5245 Dissertation

Download as pdf or txt
Download as pdf or txt
You are on page 1of 34

Determinants of individual investors behavior in portfolio decisions

Dmitry Salimov

Universite Paris 1 Pantheon-Sorbonne 02 UFR Economics Master 2 Economie Theorique et Empirique Supervisor: Jean-Marc Tallon June 14, 2012

Determinants of individual investors behavior in portfolio decisions


Dmitry Salimov Universite Paris 1 Pantheon-Sorbonne 106 - 112 Boulevard de lHopital, 75013 Paris, France

Abstract In my thesis I study aggregate aspects of individual investors behavior such as choices of the share of risky assets and amount of investment, choice of investment instruments and the duration of relationship with the investment company. I attempt to explain the variations in these parameters by using demographic, socioeconomic and, most importantly, personality trait variables such as aversion to risk, cognitive skills and several others. I make a particular emphasis on the decisions related to the share of risky assets in the overall portfolio of the investor. I eliminate standard risk and ambiguity aversion measures as possible explanatory variables for the share of risky assets but provide an intermediary measure that is dened over actual behavior of the subjects in situations involving risk. I nd out that the choice of the aggregate level of risk by the investor is actually quite rational and relies mostly on the ability of the investor to quantify and control the risk meaning that the irrationality appears only on the specic level.

The University of Paris 1 Pantheon-Sorbonne does not intend to give any approval or disapproval to the opinions expressed in this paper and must be considered the authors own.

Contents
1 Introduction 1.1 1.2 1.3 1.4 1.5 Behavior of individual investors . . . . . . . . . . . . . . . . . . . . Goals of the study . . . . . . . . . . . . . . . . . . . . . . . . . . . Other studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Aversion to Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . Aversion to Ambiguity . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 2 3 3 4 5 5 6 7 7 8 10 11 12 13 14 15

2 Data 2.1 2.2 2.3 2.4 2.5 2.6 Data transformation . . . . . . . . . . . . . . . . . . . . . . . . . . Bias in the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . AXA data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Questionnaire data . . . . . . . . . . . . . . . . . . . . . . . . . . . Riskav variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Other constructed variables . . . . . . . . . . . . . . . . . . . . . .

3 Basic Analysis 3.1 3.2 3.3 3.4 Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Amount of investment . . . . . . . . . . . . . . . . . . . . . . . . . Active / passive instruments . . . . . . . . . . . . . . . . . . . . . . Duration of relationship with AXA . . . . . . . . . . . . . . . . . .

CONTENTS 4 Models 4.1 4.2 4.3 Linear model for uc share . . . . . . . . . . . . . . . . . . . . . . . Robustness check . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hurdle model for uc share . . . . . . . . . . . . . . . . . . . . . . .

3 17 17 18 18 20 22 29

5 Conclusions Appendix: Tables and gures References

Chapter 1 Introduction
The question of quantifying human behavior was always one of my favorite ones. The methods for collecting data on human behavior have rapidly evolved in the past decade and nowadays we can perform empirical studies of actual human behavior instead on relying on the articial experimental data. In my analysis I will attempt to provide explanation for some basic decisions that are made by individual investors by using their demographic and socioeconomic characteristics and their personality traits.

1.1

Behavior of individual investors

Analyzing every small decision made by investors like the decision to invest a specic amount of money to buy stock of a specic company on a specic date is virtually impossible. Small decisions almost always depend on the context: the investor may have seen in the news that this particular company had some major breakthrough or maybe he just discovered that he will not be able to go on a holiday and decided to invest this money instead and this particular stock was recommended to him by the agent. Because of this uncertainty I do not attempt to study individual decisions but instead construct aggregate measures that characterize the aggregate state of the investors portfolio reached as the nal result of these small decisions made by the investor. Due to the specics of the data and my interest the emphasis of the analysis will be made on the measure of the share of risky assets (e.g. stocks, hedge funds shares) in the overall portfolio of the investor.

CHAPTER 1. INTRODUCTION

I would also like to note that when I study the share of risky assets I implicitly use the law of large numbers since investors characteristics do not inuence the share as is but instead make an impact on the separate small decisions made by the investor.

