
. .
. .
. .
. .
Leverage using
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
Leverage
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
71
72
73
73
76
80
85
86
Model using
. . . . . . . . 89
. . . . . . . . 90
. . . . . . . . 93
. . . . . . . . 95
. . . . . . . . 96
. . . . . . . . 96
. . . . . . . . 97
. . . . . . . . 104
. . . . . . . . 105
List of Figures
3.1
The flow-chart of the sampling process of the proxy and the real
geometric mean. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 The series of prices of UNH between 17/01/2001 and 04/08/2011. .
3.3 The series of returns of UNH between 17/01/2001 and 04/08/2011.
3.4 The cumulative equity returns of leveraged and unleveraged investment in of UNH between 17/01/2001 and 04/08/2011. . . . . . . .
3.5 Plot of the series of the proxy and the real geometric mean of the
unleveraged investments. . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Plot of the difference between the proxy and the real geometric mean
of the unleveraged investments. . . . . . . . . . . . . . . . . . . . .
3.7 The boxplot of the proxy of the geometric mean and the real geometric mean of the returns of unleveraged investments. . . . . . . . . .
3.8 Plot of the series of the proxy and the real geometric mean of the
leveraged investments. . . . . . . . . . . . . . . . . . . . . . . . . .
3.9 Plot of the difference of the series of the proxy and the real geometric
mean of the leveraged investments. . . . . . . . . . . . . . . . . . .
3.10 The boxplot of the proxy of the geometric mean and the real geometric mean of the returns of leveraged investments. . . . . . . . . . . .
4.1
4.3
4.2
4.4
4.5
4.6
5.1
5.2
5.3
. 43
. 45
. 46
. 47
. 49
. 49
. 50
. 50
. 51
. 51
. 62
. 63
. 63
. 65
. 66
. 67
List of Figures
5.4
5.5
5.6
5.7
6.1
6.2
6.3
6.4
6.5
6.6
79
80
81
84
. 99
. 99
. 101
. 101
. 103
. 103
List of Tables
3.1
Statistical analysis of the series of the proxy and the real geometric
mean of the unleveraged investments. . . . . . . . . . . . . . . . . . . 48
3.2
3.3
Statistical analysis of the series of the proxy and the real geometric
mean of the leveraged investments. . . . . . . . . . . . . . . . . . . . 52
3.4
4.1
4.2
Comparison between geometric Brownian motion approach and numerical methods of the geometric mean using the optimal level of
leverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3
4.4
Comparison between geometric Brownian motion approach and numerical methods of the optimal level of leverage. . . . . . . . . . . . . 64
5.1
Statistical analysis of the log of the end of period equity return of unleveraged ln(EoP (1)), geoemtric Brownian motion leverage ln(EoP (lgbm )),
optimally leveraged ln(EoP (l )) and excessively leveraged investment
ln(EoP (10)). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.2
T-Test of the difference between the log of the end of period equity
return of optimally leveraged investment and geotmric Brownian motion leveraged investment (GM (l ) = GM (lgbm )); the difference between the log of the end of period equity return of optimally leveraged investment and unleveraged investment (GM (l ) = GM (1));
and the difference between the log of the end of period equity return
of optimally leveraged investment and excessive leveraged investment
(GM (l ) = GM (1)). . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.1
List of Tables
6.2
6.3
6.4
6.5
6.6
T-test results between the log of the end of period return of the
optimally leveraged agent (Opt) and the unleveraged agent (Unl);
and the optimally leveraged agent and the excessively leveraged agent
(Exc). N = 10 agents. . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Statistical analysis of the log of the ends of period returns, N = 10o
agents. Opt = optimally leveraged, Unl = unleveraged and Exc =
excessively leveraged. . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
T-test results between the log of the end of period return of optimally
leveraged agent (Opt) and unleveraged agent (Unl) and optimally
leveraged agent and excessively leveraged agent (Exc). N = 100 agents.100
Statistical analysis of the log of the ends of period returns, N = 1000
agents. Opt = optimally leveraged, Unl = unleveraged and Exc =
excessively leveraged. . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
T-test results between the log of the end of period return of optimally
leveraged agent (Opt) and unleveraged agent (Unl) and optimally
leveraged agent and excessively leveraged agent (Exc). N = 1000
agents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Chapter 1
Introduction
This thesis uses numerical methods to calculate the optimal level of leverage and
tests the ability of the optimal level of leverage using numerical methods to improve
individuals investment performance. I assume that investment performance is the
end of period equity and equity is the amount of capital of the agent. For example,
if an investor with an initial equity of $1,000 opens a long position of 100 shares of
security A, which its price per share is $100, the leverage is 100 x $100 - $1,000 =
$9,000 and the level of leverage is 100 x $100 / $1,000 = 10.
This exposure expands the investors equity returns and can have a significant
impact of stock returns on equity returns. For example, suppose there is a scenario in
which two period returns of an investment in a single stock with a value of $100 and
with a stock returns of 8% and -10%. We also assume that the capital is reinvested
in the end of each period. If the level of leverage is 1, this means that the leverage is
0, and the equity increases to $108 in the first period, and, in the second period, the
equity decreases -10% to $97.20. Consequently, the cumulative equity return would
be -2.8%. However, if the level of leverage is 10, the equity increases to $180 [100*(1
+ 8%*10)] in the first period, and, in the second period, the equity decreases to $0
[180 * (1 - 10% * 10]. Consequently, the cumulative equity return would be -100%.
Hence, what dictates the equity return is the combination of stock return and the
level of leverage. Thus, in order to improve the individual investment performance,
Chapter 1. Introduction
the appropriate use of leverage is as crucial as the appropriate selection of the stock.
In this chapter, I present the introduction of this thesis. This chapter is organized
as follows. In Section 1.1, I present the motivation of this thesis. In Section 1.2,
I present the research objectives. In Section 1.3, I describe the contributions of
this thesis. In Section 1.4, I present the outlines the general structure. Finally, in
Section 1.5, I present the publications originated from this thesis.
1.1
Motivation
In this section, I present the motivation of this thesis. Nowadays, leverage plays a
very important role in the financial markets. New instruments designed to facilitate
leveraged investments such as Contract for Difference (CFD) in conjunction with
technological advances in the electronic markets make leverage accessible for every
type of investor, from small individual investors up to big pension funds. CFD
is an informal contract between a broker and his client (Norman 2009). Using
CFD, broker controls the level of exposure of the client and is more interested in
allowing clients to take leverage than using stocks. For example, in 2007, according
to UK Financial Service Authority (FSA), 30% of the volume on the London Stock
Exchange was driven by leveraged investments (FSA 2007).
However, to the best of my knowledge, there are relatively few studies relating
leverage and Computational Finance. The existing papers analyse the implication
of leverage on the financial market as a whole (Geanakoplos 2009; Thurner, Farmer,
and Geanakoplos 2011). However, no previous research has analysed the implication
of leverage to an individuals investment performance.
Computational Finance provides tools to experiment and to simulate hypotheses
regarding different aspects of finance. There is a considerable amount of research
on trading systems using Computational Finance tools (Lo and MacKinlay 1990;
LeBaron 2000; LeBaron 2002; LeBaron 2006; Martinez-Jaramillo and Tsang 2009;
Iori and Chiarella 2002), but such research has not taken into account leverage.
1.1. Motivation
Studies considering leverage and individual investment performance in Computational Finance are relevant for three reasons.
Firstly, recent studies about leverage demonstrate that an excessive level of leverage is considered irrational behaviour (Geanakoplos 2009; Thurner, Farmer, and
Geanakoplos 2011). When agents obtain positive equity returns, they can be frustrated if the level of leverage is low because such returns could be higher if the
agent was more leveraged. This belief encourages the agent to increase their level
of leverage in the next period. However, this behaviour is based on a naive belief
rather than rational reasoning because there is no model to indicate whether the
increment of the level of leverage is appropriate.
Secondly, the equity return formula has two components: the stock return and
the level of leverage. Additionally, the level of leverage is linear in the equity return
formula (Peters 2010). This means that leverage has the ability to modify the
impact of the stock return on the equity return and, consequently, on an individuals
investment performance. For example, a positive stock return can lead to negative
equity return if the level of leverage is negative.
Thirdly, leverage can lead the investor to ruin. For example, suppose that an
investor is 10 times leveraged and the stock return is -10%, the equity return is 100%. Therefore, the end of period return is -100%, and then after the -100% equity
return, the investor has no more capital to invest and, consequently, would not be
able to open any position.
Leverage has the ability to impact the investment performance. However, if used
on an excessive level, leverage can lead investor to bankruptcy. Thus, if excessive
level of leverage is bad for investments, could leverage be good on some level? Is
there an optimal level of leverage? If yes, could it be measured?
Kelly (1956) demonstrated that in binomial games with positive expectation,
there is one specific value of leverage which maximises the geometric mean or the
end of period wealth. He proposed a model to obtain this value and this model was
called the Kelly criterion. Today, one of the possible uses of the Kelly criterion is
Chapter 1. Introduction
1.2
Research Objectives
The research objectives of this thesis are to introduce a numerical method to calculate the optimal level of leverage that allows the relaxation of the assumption of
normal distribution of the series of financial returns; and to test the ability of this
method to improve investment performance. Thus, the hypotheses of this research
are that using numerical methods, the geometric mean of the historical financial
returns using numerical methods is superior to the geometric mean of the historical
financial returns using analytical methods; and that the use of the optimal level of
leverage using numerical methods improves investment performance.
I use two different approaches in this thesis. In the first approach, in order to
test the hypothesis that the value of the optimal level of leverage is more appropriate
than using analytical methods, I assume that the agents objective is to maximise
the end of period return. Maximising the end of period return is the same as
maximising the geometric mean (Williams 1936, Latane 1959). Thus, similarly to
analytical models, I assume that the optimal level of leverage is the level of leverage
that maximises the geometric mean of the series of the historical financial returns.
Thus, the test is a T-test of difference between the series of the geometric mean
of the series of the leveraged historical financial returns calculated using numerical
methods and the series of the geometric mean of the series of the leveraged historical
financial returns calculated using an analytical model with a closed-form solution.
In the second approach, I compare the individual end of period equity return
under three different experimental treatments using an agent-based model of an
order-driven market. In the first treatment, the agent maintains her level of leverage
equal to 1. In the second treatment, the agent maintains her level of leverage equal
to the value calculated using numerical methods. Finally, in the third treatment,
Chapter 1. Introduction
1.3
In this section, I describe the contributions of this thesis. The general contributions
of this thesis are the proposal of numerical methods to optimize the level of leverage
and the demonstration that the use of the optimal level of leverage using numerical
methods improves investment performance. The main advantage is that the use of
numerical methods allows the calculation of the optimal level of leverage directly
using the formula for the geometric mean. This allows me to relax the assumption
of the normal distribution of financial returns and, consequently, reduces the risk of
to underestimating extreme returns.
In order to do this, I assume that the investors objective is to maximise the end
of period equity return. Thus, this serves to measure investment performance, and
the end of period equity return can be represented by the geometric mean of the
equity returns (Williams 1936, Latane 1959).
There are three specific contributions of this research. The first contribution is
to show that a commonly-used proxy of the geometric mean, which is widely used
in the current literature, fails to accurately estimate leveraged returns. This proxy
uses the arithmetic average and the variance of the stock returns (Parramore and
Watsham 1997). However, this proxy assumes that returns can be characterised
by the first two moments and ignores the fat-tailed return distributions that are
observed in actual empirical data (Cont 2001).
1.4
This thesis consists in seven chapters, including the current one. In Chapter 2, I
present the literature review of this thesis.
10
Chapter 1. Introduction
In Chapter 3, I discuss the estimation of the returns on leverage. In the literature,
there is a widely used approximation of geometric mean that employs the arithmetic
average and the variance of the stock return (Parramore and Watsham 1997) but
ignores the kurtosis. I discuss the use of this formula for leveraged equity return
instead of stock return.
In Chapter 4, I propose a numerical method in order to calculate the optimal
level of leverage. The advantage of the use of numerical methods is that, differently
of the analytical models existing on the literature, the method proposed in this thesis
allows me to relax the assumption of normal distribution of the series of financial
returns and, consequently, undermine the risk of underestimation of extreme returns.
In Chapter 5, using the historical daily price of the components of the S&P 500
Index, I back-test the use of the optimal level of leverage using numerical methods. I
compare the end of period returns of the optimally leveraged investment to geometric Brownian motion leveraged, unleveraged and excessively leveraged investment,
respectively.
In Chapter 6, I employ an agent-based model in order to experiment the use of
the optimal level of leverage in a totally controlled environment. I introduce three
leveraged agents regarding their approach of leverage, the optimal level of leverage
using numerical methods, the unleveraged and the excessive level of leverage, in
an order-driven model proposed by Iori and Chiarella (2002). Then, I simulate it
regarding to three different sizes of market, 10, 100 and 1000 agents in the JASA
platform (Phelps 2007).
Finally, in Chapter 7, I summarize the main findings of this thesis and conclude.
I also discuss ideas for future research.
In the next chapter, I present the publications originated from this thesis.
1.5. Publications
1.5
11
Publications
There are two publications from this thesis. Firstly, material from Chapter 4 has
been peer-reviewed and published as:
Sbruzzi, E. and S. Phelps: 2011. Optimal Level of Leverage using Numerical
Methods. In, 7th Conference of the Portuguese Finance Network, Aveiro,
Portugal.
Secondly, material from Chapter 6 has been submitted as:
Sbruzzi, E. and S. Phelps: 2013. Testing leverage-based trading strategies
under an adaptive-expectations agent-based model. In, 12th International
Conference on Autonomous Agents and Multiagent Systems (AAMAS 2013).
In the next chapter, I review the literature of the thesis.
Chapter 2
Literature Review
In this chapter, I present the literature review of this thesis. Firstly, in Section 2.1,
I review the literature on leverage. Secondly, in Section 2.2, I review the literature
on models to estimate returns. Thirdly, in Section 2.3, I review the literature of the
investors objective function. Fourthly, in Section 2.4, I demonstrate how to relate
geometric mean, end of period return and level of leverage. Fifthly, In Section 2.5,
I present the existing models to optimize the level of leverage. Sixthly, in Section
2.6, I review the literature on the agent-based modelling. Finally, in Section 2.7, I
present the summary of the chapter.
2.1
Leverage
In this section, I present the literature review of leverage and its implication to the
individual investors performance and the financial market as a whole. Breuer (2002)
argues that leverage can be thought of as elasticity; indicating the responsiveness of
the value of equity to changes in the value of overall stocks:
kt = lt rt
(2.1)
where kt is the equity return, lt is the level of leverage and rt is the stock return. Note
that stock return, rt , is different of equity return kt . Stock return is the percentage
14
rt =
Pt
1
Pt1
where Pt is the price on the period of time t. Consequently, the impact of the stock
return on the equity return depends on the level of leverage of the position.
As changes in the value of equity are equal to changes in the stock portfolio,
leverage is conventionally defined as the ratio of assets to equity. Furthermore, by
assuming that the stock return expectation is approximately zero and that variance
is the representation of risk, Leverage increases the variance of the equity return
and, consequently, increase the risk of the equity return:
(2.2)
2.1. Leverage
15
by the funds competition for new investors. During boom times, higher leverage
hedge funds show superior returns and these hedge funds use such returns to attract
new investors. This competition creates a bubble in leverage and, consequently, a
bubble in the financial markets.
Khandani and Lo (2007; 2008) demonstrate that the use of leverage was quite
common among quantitative equity market-neutral strategies, where the higher
leverage ratios were used by those managers engaged in high-frequency margin overflow strategies because those strategies exhibited the lowest volatilities and highest
Sharpe ratios.
Adrian and Shin (2010) point out the relationship between leverage and liquidity; there is a positive correlation between leverage and the financial intermediaries
and commercial bank balance sheet. They demonstrate that, on one hand, the commercial banks and financial intermediaries target a fixed leverage ratio; and, on the
other hand, for households, there is a strongly positive relationship between changes
in total assets value and changes in leverage. Furthermore, they state that leverage
is pro-cyclical. The adjustments of the leverage and price reinforce each other amplifying the financial cycle and, due to the size of the financial intermediaries, this
strongly influences the liquidity, possibly leading to unwinding as observed during
the period between the crisis in July and August 2007 (Khandani and Lo 2007).
Brunnermeier and Pedersen (2005) explore the predatory characteristics of the
combination of leverage, liquidity and the overreactions of strategies. The bear
market can force traders and financial intermediaries to liquidate their position to
cover their margin calls, and provided that others players have this information, it
can lead other players to trade in the same direction, resulting in a withdraw of
liquidity from the market. This activity makes liquidation costly and leads to price
overshooting.
According to Loeys and Panigirtzoglou (2005), leverage is considered an important driver of how much the market reacts to news. Leverage increases volume
without a similar increment in the number of participants. This leads to the con-
16
centration of the risk in fewer hands. Leverage also increases the probability of
bankruptcy and can force investors to close the position in hostile conditions, which
would force the price down, exacerbating volatility. Furthermore, volatility and
leverage are positively correlated and volatility causes leverage.
The conclusion that leverage and volatility are positively correlated is consistent
with the optimal leverage model for non-ergodicity as introduced by Peters (2010).
Using geometric Brownian motion, he demonstrates that the optimal level of leverage
is simply the estimation of log-return divided by variance of log-return. Hence, the
objective is to estimate the appropriate return and variance.
Domian, Racine and Wilson (2003) use a continuous time model to derive distributions for leveraged portfolios over long periods. They demonstrate that expected
return is linear in leverage in a single-period CAPM model, but concave in the
continuous time model.
From the literature, leverage is a way to expand profitability. These characteristics create the expectation to gain huge amount of money in shorter period.
Therefore, it can affect the investors behaviours with consequences to the financial
market. The point is that, together with the expansion of the profitability, there is
a natural increase in the probability of bankruptcy.
I assume that behaviour is irrational if the investor increase the level of leverage
without any mathematical optimization of the level of leverage dictating this increment. Hence, the combination of irrational behaviour and probability of bankruptcy
can be extremely harmful for the individual investor and, consequently, to the financial market. If an individual agent is leveraged and in bankruptcy, there is a
probability that this agent does not repay its margin assumed. For this reason,
the counterpart of this position, the lender, can also have the consequences of this
situation. If the number of bankruptcy agent is low, the impact of this situation in
the financial system is low and, therefore, the chance of this to be absorbed by the
market is higher. However, if the number of bankruptcy agent high, the chance of
this to be absorbed by the market is low and, consequently, this can collapse the
17
financial market.
There is a competition for new investors and, particularly during boom times,
this competition is reinforced by the increase in the level of leverage. Thus, the
number of agents will increase their exposure and the level of leverage will increase
with the duration of the boom times. Naturally, this situation creates bubbles.
However, besides creating bubbles, this competition increases the impact of the
bearing market on the financial system sustainability, due to the huge number of
agents leveraged.
Although the literature on leverage focuses on the consequences to the market, it
is important to see the consequences of leverage to the individual investor. I assume
that the financial market is a result of the interaction of different agents. Hence,
instead of the contribution with techniques to use leverage in order to improve the
financial market, this thesis aims to contribute with techniques to the individual
investor to use leverage in their investment.
In the next section, I review the literature on the estimation of the returns.
2.2
Return Estimation
In this section, I review the literature on the appropriate estimator for the returns.
The reason is that in Chapter 3, I test the ability of the proxy of the geometric
mean to be used as an estimator of the real geometric mean of leveraged returns.
The debate about the appropriate estimator for returns has its origin in Williams
(1936), who demonstrated that speculators, in a multi-period framework, should be
concerned with the geometric mean instead of the arithmetic average. The main
argument is that arithmetic average does not consider the non-linearity of returns
present on exponential compound series, like financial series, which is only captured
by geometric mean (Latane 1959).
Latane (1959) and Williams (1936) argue that, empirically, the investors objective is to maximise the end of period wealth. Thus, the estimator has to be able to
18
1X
rt
n t=1
(2.3)
(2.4)
t=1
GM '
2
2
(2.5)
19
I demonstrate how the proxy of the geometric mean of the return, GM , as shown
in Eq. (2.5), is obtained using Taylor approximation. Lets define x = ln(1 + GM ).
Thus from Eq. (2.4):
n
x=
1X
ln(1 + rt )
n t=1
1X
rt 2 rt 3 rt 4
x=
(rt
+
+ ...)
n t=1
2
3
4
Thus:
E(r2 ) E(r3 ) E(r4 )
x = E(r)
+
+ ...
2
3
4
where E(a) =
1
n
Pn
t=1
x E(r)
E(r2 )
2
(2.6)
GM = x +
x2 x 3 x4
+
+
+ ...
2!
3!
4!
Assuming that if a + b > 2 then E(ra )E(rb ) = 0 and E(ra )b = 0, then the proxy
of the geometric mean is:
GM x +
x2
2!
(2.7)
E(r )
E(r2 ) (E(r) 2 )
GM E(r)
+
2
2
0
(2.8)
20
GM E(r)
E(r2 ) E(r)2
2
(2.9)
Since the arithmetic average is = E(r), and the variance is 2 = E(r2 )E(r)2 ,
I obtain Eq.(2.5):
GM
2
2
In order to obtain the result the closed-form solution Eq. (2.5), that the proxy of
the geometric mean is the arithmetic average minus half of the variance, the model
has three assumptions. The first assumption is that the returns are always greater
than minus one and less or equal to one. This means that I assume that there is
no return less or equal to -100%, i.e., there is no probability of default. The second
assumption is that the stock returns are so small that the estimation of the return
power any value superior to two is zero. Finally, the third assumption is that if the
sum of the value of the power of the return is superior to two then the product of
the estimation of the return is also zero. The consequence of the second and the
third assumption is that the model assumes that the value of any statistical moment
superior to two is 0. Thus, the value of the skewness and the kurtosis is zero.
In the next section, I review the literature on the investors objectives.
2.3
In this section, I review the literature on the investors objective function. The
reason is that, in Chapter 4, I assume that the investors objective is to use the
geometric mean approach in order to optimize the level of leverage.
Studies of optimal level of leverage are associated with debates of individual
investors objective function. In the literature, the investors objective function can
be divided in two different approaches: mean-variance (Markowitz 1952) and growth
optimal portfolio (Kelly 1956 and Latane 1959). Leverage is linear in the mean-
21
variance approach, and concave in the growth optimal portfolio approach. Due to
linearity of leverage, studies of leverage are irrelevant in the mean-variance approach,
because even when leverage is introduced into the parameters of the model, it does
not alter the model itself; i.e., the model is independent of leverage. However, due to
the non-linearity of leverage in the growth optimal portfolio approach, the study of
leverage is important because the introduction of leverage on the parameters alters
the results of the model.
There is a very important debate about the objectives of individual investors.
While Markowtiz (1952) developed the mean-variance method to determine the
optimal portfolio to the next period; Kelly (1956) and Latane (1959) argued that
the main objective of the individual investor is to maximise the end of period wealth;
hence, the investor should be concerned about the capital rate of growth which could
be measured using the geometric mean.
This debate continued during the 1960s and 1970s. On one hand, Breiman (1961)
argues that on a long sequence of trials, the game objectives are to minimise the
time to reach the target level of wealth, or to maximise the value of the end of
the period wealth. Breiman (1961) and Hakansson (1971a; 1971b) show that the
optimal strategy to attain both objectives is to maximise the expected value of the
log of the terminal wealth which is the same as maximising the geometric mean of
return.
On the other hand, Samuelson (1971; 1979) argues that geometric mean maximisation was only one among many investment rules and there is no rationale to
suppose that it is the best. He shows that the geometric mean rule leads to suboptimal expected utility and, because the end of period expected is the sum of the
utility for each period, the end of period wealth under geometric mean rule will also
be sub-optimal.
The debate was theoretical with each author advocating different rules by using complex mathematical models. The reason is that during that period, the
use of numerical methods was extremely difficult because of the primitive level
22
RR =
(2.10)
where RR is the risk-return and and are the arithmetic average and the standard
deviation of the historical returns, respectively.
The alternative methodology to calculate the return is the geometric mean formula:
n
Y
1
GM =
(1 + rt ) n 1
(2.11)
t=1
A debate about the ideal methodology to measure average return has its origin in
Williams (1936), who demonstrates that speculators, in a multi-period framework,
should be concerned about the geometric mean instead of the arithmetic mean.
The main argument is that risk-return does not consider the non-linearity of returns
present on exponential compound series, like financial series, and such characteristics
are only captured by the geometric mean (Latane 1959).
In order to introduce leverage, Eq. (2.10) can be rearranged, and consequently
the leveraged risk-return formula (RRL) is:
RRL =
l
= RR
l
(2.12)
23
where l is the level of leverage. Note that leverage is monotonic and linear in the
risk-return formula. For this reason there is no level of leverage that maximizes
the Eq. (2.12). Hence this formula is not able to be used to optimizes the level of
leverage. Rearranging Eq. (2.11), the geometric mean formula (GM(l)) is:
GM (l) =
n
Y
(1 + rt l) n 1
(2.13)
t=1
From Eq. (2.13), note that leverage is neither monotonic nor linear in the geometric mean formula. For this reason, different models have been proposed to obtain
the level of leverage that optimize the geometric mean formula.
In the next section, I review the relation between geometric mean, end of period
return and level of leverage.
2.4
In this section, I demonstrate how to relate the geometric mean, the end of period
return and the level of leverage. This section is divided in two subsections. In the
first subsection, I present the impact of the level of leverage in a single period equity
return. And, in the section subsection, I demonstrate the impact of the level of
leverage in the end of period return and on the geometric mean.
2.4.1
24
lt =
St
Kt
(2.14)
where lt represents the level of leverage, St represents the market value of the total
securities and Kt represents the equity value. For simplicity, no transaction costs
are considered. The leverage is:
Lt = St Kt
(2.15)
St = Pt Qt
(2.16)
where Pt represents the price and Qt represents the quantity. I assume that the
level of leverage is represented by Eq. (2.14) and the market value of the position
is represented by Eq. (2.16). Thus, extending the model to two time periods, t and
t + 1, the stock return to the next period is represented by rt , insomuch:
Pt+1 = Pt (1 + rt+1 )
(2.17)
where Pt+1 represents the price for the next period. Assuming that Qt+1 = Qt ,
substituting Eq. (2.17) into Eq. (2.16) and rearranging:
St+1 = St (1 + rt+1 )
(2.18)
where the market value of the portfolio in the next period is St+1 . From Eq. (2.18),
the market value of the portfolio depends on the stock return. Assuming that the
leverage is fixed from time t to time t + 1, Lt+1 = Lt , substituting Eq. (2.18) into
Eq. (2.14) and rearranging :
(2.19)
25
where the equity for the next period is Kt+1 . From the Eq. (2.19), equity is dependent on the stock return and on the level of leverage. The equity return is:
kt+1 =
Kt+1
1
Kt
(2.20)
where kt is the equity return at time t. Thus, substituting Eq. (2.19) in Eq. (2.21),
the equity return is:
(2.21)
Note that the equity return is the product of stock return and the level of leverage. This means that the level of leverage is the impact of the stock return on
the equity return. Leverage has the ability to completely modify the impact of the
stock return because the level of leverage can assume negative values. For example, assuming that the stock return is +1%, the position is short and the level of
leverage is 3. The short position means that the level of leverage can be rewritten
as -3. Thus, the equity return is -3%, i.e., even if the stock return is positive, +1%,
the equity return is negative -3%. Therefore, if the end of period equity return is
the consequence of cumulative single period equity return and the objective is to
maximise the end of period equity return, leverage is relevant because of its direct
impact on the single equity return.
In the next subsection, I demonstrate the impact of the level of leverage in the
end of period return and on the geometric mean.
2.4.2
In this subsection, I demonstrate that the end of period return can be represented by
the geometric mean of the equity return and they are functions of the stock return
dynamic process and the level of leverage dynamic process. I also demonstrate that
what maximises the end of period equity is the same that maximises the geometric
mean of the equity returns. Initially, substituting Eq. (2.21) in Eq. (2.19), the next
26
Kt+1 = Kt (1 + kt+1 )
(2.22)
From Eq. (2.22), the equity on the next period is a function of the equity return.
Proceeding similar calculation, the equity on the next period, t + 2, is:
(2.23)
From Eq. (2.23), after two periods of time the equity is a function of the cumulative equity returns over the time t and t + 1. Thus, applying similar calculations
until the end of period, t = T , and considering the initial time equal to 0, the end
of period equity is:
T
1
Y
KT = K0
(1 + kt+1 )
(2.24)
t=1
From Eq. (2.24), the end of period equity is a function of the cumulative equity
returns. the geometric mean of is:
T
1
Y
(1 + kt+1 )] T 1
GM = [
(2.25)
t=1
KT (GM ) = K0 (1 + GM )T
(2.26)
Eq. (2.26) shows that the end of period equity is a function of the geometric
mean of the equity returns.
In order to verify if the maximisation of geometric mean of equity returns leads
the maximisation of end of period equity, I differentiate Eq. (2.26):
dKT
= T K0 (1 + GM )T 1 > 0
dGM
(2.27)
27
According to Eq. (2.27), the impact of the geometric mean on the end of period
equity is positive. Thus, the maximisation of the geometric mean of the returns
leads to the maximisation of the end of period equity.
According to Eq. (2.21), equity return is a function of the stock return and the
level of leverage. Thus, Substituting Eq. (2.21) in Eq. (2.26) and considering the
initial t 1 = 0:
(2.28)
From Eq. (2.28), the end of period equity is a function of the stock return
dynamic process and level of leverage dynamic process. The end of period return is:
(2.29)
Thus, from Eq. (2.29), the end of period return also depends on the series of
stock return r and level of leverage l.
Assuming that the stock returns dynamic process is exogenous, i.e., out of the
investors control; the variable of control is the level of leverage. For this reason, in
order to maximise the geometric mean of the equity return, investors should to be
concerned on the level of leverage that maximises the geometric mean of the equity
return. Hence, the appropriate estimator of the geometric mean of leveraged returns
is relevant.
In Section 2.2, I stated that the proxy of geometric mean as Eq. (2.5) has three
assumptions. Firstly, the returns is greater than minus one and less or equal to one.
Secondly, the stock returns is so small that the estimation of the return power any
value superior to two is zero. Finally, the sum of the value of the power of the return
is superior to two then the product of the estimation of the return is also zero.
However, this proxy assumes that the returns are normally distributed and this
implicates that the value of the skewness and kurtosis is 0, i.e., extremes are barely
on the distribution. The key point is that leverage expands the impact of the stock
28
return (Breuer 2002) and, naturally, increases the kurtosis. Thus, it contradicts
the assumptions of the advocators of the approximation for the geometric mean.
Therefore, the leverage characteristics could indicate that the approximation is not
appropriate for the equity return.
For example, if we expand our original example of two periods stock returns to
four periods stock return, the stock returns are +8%, -10%, +10% and +5%. Then,
if we assume that the level of leverage is 1, the arithmetic average, the proxy of
geometric mean and the real geometric mean of the equity returns are 3.25%, 2.84%
and 2.93%, respectively. Note that the value of the estimators are considerable
near to one another. Furthermore, this result suggests that the arithmetic average
overvalues the rate of return and the proxy of the geometeric mean undervalues the
rate of return, which agrees with the findings in Jacquer et al. (2003).
However, if we assume that the level of leverage is 10, the stock return remains
unchanged, but, the equity return are +80%, -100%, +100% and +50%. Consequently, the arithmetic average, the proxy of geometric mean and the real geometric
mean of the equity returns are 32.50%, -8.63% and -100%, respectively. Note that
in this case, there is a considerable distance between the estimators. Furthermore,
this result demonstrates that the arithmetic average and the proxy of geometric
mean are not able to capture the real rate of return, which is -100%. The reason
for this failure is that both estimators do not capture the risk of default which is
increasing in leverage; i.e., the higher the level of leverage, the higher the risk of
default (Breuer 2002).
In the next section, I review the literature on optimal leverage models.
2.5
In this section, I present the three existing models to optimize the level of leverage.
The first model is the Kelly criterion which assumes normal distribution of the
returns and discrete number of possible outcomes (Kelly 1956). The second model
29
is the Rotando and Thorp model which assumes normal distribution of the returns
and continuous number of possible outcomes (Rotando and Thorp 1992). Finally,
the third model is the Peters model which assumes normal distribution of the returns
and that price follows a geometric Brownian motion (Peters 2010).