1.2

Goals of the study

The main goal of my research was to obtain some intuition behind the factors that inuence the decisions of an individual investors. There are a lot of studies that focus on the impact of demographic or socioeconomic variables on the decision of the investors or the impact of performance-based indicators. But I was more interested in the relationships between the decisions and the personal characteristics of the subject such as his cognitive abilities, his aversion to risk or his self-control. This was my main reason for including a lot of measures related to the character of the investor into my analysis. I also include several measures of nancial risk perception to check whether it aects the choices related to the risky assets in the portfolio. The reason for my focus on explaining the choice of the share of risky assets is the fact that this aspect of an investment decision is the one that is most prone to be aected by factors that are hard to observe and not directly related to the investment problem. For example the decision about the amount of money to invest will mostly vary depending on the investors wealth and the studying the impact of other factors would require a large number of observations that would allow us to study the variability for dierent xed levels of wealth. The choice of the share of risky assets on the other hand does no have any extremely prominent factors allowing us to study the relationships between the variables even in the cases of relatively small samples. I would also like to articulate the fact that because of the lack of good data on both sides of the equation (the data on actual behavior of investors and the data concerning their personality) at the same time a general analysis of these relationships between the investment decisions and their possible factors was impossible. But the appearance of AXA dataset changed that and allowed me to study the basic connections between the personality and the actual behavior of an individual investor.

CHAPTER 1. INTRODUCTION

1.3

Other studies

There are several notable studies that focus on the behavior of individual investors, on the impact of personal traits on economic decisions or on the relationships between the risk attitude measures: Barber and Odean in [BO11] study the stock trading behavior of individual investors. They found that instead of behaving rationally individual investors exhibit a lot of behavioral biases seriously aecting their nancial well being. In my study I show that risk attitude is somewhat related to the personality and behavioral traits of the individual which can explain the irrationality in choices. Burks, Carpenter, Goette, and Rustichini in [BCGR08] nd that cognitive skills signicantly aect individual preferences. In particular they nd that willingness to take risks and patience both increase with cognitive skills and in my study I also nd some support for this statement. Butler, Guiso and Jappelli in [BGJ11] make a link between the decision style (intuition vs. reasoning) and both risk and ambiguity aversion when conducting a study of a large sample of retail investors. Although I did not have any data on the decision style in my study I observed a negative relationship between risk and ambiguity aversion which by the results of the [BGJ11] can be present among relatively wealthy individuals.

1.4

Aversion to Risk

Risk is the potential that a chosen action or activity (including the choice of inaction) will lead to a loss (an undesirable outcome). Risk aversion is the reluctance of a person to accept a bargain with an uncertain payo rather than another bargain with a more certain (meaning lower potential losses), but possibly lower, expected payo. For example, a risk-averse investor might choose to put his or her money into a bank account with a low but guaranteed interest rate, rather than into a stock that may have high expected returns, but also involves a chance of losing value. Example: A person is given the choice between two scenarios, one with a guaranteed payo and one without. In the guaranteed scenario, the person receives $50. In the uncertain scenario, a coin is ipped to decide whether the person receives $100 or nothing. The expected payo for both scenarios is $50, meaning that an individual

CHAPTER 1. INTRODUCTION

who was insensitive to risk would not care whether they took the guaranteed payment or the gamble. However, individuals may have dierent risk attitudes. The average payo of the gamble, known as its expected value, is $50. A person is said to be: risk-averse (or risk-avoiding) - if he or she would accept a certain payment (certainty equivalent) of less than $50 (for example, $40), rather than taking the gamble and possibly receiving nothing risk-neutral - if he or she is indierent between the bet and a certain $50 payment risk-loving (or risk-seeking) - if the guaranteed payment must be more than $50 (for example, $60) to induce him or her to take the guaranteed option, rather than taking the gamble and possibly winning $100

1.5

Aversion to Ambiguity

Ambiguity aversion (also known as uncertainty aversion) on the other hand describes an attitude of preference for known risks over unknown risks. People would rather choose an option with fewer unknown elements than with many unknown elements. It is demonstrated in the Ellsberg paradox (i.e. that people prefer to bet on an urn with 50 red and 50 blue balls, than in one with 100 total balls but where the number of blue or red balls is unknown). The distinction between ambiguity aversion and risk aversion is important but subtle. Risk aversion comes from a situation where a probability can be assigned to each possible outcome of a situation. Ambiguity aversion applies to a situation when the probabilities of outcomes are unknown [Eps99]. The main idea behind ambiguity aversion encompasses the idea of risk aversion. A real world consequence of increased ambiguity aversion is the increased demand for insurance because the general public are averse to the unknown events that will aect their lives and property.

Chapter 2 Data
The data used in my thesis comes from two sources: AXA panel: panel data on the portfolio actions of 618 AXA clients between years 2002 and 2011. The data comes from the AXAs management system and includes contract/support level data on all the actions that were preformed on the clients accounts. Two waves of questionnaires conducted by phone via computer. They were organized by Institut dEtudes marketing by the request of Paris School of Economics. Subjects were not aware of the questionnaires relation to the AXA. The rst questionnaire has a sample of 1000 participants, second - 807 participants. Note: the legend for all ambiguously named variables is given at the beginning of appendix (see table 5.1 on page 22).