This section is divided in three subsections. In the first subsection, I describe
the Kelly criterion. In the second subsection, I describe the Rotando and Thorp
model. Finally, in the third subsection, I describe the Peters model.
2.5.1
(2.30)
i=1
Assuming p > 1/2 > q and, consequently the expected value of the end of
30
trials wealth E(Xn ) is positive, the objective is to find the amount of the available
capital which should be bet, Bi , and the amount of capital that should be saved
for later trials, Xi1 . On one hand, if the player decides to bet all the capital, this
increases the probability of ruin. On the other hand, if he decides to bet the minimal
capital, the player reduces the probability to maximise the value of the end of trials
wealth. Therefore, there is some optimal level of leverage or optimal fraction, l,
which balances the objective to maximise the expected value with the restriction to
minimise the probability of ruin.
In the coin-tossing game, since the gambling probability and the payoff at each
trial are the same, it is clear that the optimal level of leverage is the same for all
trials. This assumption of fixed leverage helps us to comprehend the purpose of the
Kelly criterion. Maximising the expected end of trial wealth is similar to maximising
the expected value of the growth rate coefficient or the geometric mean, GMk (l):
Xn 1
] n = p log(1 + l) + q log(1 l)
X0
q
p
1+l 1l
(2.32)
l2 + 2l(p q) 1
(1 l2 )2
(2.33)
GMk0 (l) =
GMk00 (l) =
(2.31)
where GMk (l) is the geometric mean approximation on the Kelly criterion and
0 l < 1. From Eq. (2.33), the function GMk (l) is concave in l, hence it can be
maximised. Furthermore, solving Eq. (2.32), the result is that the optimal level of
leverage is l = p q.
In the next subsection, I describe the Rotando and Thorp model.
2.5.2
In this subsection, I describe the Rotando and Thorp model. The Kelly criterion
cannot be directly used on investments. The main reason is that with games like coin
31
tossing there are a discrete number of possible outcomes, whereas with investments,
the number of possible outcomes is continuous. In order to use the Kelly criterion
in financial markets, the original model proposed in Eq. (2.31) should be modified
and adapted to represent continuous outcomes. This modification is proposed by
Rotando and Thorp (1992). They propose a new model incorporating continuous
outcomes, but respecting the Kellys original insight: the definition of the optimal
level of leverage in order to maximise the end of period wealth or the end of period
return, E[log(Xn/X0)].
According to Rotando and Thorp (1992), trading financial securities can be
considered a continuous game. Thus, in order to model the possible outcomes, they
assume that financial returns are Normally distributed. However, the simple use of
the unaltered Normal probability distribution is inadequate because this distribution
allows an infinite range of possible returns. Therefore, they modify the standard
normal curve using a correction term for chopping off the tails (Rotando and
Thorp 1992). This results in new parameters, h and , which serves to maintain the
mean and the standard deviation at values similar to those of the standard Normal
distribution..
Similar to binomial games, the objective in this model is to find out the level of
leverage that maximises function GMrt (l), which in this case is:
Z
GMrt (l) =
(2.34)
where GMrt is the geometric mean approximation if the Rotando and Thorp model.
Z
GMrt (l) =
log(1 + rl)[h +
1
22
2 /22
e(r)
(2.35)
where A = 3 and B = + 3.
The demonstration of first-order and the second order conditions of Eq. (2.35)
is complicated. However, numerical methods can be used to obtain the value of l
that maximises the function GMr t(l). For example, simulating the model proposed
32
for the 59 year period from 1926 to 1984, with the = 0.058 and = 0.216 by
using numerical methods, Rotando and Thorp (1992) found that the optimal level
of leverage is l = 1.17.
In the next subsection, I describe the Peters model that uses the geometric
Brownian motion.
2.5.3
In this subsection, I describe the Peters model that uses the geometric Brownian
motion. Assuming that the price follows a geometric Brownian motion, and consequently, that returns are log-normally distributed (Hull 2009), Peters proposes a
different model to optimize the level of leverage (Peters 2010). Suppose that the
price process is:
p(t) = p0 [exp (
2
)t + W (t)]
2
(2.36)
l = l
(2.37)
l2 = l2 2
(2.38)
2
)
2
Substituting Eq. (2.37) and Eq. (2.38) into Eq. (2.39) gives:
(2.39)
33
GMp (l) = (l
l2 2
)
2
(2.40)
In order to obtain the optimal level of leverage, Eq. (2.40) is differentiated with
respect to l and the result is set to zero. Then, the optimal level of leverage is:
l =
(2.41)
2.6
Agent-Based Modelling
In this section I review the literature on agent-based modelling. The reason is that
in Chapter 6, I test the ability of the optimal level of leverage to improve investment
performance in an agent-based modelling experiment.
Agent-based modelling has the ability to go beyond the traditional economics
because it considers the market as a multi-agent system wherein each agent has
particular characteristics that differs from other agents, while the traditional economics assume that every agent is similar and rational (Chakraborti et al. 2011).
The traditional economic approach with closed solutions was crucial to allow the
proposition of analytical models which would be unfeasible without the assumption
of similarity and rationality among the agents. However, advances in the technology and in the computer platforms such as StarLogo (Resnick 1996), NetLogo (Sklar
34
2007), Repast (Collier 2003), JABM (Java Agent-Based Modelling) (Phelps 2012)
and JASA (Java Auction Simulation API) (Phelps 2007) have enabled the progress
in numerical approaches and simulation instead of analytical approach in order to
model any aspects of different areas of research, for example, in the Biology and in
the Sociology (Chakraborti et al. 2011).
Naturally, agent-based modelling research has been proposed for Economics and
Finance. According to LeBaron (2000; 2006), the combination of agent-based models
and computational methods in finance creates a new sub-area called agent-based
computational finance. One use of agent-based computational finance is artificial
stock market which consists on a set of tools that intends to replicate artificially
the stock market (LeBaron 2002). This is a controlled environment which allows
one to run experiments using artificial agents in order to analyse a wide variety of
hypotheses from the literature of finance.
Different agent-based models have been proposed with the objective of investigating how different trading strategies may affect the dynamics of price, bid-ask
spreads, trading volume and volatility. The literature of agent-based modelling for
finance used in the thesis evolves along of the time as follows: Firstly, The proposal
of agent-based modelling for finance is not recent in the literature, its origins are
contained in Grossman (1976) who proposed a model with heterogeneous agents
that interact with each other. One of the most important debates in agent-based
modelling area is how to model the artificial agent, with its individual characteristics, heterogeneity and behaviour (LeBaron 2000). Lettau (1997) proposes a very
simple model where agents should decide between a risk and risk free asset according
to personal utility function.
Secondly, Bak, Paczuski and Shubik (1996) propose a model which assumes that
orders are input according to a price line. They assume that there are N agents and
N/2 stocks. The agent is a buyer if he does not hold the stock, and a seller if he
holds a stock. Agents can only sell stock if they hold one, and they can only buy
stock if they dont hold one. This means that there is no leverage in this model.
35
Agents advertise the value that they intend to buy (sell) the stock and the trade
occurs when the prices cross, i.e., the offer to buy is superior to the price to sell. At
this moment, the seller trades with the buyer that offers the highest price. In the
case of no trade, the prices are updated to the next round.
Thirdly, Maslov (2000) introduces a model with limit and market orders. This
is an important step towards the replication of the real financial market. In their
model there is no strategy pre-defined. Traders simply collect the last transaction
price, p(t) and aggregate a small variation on this price . Thus, the order to sell
is above the current price, p(t) + , and the order to buy is below the current price
p(t) . In the first moment, there is no trade, because the highest order to buy is
inferior to the lowest order to sell. However, in this model, agents are not allowed to
update or cancel their orders along of the rounds. Consequently, after some rounds,
the probability of having the order matched increases because the new current price
p(t + n) indicates one new order to sell p(t + n) + that could be inferior to the
original price to buy p(t) .
Fourthly, Lux and Marchesi (2001) propose that agents decide according to the
combination of three components. The first component is the fundamentalist. The
fundamentalist component uses the fundamental value of the security, For example,
in the case of stocks, the intrinsic value of the company is calculated using the
information in its balance sheet. The fundamentalist component indicates to sell
(buy) when the price is bigger (smaller) than the stock intrinsic value or fundamental
price.
The second component is the chartist. The chartist component uses the historical data instead of the balance sheet. The chartist component can be divided into
contrarians and trend followers. Contrarians are based on the existence of overreaction on the stock returns (Lo and MacKinlay 1990). This indicates to buy (sell)
past losers (winners). Unlike the contrarians, trend-followers are based on the existence of trends on the financial series (Brock, Lakonishok and LeBaron 1992). This
indicates to buy (sell) past winners (losers).
36
acterized. The noise trader does not follow any standard or pre-defined behaviour.
This indicates decisions using no strategy and zero-intelligence.
Finally, Iori and Chiarella (2002) introduce an order-driven market model with
heterogeneous agents trading via a central order matching mechanism in a double
auction, dispensing with the need for a market maker. At the moment of ordering,
agents face three different parameters in their decision: the price (limited at some
value or marketed), the direction (long or short) and the volume or the level of
leverage. Iori and Chiarella (2002) propose model in which agents, in combination
with three different components (the fundamentalist, the chartist, and the noise,
mentioned above) have intelligence in order to decide the price and the direction.
However, there is no intelligence in order to decide the level of leverage, rather we
simply assume that each agent can only trade one stock at a time:
i
Ert,t+n
= g1i
p f pt
+ g2i rLi + ni et
pt
where g1i is the weight of fundamentalist component which is positive, and g2i is the
weight of the chartist component which is positive in the case of trend followers,
and negative in the case of contrarians.
Each agent has different values for g1i and g2i which is normally distributed.
The parameter rLi is the moving average value of the return over Li period and is
uniformly and independently distributed. The position is opened according to the
i
i
value of the expectation of return Ert,t+n
. Positive (negative) Ert,t+n
indicates long
(short) position.
Iori and Chiarella (2009) extend the model proposed by Iori and Chiarella (2002)
introducing intelligence in the volume, proposing that agents should invest according
to their individual exponential utility of wealth function U (w) = eW . Consequently, the optimal quantity agent i is:
i (p) =
log(Epi(t+n) ) log(p)
i t p
37
where t is the standard deviation of the historical return and i is the risk-profile
value.
The risk-profile value is a parameter which measures how adverse to risk is the
agent; i.e., the higher the risk-profile value, the more risk-adverse is the agent. Note
that the optimal volume is negative correlated to the risk-profile value; the more
risk-adverse is the agent, the smaller is the volume.
For the purpose of this thesis, I use the order-driven model as proposed by Iori
and Chiarella (2002) and Lux and Marchesi (2001). I assume that this model is
the best representation of trade price expectation scheme because the order-driven
model assumes that the individual agent decision has the three components, the
fundamental component, the technical component and the noise component. The
only difference is the weight of each particular component to the agents. There are
two advantages of the model. Firstly, the components weight are randomly among
the artificial agents and all of the three component are taken into account. This
is natural that the real agents use a mix of signal during their decision process.
Intuitively, It is difficult to conceptualize that some agent is 100% technical and 0%
fundamentalist and vice-versa. The most natural assumption is that each agent is
unique and gives different weights to each component, and this situation is captured
by the model.
Secondly, agents are assumed to be boundedly rational, thus their behaviour
has simple characteristics, that can be described by the components. The model
is in line with the agents behaviour characteristics because his price expectation is
simply the combination of the three components and the historical prices.
In terms of the quantity specified in orders, Iori and Chiarella (2009) propose
an extension based on the utility function. In Section 2.2, I discussed the difference
between the approach of the investors objective and the reasons to use the geometric
mean with the leverage in this thesis. For this reason, in Chapter 6, I assume that
instead of maximising their utility function, the agents objective is to maximise the
geometric mean of the equity returns.
38
2.7
Summary
In this chapter, I reviewed the literature of this thesis. This chapter is divided in
six sections, including the current one. In the first section, I reviewed the literature
on leverage and its implication to the individual investment performance. Leverage
allows to enhancing the equity return. This characteristic encourages excessively
level of leverage which can lead the investor to default with consequences to the
financial market as a whole. Even though the investors behaviour impacts the
financial market, the literature is concentrated in the impacts of leverage on the
financial market as a whole instead of the individual investors behaviour. For this
reason. this thesis aims to contribute with techniques to the individual investor to
use leverage in their individual investment.
In the second section, I reviewed the literature on the return estimation because
in Chapter 3, I discuss the ability of that models to be employed to estimate the
leveraged returns. The proxy of geometric mean as Eq. (2.3) has two assumptions.
Firstly, all returns are greater than minus one and less or equal to one. Secondly,
the stock returns are so small that the estimation of the return power any value
superior to two is zero. Due to its assumptions, the proxy of geometric mean ignores
the presence of kurtosis. The kurtosis increases in leverage (Thurner, Farmer and
Geanakoplos 2011), and is one of the characteristics of the series of financial returns
(Cont 2001).
In the third section, I reviewed the literature on the investors objective function
because in Chapter 4, I assume that the investors objective is to use the geometric
mean approach in order to optimize the level of leverage. There are two different
approaches of the investors objective function: The mean-variance approach and
the geometric mean approach. Furthermore, leverage is monotonic and linear in the
risk-return formula, but is neither monotonic nor linear in the geometric mean. For
2.7. Summary
39
this reason, different models have been proposed to obtain the level of leverage that
optimize the geometric mean formula.
In the fourth section, I demonstrated how to relate the geometric mean, the end
of period return and the level of leverage. This section is divided in two subsections.
In the first subsection, I demonstrate that the end of period return is dependent
on the stock return and on the level of leverage. And, in the section subsection, I
demonstrate that the end of period return can be represented by the geometric mean
of the equity return and they are functions of the stock return dynamic process and
the level of leverage dynamic process.
In the fifth section, I reviewed the literature on the existing optimal leverage
models on the literature. There are three different models. The first model is the
Kelly Criterion that serves to determine the optimal level of leverage on investments
considering the possible returns on different discrete scenarios (Kelly 1956). The
second model is the Rotando and Thorp model that is an adaptation of the Kelly
criterion for continuous gambling games, such as financial returns (Rotando and
Thorp 1992). Finally, the third model is the Peters model, which assumes that
prices follow a geometric Brownian motion and, consequently, the optimal level of
leverage is a function of the arithmetic average and the variance of the stock return
(Peters 2010). One of the drawbacks of these models is that they assumes that the
financial returns is normally distributed and, consequently, ignores the presence of
kurtosis (Cont 2001). With the objective of relax this assumption, in Chapter 4, I
propose a numerical method in order to optimize the level of leverage.
Finally, In the sixth section, I reviewed the literature of agent-based modelling
because in Chapter 6, I employ this environment to test the ability of optimal level
of leverage to improve the investment performance. Agent-based modelling allows
me to replicate the stock market artificially in order to make controlled experiments
and simulations. There are no studies in the literature that employ an agent-based
model to test the optimal leveraged models, and its consequence to the individual
investors performance.
40
which is widely used in the current literature, fails to accurately estimate leveraged
returns.
Chapter 3
Proxy Vs Real Geometric Mean
on Leverage
In this chapter, I demonstrate that the proxy of geometric mean existing in the literature fail in order to estimate leveraged returns. I analyse the use of the geometric
mean approximation widely used in the literature, in leveraged equity return. The
formula of the proxy is a function of the mean and the variance, and does not take
into account the probability of default(see Section 2.2); it assumes that the value of
the kurtosis of the equity equity returns, which increases with the level of leverage,
is zero. For this reason, I test the ability of the geometric mean approximation to
estimate the real rate of return of leveraged investments. I find that the proxy is
accepted as an estimator of the geometric mean of the unleveraged returns but is
not accepted as an estimator of the geometric mean of the leveraged returns.
The chapter is structured as follows: In Section 3.1, I present the introduction
of this chapter. In Section 3.2, I present the experiment of the proxy and the real
geometric mean of the returns of unleveraged and leveraged investments. In Section
3.3, I discuss the results of the experiment. Finally, in Section 3.4, I present the
summary and the contribution of this chapter.
42
3.1
Introduction
In this thesis, I assume that the investors objective is to maximise the end of period
return; and, in the existing literature, the formula proposed to represent the end of
period equity returns is the geometric mean (Williams 1936 and Latane 1959). There
is an approximation of this model that is widely used in the literature of finance
which uses arithmetic average and variance (Parramore and Watsham 1997). This
proxy exists in order to make the calculation of the geometric mean more friendly
with a closed-form solution.
In this chapter, I analyse the reliability of this proxy for leveraged investments.
Similar to the example above, the intuitively perception is that, even this approximation could be used on stock returns or unleveraged investments, it is not possible
to be used on leveraged investments because it does not take into account the risk
of default and the value of the kurtosis of the equity returns which increases in the
level of leverage.
Therefore, I perform two simulation experiments. In the first experiment, I
test the hypothesis that the proxy can be used as a real geometric mean of the
unleveraged returns. In the second experiment, I test the hypothesis that the proxy
can be used as a real geometric mean of the leveraged returns. The results are that
the proxy can be used as a real geometric mean of the unleveraged returns, and the
proxy should not be used as a real geometric mean of the leveraged returns.
In the next section, I present the results of the experiment that consists in the
test of the ability of the proxy of geometric mean to replicate the real geometric
mean of leveraged investments.
3.2
Experiment
In this section, I present the results of the experiment that consists in the test of
the ability of the proxy of geometric mean to replicate the real geometric mean of
leveraged investments. This section is divided in three subsections. In the first sub-
3.2. Experiment
43
3.2.1
Methodology
Collect daily data of every component of S&P 500 Index between 2000 and 2010.
Randomly select one component of the S&P 500 Index, one window period with 1000
days or more and one value for the level of leverage between 1 and 10.
Figure 3.1: The flow-chart of the sampling process of the proxy and the real geometric mean.
44
the real geometric mean. As shown in Figure 3.1, in order to test the hypotheses,
the sampling process consists in three steps. Firstly, I collect daily data of every
component of the S&P 500 index between January of 2000 and December 2010, 2766
trading days. This provides a dataset with 1.383 x 106 items of data. Secondly, From
the dataset, I randomly select one component of the S&P 500 index, one window
period with 1000 days or more and one value for the level of leverage between 1 and
10. Thirdly, I calculate and record the proxy and the real geometric mean of the
unleveraged and leveraged investments. I repeat this sampling process a total of
1000 times. The number of possible selected series is 758 x 103 . Thus, our sample
represents approximately 0.13% of the total (1/758).
The sampling process produces four different sets of data each consisting of 1000
data point: the series of the proxy of the geometric means and the series of geometric
means of equity returns for unleveraged and leveraged investments using Eq. 2.4
and Eq. 2.5, respectively.
In the next subsection, I present an example of iteration in order to show how
the data point in the generated series is obtained.
3.2.2
Single Iteration
rt =
Pt
1
Pt1
(3.1)
where rt is the return, Pt is the current price and Pt1 is the previous price as showed
3.2. Experiment
45
70
60
Price
50
40
30
20
10
0
500
1000
1500
2000
2500
Figure 3.2: The series of prices of UNH between 17/01/2001 and 04/08/2011.
in Figure 3.2. As the level of leverage of unleveraged investment is 1, I use the series
of stock return as the series of equity return of the unleveraged investment.
Thirdly, I randomly select the value of the level of leverage which is uniformly
distributed between 1 and 10. In this iteration the random level of leverage is 2.42.
Combining the series of the return and the random level of leverage, and using Eq.
3.1, I generate the series of the equity return of leveraged investment as showed in
Figure 3.3.
Fourthly, using the series of return, Eq. 2.4 and Eq. 2.5, I calculate the proxy of
the geometric mean and the real of the geometric mean of the unleveraged investment
and of the leveraged investment, respectively. The proxy of the geometric mean is
0.0470% and the real geometric mean is 0.0473% for the unleveraged investment;
and the proxy of the geometric mean is 0.0273% and the real geometric mean is
0.0287% for the leveraged investment.
Finally, these results are stored in four different series of data: a) the series of the
46
0.5
0.4
Stock Return
0.3
0.2
0.1
0
0.1
0.2
0
500
1000
1500
2000
2500
Figure 3.3: The series of returns of UNH between 17/01/2001 and 04/08/2011.
proxy values of the geometric mean of unleveraged investment, b) the series of the
real geometric mean of the unleveraged investment, c) the series of the proxy of the
geometric mean of the leveraged investment and d) the series of the real geometric
mean of the leveraged investment.
In the next subsection, I present the results of the experiment.
3.2.3
Results
In this subsection, I present the results of the experiment. A single iteration only
produces four series consisting of one data point: a) proxy values of the geometric
mean of unleveraged investment, b) the real geometric mean of the unleveraged
investment, c) the proxy of the geometric mean of the leveraged investment and d)
the series of the real geometric mean of the leveraged investment.
This single iteration is obviously not sufficient to test the hypothesis that the
difference of the proxy of the geometric mean and the real geometric mean of the
3.2. Experiment
47
25
Level of Leverage = 2.42
Cumulative Equity Returns
20
15
10
5
0
Unleveraged
500
1000
1500
2000
2500
Figure 3.4: The cumulative equity returns of leveraged and unleveraged investment
in of UNH between 17/01/2001 and 04/08/2011.
unleveraged and leveraged investment are zero because of the lack of data per series. Therefore, this iteration is replicated 1000 times in order to generate 1000
data points per series. Thus, after the replication, there are four different series of
data with 1000 data points: a) the series of the proxy values of the geometric mean
of unleveraged investment, b) the series of the real geometric mean of the unleveraged investment, c) the series of the proxy of the geometric mean of the leveraged
investment and d) the series of the real geometric mean of the leveraged investment.
These series are used in the two different tests. In the first test, I attempt to
reject the null hypothesis that the average value of the difference between the series
of proxy and real geometric mean of the unleveraged investment has average value
of zero. In the second test, I attempt to reject the null hypothesis that the average
value of the difference between the series of proxy and real geometric mean of the
leveraged investment has average value of zero.
48
Real
Proxy
Table 3.1: Statistical analysis of the series of the proxy and the real geometric mean
of the unleveraged investments.
Item
P-value
Confidence Interval (Prob95%)
Null Hypothesis
GM GM = 0
0.0553
[-2.90e-08; 2.58e-06]
Accepted
Table 3.2: Statistical test of the difference between the proxy of the geometric mean
and the real geometric mean of unleveraged investments.
Test for Unleveraged Investments
From Table 3.1, the difference of the mean and the median of the proxy and the
real geometric mean of the unleveraged investments are not significant at the first
glance, which could indicate that there is no difference between the series. From
Table 3.2, the hypothesis that the difference between proxy if the geometric means
and the real geometric mean is zero is accepted with a probability of 95%. From
Figure 3.4 and Figure 3.5, it is possible to see that the values of the proxy and
the real geometric mean do not differ substantially. Finally, from Figure 3.6, it is
possible to notice that the average value of both series are very similar.
3.2. Experiment
49
Real
Geometric Mean
x 10
0
2
4
0
200
400
Geometric Mean
800
1000
600
800
1000
Proxy
600
x 10
0
2
4
0
200
400
Iterations
Figure 3.5: Plot of the series of the proxy and the real geometric mean of the
unleveraged investments.
x 10
3
2
1
0
1
2
3
4
0
200
400
600
800
1000
Iterations
Figure 3.6: Plot of the difference between the proxy and the real geometric mean of
the unleveraged investments.
50
x 10
Geometric Mean
1.5
0.5
0.5
Real
Proxy
Figure 3.7: The boxplot of the proxy of the geometric mean and the real geometric
mean of the returns of unleveraged investments.
Real
Geometric Mean
0.5
0
0.5
1
0
200
400
600
800
1000
600
800
1000
Geometric Mean
Proxy
0
0.05
0.1
0.15
0.2
0
200
400
Iterations
Figure 3.8: Plot of the series of the proxy and the real geometric mean of the
leveraged investments.
3.2. Experiment
51
0.4
0.2
0
0.2
0.4
0.6
0.8
1
0
200
400
600
800
1000
Iterations
Figure 3.9: Plot of the difference of the series of the proxy and the real geometric
mean of the leveraged investments.
0
0.1
Geometric Mean
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Real
Proxy
Figure 3.10: The boxplot of the proxy of the geometric mean and the real geometric
mean of the returns of leveraged investments.
52
Real
-0.3603
-0.0063
-1.0000
0.0041 0.0043
0.4780
-0.5921
1.3510
103
Proxy
-0.0104
-0.0048
-0.1811
0.0176
-4.5300
33.4113
103
Table 3.3: Statistical analysis of the series of the proxy and the real geometric mean
of the leveraged investments.
Item
P-value
Confidence Interval (Prob95%)
Null Hypothesis
GM (l) GM (l) = 0
0
[-0.3802; -0.3221]
Rejected
Table 3.4: Statistical Test of the difference between the proxy of the geometric mean
and the real geometric mean of leveraged investments.
Figure 3.8, it is possible to see that the values of the proxy and the real geometric
mean differ substantially from each other. Furthermore, is it possible to see the
considerable amount of cases that the geometric mean of the equity return is equal
to -100%. Finally, from Figure 3.9, it is observed that the boxplot of the real
geometric mean presents values of -1. This occurs because of a number of series
generated present a geometric mean of the equity return equal to -100%.
In the next section, I discuss the results of this experiment.
3.3
Discussion
In this section, I discuss the results of the experiment in Section 3.3. The proxy
of geometric mean is a straightforward method to relate the geometric mean to the
moments of the returns because it uses the arithmetic average and the variance. It
has been widely used in the literature in order to obtain a proxy of the stock rate
of growth. In this chapter, I demonstrate that the proxy is accepted to estimate
for this purpose. One of the advantages of the use of the proxy is that this allows
the calculation of the rate of the rate of growth of the stock using a closed-form
3.4. Summary
53
solution.
However, the stock return is only one of the parameters of the equity return.
The other parameter is the level of leverage. In this experiment, I also tested the
hypothesis that proxy can also be used to estimate the rate of growth of the equity
return. The result shows that the proxy should not be used to estimate the leveraged
equity return.
In Figure 3.7, we observe that there is a considerable number of leveraged investments that result in a geometric mean value of -100% and this fact is not captured
by equation 2.5. This means that even though the investor is ruined, the proxy of
the geometric mean shows that this investor is active.
In the next section, I present the summary and the contribution of this chapter.
3.4
Summary
In this chapter, I reviewed the use of the formula proposed as a proxy of the geometric mean of the historical financial returns in light of the leveraged investment.
The formula contains the value of the arithmetic average and the variance of the
series of the equity returns. This formula is widely used in the literature because it
facilitates a closed-form solution of the analytical model.
However, leverage has the ability to expand the equity return and, consequently,
increase extreme returns. Since the formula assumes normal distribution of the
series of financial returns, this underestimates extreme returns and, consequently,
underestimates the risk of default.
The contribution of this chapter is to show that a commonly-used proxy of the
geometric mean, which is widely used in the current literature, fails to accurately
estimate leveraged returns. This proxy uses the arithmetic average and the variance
of the stock returns (Parramore and Watsham 1997). However, this proxy assumes
that returns can be characterised by the first two moments and ignores the fat-tailed
return distributions that are observed in actual empirical data (Cont 2001).
54
equity returns in a similar way to that used to calculate the geometric mean of
unleveraged equity returns. The result is that, even though the proxy serves to
calculate the geometric mean of unleveraged equity returns, it does not serves to
calculate the geometric mean of leveraged equity returns because the proxy does
not consider the risk of default. For this reason, the proxy should not be used to
calculate the geometric mean of the series of leveraged returns.
The proxy of the geometric mean was an important approximation because before
the use of computer software, it was very difficult to work with numerical methods.
Thus, the closed-form solution was the best form to present and defend a theory in
the quantitative finance area. However, with the advance of computer software, the
use of numerical methods increased. Thus, numerical methods could also be used
to estimate the geometric mean which minimises the issues from the assumption of
normal distribution of the financial returns, necessary to use the proxy.
In the next chapter, I introduce a novel trading strategy which uses numerical
methods in order to calculate the optimal level of leverage.
Chapter 4
Optimal Level of Leverage using
Numerical Methods
In this chapter, I introduce a new model that uses numerical methods in order to
find the level of leverage that maximises the geometric mean of a historical equity
returns. The reason for this proposition is that the existing analytical models to
optimize the level of leverage in the literature assume the returns are normally
distributed. My insight is that, by assuming normal distribution of the returns,
these models are not able to capture the presence of the fat tails on the return
distribution (Cont 2001) and, consequently, their results could not represent the
real optimal level of leverage. However, by using numerical methods, I can relax
the assumption of normal distribution of the returns. Thus, when calculating the
optimal level of leverage, I can take into account the presence of the fat tails in
the return distribution. To the best of my knowledge this is the first algorithm for
optimising the level of leverage using the geometric mean which does not assume
that returns are normally distributed.
In order to test the ability of numerical methods to yield superior geometric
mean to analytical models, I run an experiment using numerical methods and the
most recent analytical models in the literature in the literature: the model of Peters
(2010) which assumes that the price follows a geometric Brownian motion. I find
56
that the use of numerical methods results in superior estimation of the geometric
mean compared with geometric Brownian motion, even though, it results in a inferior
optimal level of leverage.
The chapter is structured as follows: In Section 4.1, I present the introduction
of the chapter. In Section 4.2, I describe the model in which the optimal level of
leverage is calculated using numerical methods. In Section 4.3, I present the results
of the experiment that compares two different methods to calculate the optimal level
of leverage, geometric Brownian motion and numerical methods. In Section 4.4, I
discuss the results of the experiment. Finally, in Section 4.5, I present the summary
and the contribution of the chapter.
4.1
Introduction
In Section 2.3, I demonstrated that there are two different approaches of the investors objective: the mean-variance approach and the geometric mean approach.
I also demonstrated that leverage is monotonic and linear in the risk-return formula,
but is neither monotonic nor linear in the geometric mean. For this reason, I argue
that the geometric mean formula is the appropriate model to measure the investment performance whether the investor uses leverage. Thus, the geometric mean
formula should be used to calculate the optimal level of leverage.
In Section 2.5, I described the three different models to optimize the level of
leverage in the literature. These models propose to find the level of leverage that
maximises the geometric mean of the historical leveraged returns. The first model is
the Kelly Criterion. Assuming that the investors objective is to maximise the end
of period wealth, Kelly (1956) demonstrates that for binomial games with positive
expectation, there is one specific value of leverage that maximises the geometric
mean of the returns or the end of period wealth. He introduced a model to obtain
this value called the Kelly criterion. Today, one of the possible uses of the Kelly
criterion is to determine the optimal level of leverage on investments considering
4.1. Introduction
57
the possible returns in different discrete scenarios. However, the Kelly criterion
cannot be properly used to determine the optimal level of leverage in continuous
games where the number of possible outcomes is unlimited, as with investments in
financial markets.
The second model is the Rotando and Thorp model. In order to relax the restriction of discrete scenarios with limited number of possible outcomes, Rotando and
Thorp (1992) revise the Kelly criterion for continuous gambling games. However,
the use of Kelly criterion for continuous gambling games is complicated. Continuous
games assume an infinite number of outcomes and the Kelly criterion was originally
designed to assume a limited number of outcomes. Hence, in order to limit the number of outcomes and respect the Kelly criterion assumptions, Rotando and Thorp
assumed that the returns in the financial market follow a normal distribution and
such distribution should be transformed to quasi-normal distribution by cutting off
the tails. This allows the use of the Kelly criterion in continuous games. However,
an important issue remains, namely the assumptions that the returns are normally
distributed.
Finally, the third model is the Peters model. In order to establish a more realistic approach, Peters (2010) introduces a new model to obtain the optimal level
of leverage. He assumes that asset prices follow a geometric Brownian motion and
demonstrates that the optimal level of leverage is simply the estimation of return divided by its variance. However, by assuming that prices follow a geometric Brownian
motion, Peters also assumes that returns are normally distributed.
The key point is that all three models assume that returns are normally distributed. Cont (2001) demonstrates that there is a presence of fat-tails on daily
returns, i.e, value of kurtosis superior to three. Hence, we see from empirical data
that actual returns in the real market are not normally distributed. Therefore, models that assume the normality of the returns distributions, underestimate extreme
returns and, consequently, underestimate the risk of default.