2.1

Data transformation

Initially the data from the AXA panel was collected at two dierent networks: AA (external agents) and RCS (employees of AXA). Because of dierent data management protocols the information was presented in the following structures: client id > contract id > support id > year > month for AA network client id > contract id > year > month for RCS network 5

CHAPTER 2. DATA

Since I was only interested in the eects on the client level I aggregated the two datasets into one by taking sequential averages/sums over the variables of interest. For example uc share variable was obtained by calculating averages (over time) of the amounts of risky and total assets for every contract, summing up the contract averages for every client and then dividing the amount of risky assets over the total assets. There is an implicit assumption of the fact that the riskiness of dierent risky assets is more or less the same. I am backing up this assumption by the fact that unlike a stand-alone investment opportunity the assets that we are talking about are oered by an insurance company. Most of them are complex nancial instruments and oering any assets with exceptional risks&returns would likely be unprotable for the company. Also, I would like to mention that although the preliminary analysis of the missing observations was conducted before I received the data I did not use any variables with a signicant share of missing values. Because of this precaution I felt secure excluding the missing observations when I performed the tests.

2.2

Bias in the data

From my point of view there are two main potential sources of bias in the sample: questionnaire data comes from people who accepted the request to participate in the questionnaires and AXA data comes from people who have a assurance vie contract in AXA. I believe this biases to be of small signicance because of the following reasoning: According to the organization conducting the questionnaires almost nobody rejected the request to participate in the rst questionnaire and almost 81% of of these people also participated in the second one. This means that the rejection rate is small enough for our sample to be similar in distribution to the general population. A majority of people in France hold assurance vie contracts because they provide signicant tax benets to the holder. AXA is one of the biggest insurance companies in France, its clients are well distributed both geographically and socioeconomically. AXA Group ranks as the 9th largest company in the world (based on revenue) on the 2010 Fortune Global 500 list.

CHAPTER 2. DATA

Based on the information above I make an assumption that that there is no significant deviation in the distribution of my sample with respect to the general population.

2.3

AXA data

The information coming from AXA panel mainly concerns clients balance, actions on the contracts and some basic demographic data. Based on this information I created several variables that allow me to identify clients behavior using several axis: share of risky support in the portfolio; the amount of investment; usage of active/passive instruments; prevalence of withdrawals.

In my thesis I will pay most attention to the share of risky support in the portfolio. For the other points of interest I will only provide some interesting ndings obtained during the analysis of the data.

2.4

Questionnaire data

Among the data that was collected during the questionnaires I selected the demographic data (education, employment, income) and data on personal traits (cognitive and mathematical abilities, nancial literacy, character traits). One of the novelties of my work is the fact that this is the rst time when character traits such as (self-condence, thoughtfulness, negligence) and attitude to managing personal nances are used in an attempt to explain the actual decisions of investors because such data was not available before. Also included in the questionnaires are some standard measures of risk and ambiguity aversion including the risk tolerance measure of Barsky [Bea97] that is prevalent in studies of decisions under risk. Here is an example of such measure: You have a choice between two options: (a) Earn 400 euros for sure. (b) Have a 50% chance to win 1000 euros and a 50% chance of winning nothing.

CHAPTER 2. DATA

If the respondent picked (a), the survey continues to ask: (c) Earn 300 euros for sure. (d) Have a 50% chance to win 1000 euros and a 50% chance of winning nothing. If the respondent picked (b), the survey continues to ask: (e) Earn 500 euros for sure. (f) Have a 50% chance to win 1000 euros and a 50% chance of winning nothing. From the answers to this question we obtain a categorical variable taking values from 1 to 4 that is increasing in the degree of relative risk aversion. In the main section of my thesis I will show that such measures have little explanatory power on the agents decision about the share of risky assets in his portfolio. Similar result but in a more theoretic framework were obtained by Kapteyn and Teppa in [KT02].

2.5

Riskav variable

As stated in the previous paragraph standard measures of risk aversion have almost no explanatory power on the share of risky assets in the portfolio. But I was able to construct an aggregate ad hoc measure of risk attitude named riskav that has a signicant relationship with the share of risky assets and an even closer relationship with standard risk aversion measures. This means that riskav can serve as an intermediary variable connecting the standard measures and share of risky assets. The components for riskav measure come from Questionnaire 1 and answer following questions: Q7: If you have or if you would have children, do you or would you be willing to encourage them to take risks? Q9: Did you have a health check over last 5 years? Q12: How often do you take a raincoat / an umbrella when weather is uncertain? Q16a: Did you ever happen to not pay your parking at the parking ticket machine for less than one hour? Q16b: Did you ever happen to park outside of authorized areas? Q17: Did you play following in the past 12 months: a) PMU, b) lotto, c) scratch type lotteries, d) in casino, e) online poker? S8a: Do you own any shares?