In this chapter, in order to avoid any discussion about the return distribution and
58
relax the assumption of normal distribution of the returns, I propose to estimate the
optimal level of leverage using a numerical non-parametric method. In this case, I
use a line-search method to obtain the level of leverage that maximises the geometric
mean of a series of historical returns.
In order to test the hypothesis that numerical methods yield superior geometric
mean to the Peters model, I calculate the optimal level of leverage using geometric
Brownian motion and compare this with the optimal level of leverage obtained using
numerical methods of 104 series of daily returns of the components of the S&P 500
index. In the experiment, the stock, the length of the series and the period are
randomly selected. I find that the numerical method shows superior geometric
mean to the Peters model based on the geometric Brownian motion.
In the next section, I describe the model optimal level of leverage using numerical
methods.
4.2
Numerical Methods
In this section, I describe the model in which the optimal level of leverage is calculated using numerical methods. In general, the models reviewed in Section 2.5
present the following three characteristics of leverage. Firstly, the geometric mean is
concave in leverage, and consequently, there is an optimal level of leverage. Secondly,
the optimal level of leverage is achieved analytically using those models. Finally,
the models assume the normal distributions of the returns and ignores the presence
of fat tails (Cont 2001).
The proposal of this section is that, instead of using proxies and analytical models
to obtain the optimal level of leverage, we should return to the original model of
geometric mean as described in Eq. (2.13) and use a numerical method, in this case
line-search, in order to obtain the optimal level of leverage. By so doing I am able
to relax the assumption of normal distribution of the returns.
In the line-search method (Kelley 1999), I randomly select two levels of leverage,
4.3. Experiment
59
l1 and l2 , in which each point is on a different side of the maximum value. Next, I
select two new levels of leverage, l3 and l4 . This will increase the geometric mean
value. This procedure is performed iteratively to obtain the specific level of leverage
that no other level of leverage results in a superior value of geometric mean. The
resulting level of leverage is considered the optimal.
In order to use the numerical method, I transform Eq. (2.13) and introduce the
term leverage in the geometric mean formula:
n
Y
1
GMnm (l) = [ (1 + rt l)] n 1
(4.1)
t=1
where rt l is the leveraged return at the time t. Iteratively, I use the line-search
method, in order to obtain the value of l that maximises Eq. (4.1):
n
Y
1
Max GMnm (l) = Max[ (1 + rt l)] n 1
l
(4.2)
t=1
4.3
Experiment
In this section, I present the results of the experiment that compares two different
methods to calculate the optimal level of leverage, geometric Brownian motion and
numerical methods. This section is divided in two subsections. In the first subsection, I describe the methodology of the experiment. And, in the second subsection,
60
4.3.1
Methodology
4.3. Experiment
4.3.2
61
Results
In this subsection, I present the results of the experiment. In order to compare the
methods and their results, I test two different null hypotheses. Firstly, I test the
null hypothesis that the optimal level of leverage using numerical method is similar
to the optimal level of leverage using on geometric Brownian motion. Secondly, I
test the null hypothesis that the leveraged geometric mean using numerical methods
is similar to the leveraged geometric mean using on geometric Brownian motion.
Geometric Mean
Item
GMN M
GMGBM
Mean
3.9004x104
3.8932x104
Median
2.0925x104
2.0922x104
Minimum
-3.9415x104
Maximum
0.0073
0.0073
5.6597x104
5.6463x104
Skewness
3.8271
3.8207
Kurtosis
25.5245
25.4607
104
104
Standard Deviation
Size
Table 4.1: Statistical analysis of the geometric mean of using the geometric Brownian motion approach and the numerical methods to calculate the optimal level of
leverage.
From Table 4.1, the mean and the median of the series of geometric means calculated
using the numerical methods are superior to the series of geometric means using the
geometric Brownian motion. It indicates that the numerical methods could yield to
a higher value of geometric mean than the geometric Brownian motion.
62
GMN M GMGBM
Item
Confidence Interval (P95%)
[1.66x104 ;2.16x104 ]
Null Hypothesis
Rejected
Table 4.2: Comparison between geometric Brownian motion approach and numerical
methods of the geometric mean using the optimal level of leverage .
From Table 4.2, the confidence interval of the difference between the series of
geometric means calculated using the numerical methods and the series of geometric
means using the geometric Brownian motion is positive. This demonstrates that
numerical methods yield to a higher value of geometric mean than the geometric
Brownian motion with a probability superior to 97.5%.
Numerical Methods
Geometric Mean
x 10
6
4
2
0
0
2000
Geometric Mean
6000
8000
10000
8000
10000
4000
x 10
6
4
2
0
2
0
2000
4000
6000
Iterations
Figure 4.1: The plot of the geometric means calculated using numerical methods
and geometric Brownian motion.
From Figure 4.1, there is no negative value of the geometric mean of the historical
returns of the use of numerical methods. However, there is one negative value for
the geometric mean of the use of the geometric Brownian motion.
4.3. Experiment
63
4
Geometric Mean
15
x 10
10
5
0
Numerical Methods
Figure 4.3: The boxplot of the geometric mean using numerical methods and based
on the geometric Brownian motion.
10
x 10
2
0
2000
4000
6000
Iterations
8000
10000
Figure 4.2: The plot of the difference of the geometric means between numerical
methods and geometric Brownian motion.
From Figure 4.2, there is no negative difference between the series of geometric
means calculated using the numerical methods and the series of geometric means
using the geometric Brownian motion. This demonstrates that numerical methods
yield a value of geometric mean superior to geometric Brownian motion.
From Figure 4.3, we see that the difference between the series of geometric means
64
calculated using the numerical methods and the series of geometric means using the
geometric Brownian motion is not clear. However the bottom value of the numerical
methods is superior to the minimum value of the geometric Brownian motion.
Level of Leverage
Item
LN M
LGBM
Mean
0.9621
0.9639
Median
0.7903
0.7924
Minimum
-4.7141 -4.7611
Maximum
8.0387
9.0770
Standard Deviation
1.0202
1.0292
Skewness
1.6113
1.6790
Kurtosis
8.7796
9.2992
104
104
Size
Table 4.3: Statistical analysis of the optimal level of leverage of the use of the
geometric Brownian motion approach and the numerical methods.
From Table 4.3, the mean and the median of the series of the optimal level of leverage calculated using the numerical methods are inferior to the series of the optimal
level of leverage using the geometric Brownian motion. This indicates that numerical methods could yield a lower value of optimal level of leverage than geometric
Brownian motion.
Item
Confidence Interval (P95%)
Null Hypothesis
LN M LGBM
[0.1934; 0.2428]
Rejected
Table 4.4: Comparison between geometric Brownian motion approach and numerical
methods of the optimal level of leverage.
From Table 4.4, the confidence interval of the difference between the series of op-
4.3. Experiment
65
timal level of leverage calculated using numerical methods and the series of optimal
levels of leverage using geometric Brownian motion is positive. This demonstrates
that numerical methods yield to a lower value of the optimal level of leverage than
geometric Brownian motion with a probability superior to 97.5%.
Numerical Methods
Geometric Mean
10
5
0
5
0
2000
4000
6000
8000
10000
8000
10000
10
5
0
5
0
2000
4000
6000
Iterations
Figure 4.4: The plot of the level of leverage calculated using numerical methods and
geometric Brownian motion.
From Figure 4.4, the range of the possible values of the optimal level of leverage varies between -5 and 10 in both approaches, numerical methods and geometric
Brownian motion. This demonstrates that the excessive levels of leverage, e.g. tenfold, are not rational because it has not been between the range of the level of
leverage that maximises the historical equity returns.
66
0.5
0.5
1.5
2
0
2000
4000
6000
Iterations
8000
10000
Figure 4.5: The plot of the difference of the level of leverages between numerical
methods and geometric Brownian motion.
From Figure 4.5, the difference between the series of the optimal level of leverage
calculated using numerical methods and the series of geometric means using geometric Brownian motion is distributed among positive and negative values. Hence,
visually, it is not possible to distinguish the optimal level of leverage using numerical
methods to the optimal level of leverage using geometric Brownian motion.
4.4. Discussion
67
3.5
3
Level of Leverage
2.5
2
1.5
1
0.5
0
0.5
1
1.5
Numerical Methods
Figure 4.6: The boxplot of the level of leverage using numerical methods and based
on the geometric Brownian motion.
From Figure 4.6, the difference between the series of geometric means calculated
using the numerical methods and the series of geometric means using the geometric
Brownian motion is not clear.
In the next section, I discuss the results of this experiment.
4.4
Discussion
The results of the experiment show that the use of numerical methods allows us
to calculate the optimal level of leverage that results in superior estimation of the
geometric mean of the historical returns compared with the use of geometric Brownian motion. The confidence interval of the difference of geometric mean of the
daily historical returns is between 1.66x104 and 2.16x104 . By using the formula
(1 + dgm )252 1, where the parameter dgm can be replaced by the value of the limits
of the confidence interval, this values represent a difference in the annual historical
returns of 4.27% and 5.59%.
68
are normally distributed and ignores the presence of fat-tails. Consequently, this
underestimates extreme returns. However, numerical methods relax the assumption
of normal distribution of the returns, and do not require a closed-form solution.
This makes the calculation of the optimal level of leverage more reliable, and allows
me to obtain the correct level of leverage that maximises the geometric mean of the
historical return.
In the next section, I present the summary and the contribution of this chapter.
4.5
Summary
In this chapter, I introduced a new model that uses numerical method in order
to find the level of leverage that maximises the geometric mean of the series of
historical equity returns. My method uses line-search to numerically search for the
optimal level of leverage. To the best of my knowledge this is the first algorithm for
calculating the optimal level of leverage using the geometric mean approach which
does not assume that returns are normally distributed or that the price process
follows geometric Brownian motion.
Furthermore, I tested the ability of numerical methods to yield superior geometric mean to geometric Brownian motion. In order to do that, I calculated the
optimal level of leverage using numerical methods and using geometric Brownian
motion of 104 series of daily returns of the components of the S&P 500 Index. In
the experiment, the stock, the length of the series and the window period are randomly selected. I found that the numerical methods yield superior geometric mean
of the series of historical equity returns to geometric Brownian motion, even though,
when it shows inferior level of leverage.
The results of the experiments in Chapter 3 and in this chapter show the advances
caused by the use of numerical methods in the leverage models. For this reason, I
argue that the proxy and the analytical models used to calculate the unleveraged
4.5. Summary
69
equity returns cannot be directly used to calculate the leveraged equity returns
because the proxy and the models assume that the returns are normally distributed,
and ignore the presence of fat-tails and, consequently, underestimate extreme returns
present on the leveraged investments.
The contribution of this chapter is the introduction of a novel trading strategy
which uses numerical methods in order to calculate the optimal level of leverage.
Previous studies have proposed analytical models in order to calculate the optimal level of leverage. However, these models assume that returns are normally
distributed (Kelly 1952; Rotando and Thorp 1992, and Peters 2010). By using numerical methods in order to optimise the level of leverage, I am able to avoid this
assumption and hence avoid under-estimating extreme returns (Cont 2001).
In the next chapter, I show that my trading strategy exhibits superior performance to a strategy which does not use leverage, under an analysis which uses
historical data from all components of the S&P 500 index.
Chapter 5
A Back-testing of the Optimal
Level of Leverage using Numerical
Methods
In this chapter, I back-test the optimal level of leverage using numerical methods. Using empirical historical data from the components of the S&P 500 index, I
compare the investment performance of an optimally leveraged investment to the
geometric Brownian motion (Peters 2010), the unleveraged investment and the excessively leveraged investment. I find that the optimally leveraged investment outperforms the other three leveraged investment approaches. This is interesting to
back-test the effect of leverage using empirical data because it reinforce the ability of the optimal level of leverage using numerical methods to improve investment
performance.
The chapter is structured as follows: In Section 5.1, I present the introduction
of the chapter. In Section 5.2, the experiment is presented. In Section 5.3, I discuss
the results of the experiment. Finally, in Section 5.4, I present the summary and
the contribution of the chapter.
72
5.1
Introduction
In this section I present the introduction of this chapter. The first question that
emerges from every financial theory or financial model is: Does it work?. One
possible way to answer this question is through a back-testing of the model using the
historical data. In this chapter, in order to answer the question Does the optimal
level of leverage using numerical methods work?, I back-test the ability of this
model to improve the investment performance using the historical data.
According to Williams (1936) and Latane (1959), in a multi-period framework,
the investors objective is to maximise the end of period equity return, or the geometric mean. In Chapter 3, I demonstrated that the equity return formula has
two components: the stock return and the level of leverage. Furthermore, I demonstrated that the level of leverage is the impact of the stock return in the equity
return. Assuming that the stock return is out of the agents control (being determined by the market), in Chapter 4, I proposed a numerical method in order to find
the level of leverage that maximises the geometric mean of the equity returns of an
investment in a single stock.
In this chapter, I back-test the numerical method proposed in Chapter 4. In
order to do that, I test the ability of the use of optimally leveraged investment
to yield superior end of period return to geometric Brownian motion, unleveraged
investments and excessively leveraged investments. In the test, I use daily data of
every component of the S&P 500 index between January 2000 and December 2010
and compare four series of 1000 data points: The series contain the geometric means
of the unleveraged investments, the optimally leveraged investments, the geometric
Brownian motion leveraged investments and the excessively leveraged investment.
There are three results from the experiment. The first result is that the optimally leveraged investments results in a superior end of period equity return to
the geometric Bronian motion. The end of period equity return of the optimally
leveraged investments is approximately between 3.9 and 6.4 times the end of period
equity return of the unleveraged investments with a statistical probability of 95%.
5.2. Experiment
73
The second result is that the optimally leveraged investments results in a superior
end of period equity return to the unleveraged investments. The end of period
equity return of the optimally leveraged investments is approximately between 3.9
and 6.4 times the end of period equity return of the unleveraged investments with
a statistical probability of 95%.
The third result is that the optimally leveraged investments results in a superior
end of period equity return to the excessively leveraged investments. Considering
that the average value of the end of period equity return of the excessively leveraged
investments is -100%, i.e., default, it is not possible to calculate the improvement
caused by the use of optimal level of leverage using numerical methods. Even though,
the improvement exists because any result is better than default.
In the next section, I present the results of the experiment in which I run a
back-test of the optimal level of leverage using numerical methods.
5.2
Experiment
In this section, I back-test the ability of this model to improve the investment performance compared to the geomoetric Brownian motion, the unleveraged investment
and the excessively leveraged investment. This section is divided in three subsection. In the first subsection, I describe the methodology of the experiment. In the
second subsection, I present the example of a single iteration. Finally, in the third
section, I present the results of the experiment.
5.2.1
Methodology
t=1
74
GM (kt )
n
Y
1
= [ (1 + kt )] n 1
(5.1)
t=1
where k is the equity return. Substituting Eq. (2.21) in Eq. (5.1), the geometric
mean of the historical equity return can be expressed in terms of leverage:
n
Y
1
GM (l ) = [ (1 + rt l )] n 1
(5.2)
t=1
where l is the level of leverage that maximises the historical geometric mean and
rt is the stock return on the period t. Thus, from Eq. (5.2) the formula to calculate
the initial optimal level of leverage at time 1000 is:
5.2. Experiment
75
GM (l1 )
1000
Y
(1 + (rt l1 )] 1000 1
=[
(5.3)
t=1
Secondly, the second portion serves to obtain the back-tested investment performance, i.e., the end of period equity return. The optimal level of leverage is
calculated using the previous 1000 days stock returns at each period of the backtesting. Iteratively, this calculation is made over 250 periods. The result is the end
of period equity return:
250
Y
EQ (li , ri ) = [ (1 + li ri+1000 )] 1
(5.4)
i=1
ligbm =
i
i2
(5.5)
where lip is the ideal level of leverage using geometric Brownian motion, i and i2
are the average value and the variance of the series of the previous 1000 returns,
respectively. Thus, from Eq. (5.5) the formula to calculate the initial ideal level of
leverage using geometric Brownian motion at time 1000 is:
l1gbm =
1
12
(5.6)
Consequently, the model for the geometric Brownian motion leveraged investment is:
250
Y
EQgbm (li , ri ) = [ (1 + ligbm ri+1000 )] 1
(5.7)
i=1
where EQgbm is the end of period return of the geometric Brownian motion leveraged
investment.
Furthermore, I assume that the level of leverage of unleveraged investments is
76
one, li = 1:
250
Y
EQ1 (1, ri ) = [ (1 + ri+1000 )] 1
(5.8)
i=1
where EQ1 is the end of period return of the unleveraged investment; and I also
assume that the level of leverage of unleveraged investments is ten, li = 10:
250
Y
EQ10 (10, ri ) = [ (1 + 10ri+1000 )] 1
(5.9)
i=1
where EQ10 is the end of period return of the excessively leveraged investments.
In the next subsection, I describe the example of a single iteration.
5.2.2
Single Iteration
5.2. Experiment
77
50
45
40
Price
35
30
25
20
15
10
5
0
200
400
600
t
800
1000
1200
Figure 5.1: The series of prices of NDAQ between 01/11/2002 and 19/10/2007.
0.3
0.25
0.2
Stock Return
0.15
0.1
0.05
0
0.05
0.1
0.15
0.2
0
200
400
600
t
800
1000
1200
Figure 5.2: The series of returns of NDAQ between 01/11/2002 and 19/10/2007.
78
3
Geometric Brownian motion
Optimal level of leverage
2.8
Level of Leverage
2.6
2.4
2.2
2
1.8
1.6
0
50
100
150
200
250
Figure 5.3: The series of the optimal level of leverage and of the geometric Brownian motion level of leverage of an investment in NDAQ between 01/11/2002 and
19/10/2007.
5.2. Experiment
79
50
100
150
Geometric Brownian Motion
200
250
50
100
150
Excessive Level of Leverage
200
250
50
100
150
Unleverage
200
250
50
100
200
250
1
0
1
0
1
0
1
0
0.5
0
0.5
0
150
Figure 5.4: The series of cumulative equity return of an investment in NDAQ between 01/11/2002 and 19/10/2007 considering three different approaches of leverage:
unleveraged, optimally leveraged, and excessive leveraged.
motion leveraged investment, unleveraged investment and excessively leveraged investment are +8.17%, +7.46%, +3.38% + 36.5%, respectively.
Applying similar calculations until the end of the period, day 1250, which is,
in fact, day 250 of the back-testing (the first 1000 days is used to calculate the
optimal level of leverage), on 19/10/2007, this generates four different series of
the end of period equity returns shown in Figure 5.4, the series of the end of period equity return of the optimally leveraged, geometric Bronian motion leveraged,
unleveraged and excessively leveraged investment, respectively. As a result, it is
possible to collect four different metrics: the end of period equity return of the optimally leveraged, the geometric Bronian motion leveraged, the unleveraged and the
excessively leveraged investment that are +28.7%, +25.9%, +10.3% and -100%, respectively. In this example, there is three differenct results. Firstly, the investment
performance of optimally leveraged investments is superior to the investment performance of unleveraged investment and the improvement is approximately 1.10 times
(+0.287/ + 0.259). Secondly, the investment performance of optimally leveraged
80
10
8
6
4
2
0
2
4
6
0
50
100
150
200
250
5.2.3
Results
In this subsection, I present the results of the experiment. A single iteration only
provides three metrics of interest, the end of period equity returns of the optimally
leveraged, the geometric Brownian motion leveraged, the unleveraged, and the excessively leveraged investment. This single iteration is obviously not sufficient to
test the ability of the optimal level of leverage using numerical methods to improve
the investment performance because it generates four different series of end of period
equity returns with only one single data point. Therefore, in this experiment, this
single iteration is replicated 1000 times with random stocks and window periods.
Figure 5.5 shows the leverage dynamic after 1000 iterations. Note that there is
5.2. Experiment
81
10
8
Average Value
6
4
2
0
2
4
0
200
400
600
800
1000
Iterations
Figure 5.6: Plot of the average value of the optimal level of leverage per iteration.
82
Table 5.1: Statistical analysis of the log of the end of period equity return of unleveraged ln(EoP (1)), geoemtric Brownian motion leverage ln(EoP (lgbm )), optimally
leveraged ln(EoP (l )) and excessively leveraged investment ln(EoP (10)).
the difference between the series of the end of period equity returns of the optimally
leveraged and the geometric Brownian motion leverage investment, the difference
between the series of the end of period equity returns of the optimally leveraged and
the excessively leveraged investment, and the difference between the series of the end
of period equity returns of the optimally leveraged and the unleveraged investment.
These tests are in order to test the null hypothesis that the optimally leveraged
investment shows end of period equity returns equal to geometric Brownian motion
leverage investment, the optimally leveraged investment shows end of period equity
returns equal to excessively leveraged investment, and to test the null hypothesis
that the optimally leveraged investment shows end of period equity returns equal
to unleveraged investment.
From Table 5.1, the mean of the log of the end of period equity return of the
unleveraged, the geoemtric Brownian motion leveraged, the optimally leveraged and
the excessively leveraged investment are +0.0311, +0.1426, +0.1543 and -16.94, respectively. Furthermore, the median of the geometric mean of the equity return
of the unleveraged, the geoemtric Brownian motion leveraged, the optimally leveraged and the excessively leveraged investment are +0.0732, +0.1523, +0.1662 and
-4.1145, respectively. The values of the mean and the median of different use of leverage indicate that the optimal leveraged investment shows the greatest performance
among the approaches.
5.2. Experiment
Item
GM (l ) = GM (lgbm ) GM (l ) = GM (1)
P-value
0.0024
1.1465x1010
Confidence [+0.0012; +0.0057] [+0.0878; +0.1586]
Interval
(P95%)
Null HyRejected
Rejected
pothesis
83
GM (l ) = GM (10)
0
[+14.8602 ; +19.3474]
Rejected
Table 5.2: T-Test of the difference between the log of the end of period equity return of optimally leveraged investment and geotmric Brownian motion leveraged
investment (GM (l ) = GM (lgbm )); the difference between the log of the end of period equity return of optimally leveraged investment and unleveraged investment
(GM (l ) = GM (1)); and the difference between the log of the end of period equity return of optimally leveraged investment and excessive leveraged investment
(GM (l ) = GM (1)).
From Table 5.2, the null hypothesis that the difference between the end of period
equity returns of the optimally leveraged investments and the excessively leveraged
investments is null is rejected with a probability of 95%; and the null hypothesis
that the difference between the end of period equity returns of the optimally leveraged investments and the unleveraged investments is null is also rejected with a
probability of 95%. This fact, combined with the confidence interval with positive
values in both tests, demonstrates that the use of optimal level of leverage results
in a statistically-significant improvement in investment performance.
Figure 5.7 shows the boxplot of the series of the log of the end of period equity
return. From this figure, note that the performance of the optimally leveraged is
superior to the geometric Brownian motion leveraged, unleveraged and excessively
leveraged investment.
From the experiment, it is possible to obtain three different results. Firstly, the
average value of the log of the end of period equity returns of geometric Brownian
motion leveraged investments is 14.26%. From Table 5.2, the confidence interval of
the difference between the optimally leveraged investments and geometric Brownian
motion leveraged investments is between +0.0012 and +0.0057, which indicates
interval of the estimated improvement between +0.84% and +4.00%. Therefore,
the improvement of the use of the optimally leveraged investments is aproximately
84
1.5
1.5
10
0.5
0.5
10
20
0
0
30
0.5
1 Opt
0.5
GBM
40
Opt
Unl
50
Opt
Exc
Figure 5.7: The boxplot of the log of the end of period equity return of the optimally
leveraged vs. the geometric Brownian motion leveraged investment; the boxplot
of the log of the end of period equity return of the optimally leveraged vs. the
unleveraged investment; and the boxplot of the log of the end of period equity
return of the optimally leveraged vs. the excessively levearaged investment.
5.3. Discussion
85
5.3
Discussion
In this section, I discuss the result of the experiment in Section 5.2. Leverage is
concave in the geometric mean of the historical returns (see chapter 3), hence, there
is some level of leverage that maximises the geometric mean. In Section 4.4, I
proposed a numerical method in order to find optimal level of leverage. I assumed
that the main objective of the investor is to maximise the end of period equity return
(Williams 1936; Latane 1959).
In Section 5.2, I back-tested the ability of an investment employing the optimal
level of leverage using numerical methods to show a performance superior to geometric Brownian motion leveraged, unleveraged and excessively leveraged investment,
i.e., to show end of period of the equity returns superior to the geometric Brownian
motion leveraged, unleveraged and excessively leveraged investments.
There are three different results from the back-testing. The first result is that
optimal level of leverage results in a superior investment performance to geometric
Brownian motion leveraged investments. Furthermore, considering that the average
value of the end of period equity return of the geometric Brownian motion leveraged
investments is +14.26%, and that the confidence interval of the difference between
the end of period equity returns of optimally leveraged and geometric Brownian
motion leveraged investments is between +0.84% and +4.00%; the end of period
equity return of the optimally leveraged investments is approximately between 1.1
and 1.3 times the end of period equity return of the geometric Brownian motion
leveraged investments with a statistical probability of 95%.
The second result is that optimal level of leverage results in a superior investment
performance to unleveraged investments. Furthermore, considering that the average
86
value of the end of period equity return of the unleveraged investments is +3.16%,
and that the confidence interval of the difference between the end of period equity
returns of optimally leveraged and unleveraged investments is between +9.18% and
+17.19%; the end of period equity return of the optimally leveraged investments
is approximately between 3.9 and 6.4 times the end of period equity return of the
unleveraged investments with a statistical probability of 95%.
Finally, The third result is that optimal level of leverage results in a superior
investment performance to excessively leveraged investments. Considering that the
average value of the end of period equity return of the excessively leveraged investments is -100%, i.e., default, it is not possible to calculate the improvement
caused by the use of optimal level of leverage using numerical methods. However,
the improvement exists because any result is better than default.
Therefore, the intelligent use of leverage cannot be ignored because the optimal level of leverage using numerical methods result in a superior performance to
geometric Brownian motion leveraged, unleveraged and excessively leveraged investments.
In the next section, I present the summary and the contribution of this chapter.
5.4
Summary
In this chapter, I back-tested the optimal level of leverage using numerical methods.
The advantage of the back-testing is the use empirical evidence to test the ability
of the optimal level of leverage using numerical methods to improve the investment
performance. I compared the investment performance of the optimal leveraged investments to the geometric Brownian motion leveraged, the unleveraged and the
excessively leveraged investments. I assumed that the excessive level of leverage is
the level of leverage of ten because, in Figure 5.6, there was no average value of
optimal level of leverage which is superior to this value, ten.
I randomly selected 103 times, one component of the S&P 500 Index and one
5.4. Summary
87
window period of 1250 days between 2000 and 2011. The window period is split
in two parts, one part with 1000 data and another part with 250 data. The first
part serves to calculate the optimal level of leverage and the second part serves to
back-test the use of the optimal level of leverage calculated using numerical methods.
There are three different results from the experiment. Firstly, the performance
after employing the optimal level of leverage using numerical methods is approximately between 1.1 and 1.3 times the performance of the geometric Brownian motion
leveraged investment, with a probability of 95%. Secondly, the performance after
employing the optimal level of leverage using numerical methods is approximately
between 3.9 and 6.4 times the performance of the unleveraged investment, with a
probability of 95%, that is, the investor who uses the optimal level of leverage has
the expectation to triplicate or more their equity returns compared to unleveraged
investment. Finally, thirdly, the use of the optimal level of leverage using numerical
methods is superior to the excessively leveraged investment because the usage of the
excessively leveraged investment results in default.
The results of the back-testing show that, compared to geometric Brownian
motion leveraged, unleveraged and excessively leveraged investment, the optimal
level of leverage using numerical methods improves the investment performance.
Thus, the appropriate level of leverage is an important parameter to be taken into
account when the objective is to maximise the end of period equity return.
The contribution of this chapter is to show that the model proposed in Chapter 4
allows to reach to a superior performance to different approaches of leverage, under
an analysis which uses historical data from all components of the S&P 500 index. By
performing this back-testing, I am able to avoid making any assumptions about the
distribution of returns. This is in contrast to existing studies which simply assume
that returns are normally distributed, and do not evaluate their performance using
empirical data (Kelly 1952; Rotando and Thorp 1992, and Peters 2010).
In the next chapter, I introduce a leveraged component of agents decision making
into an agent-based model of a financial market.
Chapter 6
Agent-Based Modelling of
Optimal Leverage Model using
Numerical Methods
In this chapter, I introduce a leveraged component of agents decision making into
an agent-based model of a financial market. Previous models have used expected
utility optimisation as a framework for agents trading decisions. The motivation
of this chapter is that there are two different issues on the experiment of leverage.
Firstly, previous studies have shown that similar strategies can perform well when
tested against theoretical models of asset price processes such as geometric Brownian
motion (Peters 2010). However, the former approach fails to account for phenomena
such as non-Gaussian return distributions which are observed in real markets.
Secondly, in Chapter 5, I back-tested the optimal level of leverage using numerical methods against historical empirical price data. However, this cannot take
into account how other participants in the market would likely respond to a newly
introduced trading strategy; i.e., this ignores the existence of market impact.
In order to account for these issues, I test my strategy using an existing agentbased model of financial markets (Iori and Chiarella 2002) which has been shown to
replicate many of the statistical features observed in empirical financial time-series
6.1
Introduction
In Section 2.6, I reviewed the literature on agent-based modelling, and described that
agent-based modelling consist in a set of tools that model individual agent behaviour
and its consequences to the system. These tools allow designing heterogeneous
agents in the same system. Consequently, they also allow us to analyse investor
(agent) behaviour and its consequence to the financial market (system) (LeBaron
2000; Iori and Chiarella 2002; and Lo and MacKinlay 2001).
Agent-based modelling allows us to replicate the stock market artificially in order
to make controlled experiments and simulations (Chakraborti et al. 2011). One of
the advantages of this artificial environment is the control of every aspect of the
simulation, including the individual agent behaviour (Lux and Marchesi 2001; and
Iori and Chiarella 2002). For example, using agent-based modelling, it is possible
to run experiments in high-frequency environment taking into account the market
impact.
Iori and Chiarella (2002) introduce an agent-based model in which they assume
6.1. Introduction
91
that each agent has zero-intelligence in leverage since each agent can only trade one
unit of stock at a time. In order to introduce leverage intelligence in agent-based
modelling, Iori and Chiarella (2009) propose an extension to Iori and Chiarella
(2002) in which agents decide the volume according to a utility function. The
difference among the agents relates to their risk profile: the higher the level of
risk-aversion, the smaller the quantity ordered.
In terms of leverage, in Section 2.3, I highlighted that there is a considerable debate in the literature on the investors objective. There are two different approaches,
the mean-variance approach and the geometric mean approach. The mean-variance
approach advocators argue that the investors objective is to maximise the utility
function according to her risk-profile (Markowitz 1952), and the geometric mean approach advocators argue that the investors objective in the multi-period framework
is to maximise the geometric mean (Williams 1936).
Furthermore, in Section 2.5, I described the three analytical models existing in
the literature to calculate the level of leverage that maximises the geometric mean
of the historical returns. The first model is the Kelly Criterion (Kelly 1956), which
assumes a discrete number of possible outcomes. The second model is the Rotando
and Thorp model (Rotando and Thorp 1992), which incorporates continuous outcomes, but respects the Kelly original insights. Finally, the third model is the Peters
Model (Peters 2010), which assumes that the prices follow a geometric Brownian
motion.
These three analytical models assume a normal distribution of the returns that
undermines the probability of extreme returns (Cont 2001). For this reason, in
Chapter 4, I proposed the use of numerical methods in order to calculate the optimal level of leverage instead of the analytical models. The main advantage of the
numerical methods is that I can relax the assumption of normal distribution of the
returns.
In contrast to Iori and Chiarella (2009), in this experiment, I propose the use of
geometric mean in order to decide the quantity to trade instead of the use of utility
6.2. Platform
93
6.2
Platform
6.3. Model
6.3
95
Model
In this section, I extend the original model introduced Iori and Chiarella (Iori and
Chiarella 2002) which was reviewed in Section 2.6, by allowing the agents to specify
the quantity of their orders, rather than assuming that all orders have a quantity
of 1; and also by introducing three additional agents into the population which use
different levels of leverage. The first agent uses the optimal level leverage using
numerical methods (see Chapter 4) to calculate the quantity in the order:
max
l
[1 + kt (l)] T 1
qtopt =
Kt l
pt
where qtopt is the quantity ordered by the optimal leveraged agent, and l is the
optimal level of leverage.
The second agent targets the level of leverage of 1, i.e., the agent is unleveraged.
Thus, the quantity to be hold by the unleveraged agent is:
qtunl =
Kt
pt
qtexc =
10Kt
pt
6.4
Experiment
In this section, I present the results of the experiment using agent-based modelling
that compare the optimally leveraged investments to unleveraged and excessively
leveraged investments. The section is divided in two subsections. In the first subsection, I describe the methodology of the experiment. And, in the second subsection,
I present the results of the experiment.
6.4.1
Methodology
In this subsection, I describe the methodology of the experiment. I run three different treatments according to the number of artificial agents in order to verify the
ability of the optimally leveraged agent to outperform the unleveraged agent and
the excessive leveraged agent in different sizes of market. The number of agents in
the first treatment is 10, in the second treatment is 100, and in the third treatment
is 1000.
For each treatment, in order to avoid any preference for a particular trading
strategy, I randomly select three different agents who are empowered with leverage
behaviour in contrast to the remaining agents who continue to trade 100 units per
round. The selected agents are split between the optimally leveraged agent (l = l ),
the unleveraged agent (l = 1), and the excessively leveraged agent (l = 10).
Each simulation consists of 106 rounds. On each round, agents has a chance to
re-evaluate their previous order to modify or maintain it in the order book. Every
time that the price in the order to buy is superior or equals to the price in the
counterparty order to sell, the order and the counterpart order matches, a new
trade occurs, and the quantity traded is the quantity of the agent who is offering
6.4. Experiment
97
the least quantity. The remaining quantity of the agent who was offering the highest
quantity continues in the order book until the new match until an either a new match
occurs, or the agent modifies or cancels the order.
This is important to notice that, in a market with less agents, the chance of
a particular agent to re-evaluate the order increases. However, it does not mean
that the probability of a new order to be accepted increases. The reason is that
the re-evaluation is based on the agents trading strategy instead of the order book
position. For example, there are cases where the position of the new order in the
order book is inferior to the previous order which undermines the probability of
matching. Hence, I assume that there is no influence of the rate of re-evaluation on
the level of competition.
I run 100 simulations per treatment. On every simulation I record the end of
period return of the agents. Thus, on each treatment, I collect a set of 100 end of
period returns per agent. Using this data, I test the hypothesis that the optimally
leveraged agent outperforms the unleveraged agent, and test the hypothesis that the
optimally leveraged agent outperforms the excessively leveraged agent.
I use the T-test in order to test these hypotheses. In the first T-test, I test the
difference between the log of the ends of period returns of the optimally leveraged
agent and the ends of period returns of unleveraged agents. In the second T-test,
I test the difference between the log of the ends of period returns of the optimally
leveraged agents and the ends of period returns of unleveraged agents.
The experiment is run in JASA (JAVA Auction Simulation API) (Phelps 2007),
an application of JABM (JAVA for Agent-Based Modelling) (Phelps 2012).
In the next subsection, I present the results of the experiment.
6.4.2
Results
In this subsection, I present the results of the experiment with three different treatments regarding the size of the market. In the first treatment, the size of the market
is 10; in the second treatment, the size of the market is 100; and, finally, in the third
N = 10 Agents
In this treatment, the agent-based model contains 10 agents. Three of the agents
have the ability to order quantities different to 100 per round, which corresponds to
30% of the participants.
Item
Mean
Median
Minimum
Maximum
Standard Deviation
Skewness
Kurtosis
Average Number of Orders Executed
Opt
U nl
Exc
4.4552
0.0531
-7.5764
5.7238
2.4852
-7.4434
-11.3388 -12.2820 -10.3550
6.6596
7.2056
3.3390
2.9081
0.6590
4.2618
-0.07
-0.20
2.84
10.1306
1.7821
27.8044
1772.4
3644.8
1207.9
Table 6.1: Statistical analysis of the log of the ends of period returns, N = 10 agents.
Opt = optimally leveraged, Unl = unleveraged and Exc = excessively leveraged.
Item
Confidence Interval
P-value
Null Hypothesis
Opt U nl
[3.0159 ; 5.7882]
8.2212x109
Rejected
Opt Exc
[11.1622 ; 12.9009]
4.2326x1048
Rejected
Table 6.2: T-test results between the log of the end of period return of the optimally
leveraged agent (Opt) and the unleveraged agent (Unl); and the optimally leveraged
agent and the excessively leveraged agent (Exc). N = 10 agents.
From the treatment with 10 agents, I obtain the following results. Firstly, from
Table 6.1, the average number of orders executed of optimally leveraged, unleveraged and excessively leveraged agent are 1772.4, 3644.8 and 1207.9, respectively.
Secondly, Figure 6.1 shows the logs of the ends of period returns of the optimally
leveraged and excessively leveraged agents. From Table 6.2 and from Figure 6.2,
the null hypotheses of equality of the ends of period returns between the optimally
leveraged agent and the unleveraged agent, and between the optimally leveraged
agent and the excessively leveraged agent are rejected. Furthermore, due to the
positive values of the confidence interval, the end of period return of the optimally
6.4. Experiment
99
Optimally Leveraged
20
0
20
0
20
40
60
Unleveraged
80
100
20
40
60
Excessively Leveraged
80
100
80
100
20
0
20
0
20
0
20
0
20
40
60
Figure 6.1: The plot of the log of the end of period returns of leveraged agents,
N=10 agents.
10
Optimal
Unleveraged
Excessive
Figure 6.2: The plot of the log of the end of period returns of leveraged agents,
N=10 agents.
N = 100 Agents
In this treatment, the agent-based model contains 100 agents. Three of the agents
have the ability to order quantities different to 100 per round, which corresponds to
3% of the participants.
Item
Mean
Median
Minimum
Maximum
Standard Deviation
Skewness
Kurtosis
Average Number of Orders Executed
Opt
U nl
Exc
-0.2892 -1.3826 -5.3485
2.5879 0.6952 -7.5082
-9.4852 -9.3668 -11.0809
4.6248 7.3715
4.9791
5.1023 4.9885
4.8125
-0.7781 -0.2663 1.2220
1.7296 1.6569
2.7902
172.9
150.3
173.3
Table 6.3: Statistical analysis of the log of the ends of period returns, N = 10o agents.
Opt = optimally leveraged, Unl = unleveraged and Exc = excessively leveraged.
Item
Confidence Interval
P-value
Null Hypothesis
Opt U nl
Opt Exc
[ -0.3578 ; 2.5516] [3.6257 ; 6.4618]
0.1378
2.3529x1010
Accepted
Rejected
Table 6.4: T-test results between the log of the end of period return of optimally
leveraged agent (Opt) and unleveraged agent (Unl) and optimally leveraged agent
and excessively leveraged agent (Exc). N = 100 agents.
From the treatment with 100 agents, I obtain the following results. Firstly, from
Table 6.3, the average number of orders executed of optimally leveraged, unleveraged
and excessively leveraged agent are 172.9, 150.3 and 173.3, respectively. Secondly,
Figure 6.3 shows the logs of the ends of period returns of the optimally leveraged and
excessively leveraged agents. Finally, from Table 6.4 and from Figure 6.2, the null
hypothesis of equality of the end of period return between the optimally leveraged
and the unleveraged agent is accepted. However, the null hypothesis of equality
6.4. Experiment
101
Optimally Leveraged
20
0
20
0
20
40
60
80
100
80
100
80
100
Unleveraged
20
0
20
0
20
40
60
Excessively Leveraged
20
0
20
0
20
40
60
Iterations
Figure 6.3: The plot of the log of the end of period returns of leveraged agents,
N=100 agents.
10
Optimal
Unleveraged
Excessive
Figure 6.4: The boxplot of the log of the end of period returns of leveraged agents,
N=100 agents.
N = 1000 Agents
In this treatment, the agent-based model contains 1000 agents. Three of the agents
have the ability to order quantities different to 100 per round, which corresponds to
0.3% of the participants.
Item
Mean
Median
Minimum
Maximum
Standard Deviation
Skewness
Kurtosis
Average Number of Orders Executed
Opt
-2.6577
-0.0207
-9.8221
6.8663
4.8286
-0.0120
1.4469
7.7
U nl
-1.7049
0.0006
-11.4798
7.7735
5.2910
-0.1815
1.6598
15.8
Exc
-4.3978
-7.2105
-11.2290
6.3916
5.1774
0.6895
1.8010
14.1
Table 6.5: Statistical analysis of the log of the ends of period returns, N = 1000
agents. Opt = optimally leveraged, Unl = unleveraged and Exc = excessively leveraged.
Item
Confidence Interval
P-value
Null Hypothesis
Opt U nl
[-2.4527 ; 0.5470]
0.2104
Accepted
Opt Exc
[0.4991 ; 2.9811]
0.0065
Rejected
Table 6.6: T-test results between the log of the end of period return of optimally
leveraged agent (Opt) and unleveraged agent (Unl) and optimally leveraged agent
and excessively leveraged agent (Exc). N = 1000 agents.
From the treatment with 1000 agents, it is possible to obtain the following results. Firstly, from Table 6.5, the average number of orders executed of optimally
leveraged, unleveraged and excessively leveraged agent are 172.9, 150.3 and 173.3,
respectively. Secondly, Figure 6.3 shows the logs of the ends of period returns of
the optimally leveraged and excessively leveraged agents. Finally, similar to the
6.4. Experiment
103
Optimally Leveraged
10
0
10
0
20
40
60
80
100
80
100
80
100
Unleveraged
20
0
20
0
20
40
60
Excessively Leveraged
20
0
20
0
20
40
60
Iterations
Figure 6.5: The plot of the log of the end of period returns of leveraged agents,
N=1000 agents.
10
Optimal
Unleveraged
Excessive
Figure 6.6: The plot of the log of the end of period returns of leveraged agents,
N=1000 agents.
6.5
Discussion
In this section, I discuss the implication of the results of the experiment presented in
Section 6.4. The results of the experiment show that in an agent-based model with
10 agents, where 30% of them are empowered with leverage intelligence, the optimal
leveraged agent outperforms unleveraged agent and excessive leveraged agent with
statistical probability of 95%. In an agent-based model with 100 agents, where 3% of
them are empowered with leverage intelligence, the optimally leveraged agent outperforms the excessively leveraged agent but does not outperform unleveraged with
statistical probability of 95%. Finally, in an agent-based model with 1000 agents,
where only 0.3% of them are empowered with leverage intelligence, similar to an
agent-based modelling with 100 agents, the optimally leveraged agent outperforms
the excessively leveraged agent but does not outperform unleveraged with statistical
probability of 95%.
The use of optimal level of leverage improves the agents performance in the
market with 10 agents. However, the analysis of the results suggest that the ability
of the optimal level of leverage to improve the individual agents performance is
positively correlated with the size of the market; the greater the number of agents,
the lower the agents ability.
These results highlight questions about the importance of the market size. For
6.6. Summary
105
example, is the size of the market interesting to the individual agent regardless of
the agents trading strategy or leverage approach? Assuming that there is positive
correlation between the size and efficiency of the market, the results suggest that
there could be a negative correlation between market efficiency and the ability of
optimal level of leverage to improve the individual trading system performance,
because if the market is efficient the returns are expected to be so small that there
is no winner or loser strategy or approach on leverage.
In the next section, I present the summary and the contribution of this chapter.
6.6
Summary
Chapter 7
Conclusion and Future Work
Leverage plays a very important role in the financial markets. It has the ability
to expand returns and, consequently, profits and end of period wealth. However,
the connection between leverage and cumulative return is not linear. High level of
leverage does not necessary means high level of cumulative equity return. In fact,
an excessive level of leverage increases the risk of default. Hence, studies of leverage
are relevant because this tool can encourage irrational behaviour which can impact
the market and can lead the investor to ruin. In order to comprehend leverage as a
whole, it is necessary to comprehend the investors use of leverage besides its impact
on the financial market.
In this thesis, I proposed a new method to calculate the optimal level of leverage
which uses numerical methods instead of analytical methods. This was not feasible
without advances in computing technology. Furthermore, I tested the ability of the
method to improve the individual investment performance employing two different
approaches: back-testing and agent-based modelling. Back-testing allows me to test
the optimal level of leverage using numerical methods using empirical evidence, and
agent-based modelling allows me to test the optimal level of leverage using numerical
methods in a totally controlled environment.
This chapter is structured as follows. In Section 7.1, I present the summary
of this thesis. And, in Section 7.2, I discuss the possible future research from this
108
work.
7.1
This thesis consists in seven chapters, including the current one. Chapter 1 and 2,
are the introduction to the thesis and the background of the leverage and agentbased modelling, respectively.
In Chapter 1, I introduced the motivation, the research objectives and the approach, and the contribution of thesis. In the motivation, I described the relevance
of the studies of leverage for the individual investments. Additionally, I highlighted
that there are three main reason for the studies of leverage. Firstly, recent studies
about leverage demonstrate that excessive level of leverage is considered irrational
behaviour (Geanakoplos 2009; Thurner, Farmer, and Geanakoplos 2011). Secondly,
the equity return formula has two components: the stock return and the level of
leverage, and the level of leverage is linear in the equity return formula (Peters
2010). Thirdly, leverage can lead the investor to ruin. I demonstrated that the
existing analytical models in the literature to optimize the level of leverage assumes
a normal distribution of the financial returns and it undermines the risk of default.
One of the research objectives of this thesis was to design a numerical method to
calculate the optimal level of leverage that allows the relaxation of the assumption
of normal distribution of the series of financial returns, and to test the ability of
this method to improve investment performance. The approach that I took was
to assume that the agents objective is to maximise the end of period return, and
compare the individual end of period equity return under three different levels of
leverage: the unleveraged (level of leverage = 1), optimally leveraged, excessively
leveraged investment (level of leverage = 10).
There are three specific contributions of this thesis. The first contribution is to
show that a commonly-used proxy of the geometric mean, which is widely used in
the current literature, fails to accurately estimate leveraged returns. The second
109
110
returns. Thus, I tested the hypothesis that the proxy of the geometric mean represents the geometric mean of the leveraged returns. In order to perform the test,
I simulated the investment in the components of the S&P 500 Index on different
windows periods between 2000 and 2011 using different levels of leverage with values
between 1 and 10.
The result is that even though the proxy is accepted to calculate the geometric
mean of unleveraged equity returns, it is rejected to calculate the geometric mean of
leveraged equity returns. The proxy has the advantage of allowing the calculation
of the geometric mean of the stock return in a closed-form solution. However, stock
return is only one of the parameters of the equity return, the other parameter is the
level of leverage.
Thus, the contribution of Chapter 3 is to show that a commonly-used proxy of
the geometric mean, which is widely used in the current literature, fails to accurately
estimate leveraged returns. This proxy uses the arithmetic average and the variance
of the stock returns (Parramore and Watsham 1997). However, this proxy assumes
that returns can be characterised by the first two moments and ignores the fat-tailed
return distributions that are observed in actual empirical data (Cont 2001).
Secondly, in Chapter 4, I proposed a numerical method in order to calculate the
optimal level of leverage. The advantage of the use of numerical methods is that, in
contrast to the analytical models existing on the literature, these methods do not
assume that the series of financial returns are normally distributed.
I tested the hypothesis that numerical methods yield superior geometric mean to
the geometric Brownian motion. In order to do that, I calculated the optimal level
of leverage using numerical methods and by using geometric Brownian motion of 104
series of daily returns of the components of the S&P 500 Index. In the experiment,
the stock, the length of the series and the window period are randomly selected. The
result is that the numerical methods yield superior geometric mean of the series of
historical equity returns to geometric Brownian motion, even though when it shows
inferior level of leverage.
111
Similar to Chapter 3, the results of the experiment in Chapter 4 showed the advances caused by the use of numerical methods in leveraged models. For this reason,
I argue that the proxy and the analytical models used to calculate the unleveraged
equity returns cannot be directly used to calculate the leveraged equity returns because the proxy and the models assume that the returns are normally distributed,
ignores the presence of fat-tails and, consequently, underestimates extreme returns
present on the leveraged investments.
Thus, the contribution of Chapter 4 is the introduction of a novel trading strategy which uses numerical methods in order to calculate the optimal level of leverage. Previous studies have proposed analytical models in order to calculate the
optimal level of leverage. However, these models assume that returns are normally
distributed (Kelly 1952; Rotando and Thorp 1992, and Peters 2010). By using numerical methods in order to optimise the level of leverage, I am able to avoid this
assumption and hence avoid under-estimating extreme returns (Cont 2001).
Thirdly, in Chapter 5, using the historical price of the components of the S&P
500 Index, I back-tested the leveraged investment using the optimal level of leverage
using numerical methods. The hypothesis is that optimal leveraged investment
outperforms the geometric Brownian motion leveraged investment, the unleveraged
investment and the excessively investment.
There are three different results from the back-testing. The first result is that
optimal level of leverage results in a superior investment performance to geometric
Brownian motion leveraged investments. Furthermore, considering that the average
value of the end of period equity return of the geometric Brownian motion leveraged
investments is +14.26%, and that the confidence interval of the difference between
the end of period equity returns of optimally leveraged and geometric Brownian
motion leveraged investments is between +0.84% and +4.00%; the end of period
equity return of the optimally leveraged investments is approximately between 1.1
and 1.3 times the end of period equity return of the geometric Brownian motion
leveraged investments with a statistical probability of 95%.
112
The second result is that the null hypothesis of the equality between the series
of the end of period equity return of the optimally leveraged and the unleveraged
investment is rejected. Furthermore, considering that the average value of the end of
period equity return of the unleveraged investments is +3.16%, and the confidence
interval of the difference of the end of period equity returns is between +9.18% and
+17.19%; the end of period equity return of the optimally leveraged investments
is approximately between 3.9 and 6.4 times the end of period equity return of the
unleveraged investments with a statistical probability of 95%.
Finally, the third result is that the null hypothesis of the equality between the
series of the end of period equity return of the optimally leveraged investment and
the excessively leveraged investment is also rejected. Considering that the average
value of the end of period equity return of the excessively leveraged investments is
-100%, i.e., default, it is not possible to calculate the improvement caused by the
use of optimal level of leverage using numerical methods.
The results of the back-testing showed that, compared to geometric Brownian
motion leveraged, unleveraged and excessively leveraged investment, the optimal
level of leverage using numerical methods improves investment performance. Thus,
the appropriate level of leverage is an important parameter to be taken into account
when the objective is to maximise the end of period equity return.
The contribution of Chapter 5 is to show that my trading strategy (see Chapter
4) exhibits superior performance to a strategy which does not use leverage, under an
analysis which uses historical data from all components of the S&P 500 index. By
performing this back-testing, I am able to avoid making any assumptions about the
distribution of returns. This is in contrast to existing studies which simply assume
that returns are normally distributed, and do not evaluate their performance using
empirical data (Kelly 1952; Rotando and Thorp 1992, and Peters 2010).
Finally, in Chapter 6, I employed an agent-based model in order to experiment
with the use of the optimal level of leverage in a totally controlled environment. I
introduced three leveraged agents regarding their approach of leverage, the optimal
113
level of leverage using numerical methods, the unleveraged and the excessive level
of leverage, in an order-driven model proposed by (Iori and Chiarella 2002). Then,
I simulated it regarding to three different sizes of market, 10, 100 and 1000 agents
in the JASA platform (Phelps 2007).
The hypotheses are that the agent who uses the optimal level of leverage using
numerical methods outperforms the unleveraged agent and the excessive leveraged
agent. There are three different results from agent-based modelling experiment. In
the first treatment, in a market with 10 agents, both hypotheses are accepted. In
the second and in the third treatment, in a market with 100 and 1000 agents, respectively, the hypothesis that the optimal level of leveraged using numerical methods
improves the investment performance compared to is accepted excessively leveraged
investment; but compared to unleveraged agent the hypothesis that the optimal
level of leveraged using numerical methods improves the investment performance is
rejected.
The results of the experiment suggest that the ability of the use of the optimal
level of leverage using numerical methods to improve investment performance depends on the size of the market. It is not a guarantee of the improvement of the
investment performance regardless of the market condition and the trading strategy.
Furthermore, the results also suggest that there is a negative correlation between
the ability of the optimal level of leverage using numerical methods to improve the
investment performance and the size of the market.
The contribution of Chapter 6 is the introduction of a leveraged component of
agents decision making into an agent-based model of a financial market. Previous
models have used expected utility optimisation as a framework for agents trading
decisions. In contrast, my model uses the aforementioned trading strategy in order
to decide the quantity of orders by numerically-optimising the level of leverage. By
so doing, I am able to use this agent-based model to evaluate the potential effect of
market impact on the performance of my leverage-based trading strategy. To the
best of my knowledge, this is the first attempt to systematically evaluate the effect
114
7.2
Future Work
The work of this thesis can be extended in three directions. The first direction
is the optimization of the level of leverage of a portfolio of stocks. This thesis
proposed a numerical methods to optimize the level of leverage of an investment in
a single stock. It could be interesting to investigate and experiment the possibility
of the use of numerical methods to optimize the level of leverage of a portfolio of
stocks. Estrada (2010) proposes an analytical model to calculate the optimal weight
115
of a single stock in a portfolio of stocks. However, his model assumes the normal
distribution of the returns. Thus, it could be interesting experiments of the usage
of numerical methods in order to optimize the portfolio of stocks in terms of the
geometric mean, that relax the assumption of normal distribution of the returns.
The second direction is the use of different computational methods to optimize
the level of leverage, particularly, evolutionary computational algorithms such as
genetic algorithms and genetic programming. The evolutionary computation comes
from the Darwins idea of evolution, where survival depends on fitness. In the
context of solution fitting, the idea is to maintain a set of candidate solutions. Each
solution is evaluated for its fitness, which is measured in terms of their correctness
in classifying the data. The functions are then modified in an artificial environment
that mimics evolution, which eliminates weak individuals and allow fitter individuals
to pass their genetic material to latter generations (Tsang 2009). The difference
between genetic algorithm and genetic programming is that solutions are values
in the genetic algorithms and functions in the genetic programming. My insight
is that the calculation of the optimal level of leverage of a portfolio of stocks can
show different results of the existing results applying evolutionary computation in
its calculation. The reason is that the number of possible local optima in a portfolio
of stock is superior to this number in a single stock. Consequently, the chance of
have reach to a local optima in a portfolio of stock is superior to the chance in a
single stock.
Finally, the third field is the further development of agent-based modelling.
There are three topics that deserve further attention for the researchers. The first
topic is margin allowance. The current agent-based models do not consider the existence of margin limits and, consequently, do not take into account the possibility
of margin calls. Margin calls are orders input by the margin provider in order to
restore the margin allowance of the borrower. These kind of orders are irrational
because it was not input after the indication of some strategy, but input simply to
restore the margin. My intuition is that the existence of margin limits can modify
116
the results of the experiment, particularly, in the case of excessively level of leverage.
The second topic is the relevance of the current price to calculate the orders
price. The current model of Iori & Chiarella (2002) assumes that the order is Uniformly distributed between the current price and the valuated price. For example,
if the current price is $1000 and the valuated price is $1500, the probability that
the agents order to buy be $1001 and $1499 is the same. However, intuitively, it is
expected that the agent intend to proximate his orders price to the current price
as much as possible. Thus, new experiment with different statistical distributions
could be interesting.
Finally, the third topic is learning. In this thesis, I assumed no learning, it
means that agents does not modify their behaviour and their strategy along of the
experiment. The reason to assume no learning is that one of the objective of the
thesis was to introduce the optimal leverage models in the order-driven model (Iori
and Chiarella 2002). However, I argue that there is a considerable area of research
in learning, order-driven model and optimal leverage models.
LeBaron and Yamamoto (2007) introduce learning in the order-driven model.
Their modification is that the trading rules are repeatedly updated and adapted
via learning of the agents. The authors propose a learning process, where agents
imitate the strategy of the most successful investment among the others agents. The
learning mechanism is a genetic algorithm. After five rounds, agents review their
records and forecast, and they update their forecast methods in order to improve
the forecast ability, i. e., the agents modify the weight of the three components of
the stock return forecast in the order-driven model: the weight of the fundamental
component, the weight of the technical component and the weight of the noisy
component. Hence, differently of the original order-driven model where the value of
the weight is randomly select in the begin of the simulation and does not change over
time; the model with learning, the value of the weight varies along of the simulation
(LeBaron and Yamamoto 2007), which means that agents evolve over time.
Complementary to LeBaron and Yamamoto (2007), LeBaron (2011) argues that
117
the agent-based financial markets present two forms of learning: passive learning
and active learning. On one hand, passive learning refers to the accumulation of
wealth of the successful investment. Successful strategies tends to be a large part
of the market because the strategies that lead to an unsuccessful investment are
abandoned. On the other hand, active learning refers to the adaptations of the
agents on every new information. Instead of abandoned the strategy, such as the
passive learning, the active learner intends to modify the strategies towards to the
strategies that lead to a successful investment.
LeBaron (2011) also argues that in the real markets there are both forms of
learning, passive learning and active learning, in different scales, for three different
reasons. Firstly, passive learning is easy to model because it only requires the
elimination of the bad strategies. Secondly, active learning is not easy to model
because it can involve many degrees of freedom. Finally, active learning can adapt
to the noise in the financial returns.
There is literature of learning in the order-driven model (LeBaron and Yamamoto
2007). However, unlike this thesis, these authors also assume one stock per unit,
i.e., no intelligence in the use of leverage. For this reason, it could be interesting
researches to combine learning, order driven-model and leverage intelligence. The researches could be in three different directions. The first direction is the introduction
of the optimal leverage models in the order-driven model with learning (LeBaron
and Yamamoto 2007). The second direction is the introduction of learning in the
optimal level of leverage models, and the modification of the model in Chapter 6,
in order to empower the agents with the ability to evolve their use of the leverage
over time. Finally, the third direction is the combination of learning in leverage
and learning in the order-driven model. It could be interesting to investigate how
agents evolve on the combination of leverage and trading rules. My insight is that
different levels of leverage could be associated to different strategies. For example,
agents could adapt the level of leverage, and, consequently the size of the order, to
the strategy and evolve his approach on leverage based on the investment outcomes.
References
Adrian, T. and Shin, H. 2010. Liquidity and Leverage. Journal of Financial
Intermediation 19, no. 3 (July): 418-437.
Bak, P., M. Paczuski, and M. Shubik: 1996, Price Variations in a Stock Market
with Many Agents. Statistical Mechanics and Its Applications 246(3), 430-453.
Breiman, L.: 1961, Optimal Gambling Systems for Favorable Games. Fourth
Berk. Symp. on Math. Stat. and Prob. 1, 65-78.
Breuer, P.: 2002, Measuring Off-Balance-Sheet Leverage. Working Paper
00/202, International Monetary Fund.
Brock, W., J. Lakonishok and B. LeBaron: 1992, Simple Technical Trading
Rules and the Stochastic Properties of Stock Returns. The Journal of Finance
47(5), 1731-1764.
Brunnermeier, M. and L. Pedersen: 2005, Predatory Trading. The Journal of
Finance, 60(4), 1825-1863.
Chakraborti, A., I. Muni Toke, M. Patriarca, F. Abergel, and others: 2011,
Econophysics review: II. Agent-based models. Quantitative Finance 11(7),
1013-1041.
Christensen, M.: 2005, On the history of the Growth Optimal Portfolio. Working Paper 389, University of Southern Denmark .
Collier, N.: 2003, Repast: An extensible framework for agent simulation.
Natural Resources and Environmental Issues, Vol. 8, Article 4.
Cont, R.: 2001, Empirical Properties of Asset Returns: Stylized Facts and
Statistical Issues. Quantitative Finance 1(2), 223-236.
Domian, D., M. Racine and C. Wilson: 2003, Theoretical and Empirical Distributions for Leveraged Canadian Stock Portfolios Over Long-holding Periods.
Proceedings of the Administrative Sciences Association of Canada 24(1), 40-60..
Darvas, Z.: 2009, Leveraged Carry Trade Portfolios. The Journal of Banking
and Finance 33(5), 944-957.
Estrada, J.: 2010, Geometric Mean Maximization: An Overlooked Portfolio
Approach?. The Journal of Investing 19(4), 134-147.
FSA: 2007, Disclosure of Contracts for Difference. Consultation Paper 07/20.
120
References
Geanakoplos, J.: 2009, The Leverage Cycle. Discussion Paper 1715, Cowles
Foundation.
Grossman, S.: 1976. On the efficiency of competitive stock markets where
traders have diverse information. Journal of Finance 31, 573-585.
Guo, W., F. Wang and W. Ho-Mou: 2010, Financial Leverage and Market
Volatility with Diverse Beliefs. Economic Theory 28, 1-28.
Hakansson, N.: 1971, Capital Growth and the Mean-Variance Approach to
Portfolio Selection. The Journal of Financial and Quantitative Analysis 6(1),
517-557.
.: 1971, Multi-Period Mean-Variance Analysis: Toward a General Theory of
Portfolio Choice. The Journal of Finance 26(4), 857-884.
Hull, J.: 2009, Options, Futures and Other Derivatives. Pearson/Prentice
Hall.
Hunt, F.: 2002, Growth Optimal Investment Strategy Efficacy: An Application
on Long Run Australian Equity Data. Research Papers Series 86, Quantitative
Finance Research Centre, University of Technology, Sydney.
Iori, G. and C. Chiarella: 2002, A Simulation Analysis of the Microstructure
of Double Auction Markets. Quantitative Finance 2(5), 346-353.
Iori, G. and C. Chiarella: 2009, The Impact of Heterogeneous Trading Rules
on the Limit Order Book and Order Flows. Journal of Economic Dynamics
and Control 33(3), 525-537.
Jacquier, E., A. Kane and A. Marcus: 2003, Geometric or Arithmetic Mean:
A Reconsideration. Financial Analysts Journal 59(6), 46-53.
Norman, D.: 2009, CFDs, The Definitive Guide to Contracts For Difference.
Harriman House.
Kelley, C.: 1999, Iterative Methods for Optimization. Frontiers in Applied
Mathematics 18, Society for Industrial and Applied Mathematics (SIAM),
Philadelphia.
Kelly, J.: 1956, A New Interpretation of Information Rate. The Bell System
Technical Journal 2(3), 185-189.
Khandani, E. and A. Lo: 2007, What Happened to the Quants in August
2007?. Journal of Investment Management 5(4), 5-54.
: 2008, What Happened to the Quants in August 2007?: Evidence from
Factors and Transactions Data. Journal of Financial Markets 14, 1-46.
Latane, H.: 1959, Criteria for Choice Among Risky Ventures. The Journal of
Political Economy 67(2), 144-155.
LeBaron, B. 2000, Agent-based computational finance: Suggested readings and
early research. Journal of Economic Dynamics and Control 24(5), 679-702.
References
121
122
References
Phelps, S.: 2012, Applying Dependency Injection to Agent-Based Modelling:
the JABM Toolkit. Working Papers WP056-12, Centre for Computational Finance and Economic Agents, University of Essex.
Resnick, M.: 1996, StarLogo: an environment for decentralized modelling and
decentralized thinking. In Conference Companion on Human factors in Computing Systems, 11-12, ACM Press, New York.
Ross, S.: 1999, Adding Risks: Samuelsons Fallacy of Large Numbers Revisited. The Journal of Financial and Quantitative Analysis 34(3), 323-339.
Rotando, L. and E. Thorp: 1992, The Kelly Criterion and the Stock Market.
The American Mathematical Monthly 99(10), 922-931.
Samuelson, P.: 1971, The Fallacy of Maximizing The Geometric Mean in Long
Sequences of Investing or Gambling. Proceedings of the National Academy of
Sciences of the United States of America 68(10), 2493-2496.
: 1979, Why We Should Not Make Mean Log of Wealth Big Though Years
to Act Are Long. Journal of Banking and Finance 3(4), 305-307.
Sklar, E.: 2007, NetLogo, a multi-agent simulation environment. Artificial
Life 13(3), 303-311.
Thurner, S., J. Farmer and J. Geanakoplos: 2011. Leverage Causes Fat Tails
and Clustered Volatility. Cowles Foundation Discussion Papers 1745, Cowles
Foundation for Research in Economics, Yale University.
Tsang, E.: 2009, Forecasting - Where Computational Intelligence Meets The
Stock Market. Frontiers of Computer Science in China 3(1), 53-63.
Williams, J.: 1936, Speculation and the Carryover. The Quarterly Journal of
Economics 50(3), 436-455.