CHAPTER 2. DATA

The principal component analysis (see table 5.3 on page 23) showed that all of the seven components account for at least 9% of variability (which is a very high value considering the average of 14% per component) so all seven should be used in the aggregate. But correlational analysis of these components against the share of risky assets in the portfolio shows that Q17 responses have a very surprising pattern where part of the risky game activities correlates positively with the share while the other part correlates negatively. Due to this uncertainty Q17 component was excluded from the nal version of riskav variable. As we will see in sections 3&4 personality traits and risk perception measures will have little direct impact on the value of uc share. However (same as with risk aversion measures) they will have implicit relationship with uc share through riskav. Now I will list all notable eects of the aforementioned measures on riskav (see table 5.4 on page 23): Self-condent people are more likely to engage in risky activities.

Impulsive people are more likely to engage in risky activities ( 90% condence). Emotional people are less likely to engage in risky activities. People who associate nancial risks with enjoyment are more likely to engage in risky activities. People who associate nancial risks with gain are more likely to engage in risky activities ( 90% condence). People who are decision-makers in their household are more likely to engage in risky activities ( 90% condence).

People with higher cognitive abilities are more likely to engage in risky activities. Men are more likely to engage in risky activities. From several of the statements above I could make a reasonable assumption that the possibility of engaging in a risky activity is closely intertwined with the perceived ability of the individual to avoid or control losses that may occur as the result of the activity. Thus self-condence, cognitive abilities, condence in own nancial management skills greatly inuences the chances of an individual to engage in any risky activity. Also, speaking about the dierences in the likelihood of participating in risky activities between dierent employment categories (see gure 5.2 on page 25): ordinary

CHAPTER 2. DATA

10

employees, laborers and students are less likely to engage in risky activities. On the other hand CEOs, senior sta and people of liberal professions are more likely to engage in risky activities. In the end riskav measures the amount of risky activities that given individual participated in. For a risk averse individual we would expect the value of riskav to be low and for risk loving one to be high.

2.6

Other constructed variables

There are several other variables that did not come directly from AXA panel and questionnaires but were constructed from raw data. uc share is the share of risky assets in the portfolio of a given client obtained as an average (over time) value of risky assets in the portfolio over the similar average value of total assets d ifuc is the dummy for uc share > 0 d uc is the dummy for uc share > mean(uc share)

PMA is the average value of total assets for a given client

nbacts is the total number of observed actions performed by the client

Q2 Q15 n is the measure of nancial literacy that counts the correct answers to nancial questions in the Questionnaire 2 (e.g. about income tax in France or CAC 40 index) Q2 beta describes the type of time discounting exhibited by the subject cogn is the measure of cognitive abilities that counts the correct answers to cognitive questions in the Questionnaire 2 (e.g. coin toss results or complex interest rates) shAE and shAW are shares of active instruments used by the client for operations with the accounts (E stands for entries such as opening a new contract or making an additional installment, W stands for withdrawals)

Chapter 3 Basic Analysis


The instruments that I use to analyze the data are all very common. I used standard Pearsons correlation and rank correlation to perform the initial analysis of the relationships between the variables. In some cases I also used boxplot (box-and-whisker diagram) graphs to perform the visual analysis of the relationships. Afterwards I used linear and binary response regression models to study the explanatory power of the investors characteristics on their decisions. My reasons for using the Spearmans rank correlation test to study the variables are following. First, the rank correlation assesses how well the relationship between two variables can be described using a monotonic function (if there are no repeated data values, a perfect Spearman correlation of +1 or -1 occurs when each of the variables is a perfect monotone function of the other) because of this it is often better suited for analyzing the categorical variables which constitute the majority of my dataset. Second, unlike the standard correlation coecient Spearmans rank correlation test gives us not only the coecient but also the p-value for the test which allows us to objectively measure the signicance of the relationships between variables. I used boxplots to visually analyze the data in the cases when the statistical results were ambiguous. Unlike scatterplots boxplots dont work very well on the variables with a large support but with the categorical variables boxplots provide a a more informative representation of the data because instead of just plotting the data boxplots allow to visually compare the dierences in the main statistics of the distribution (e.g. mean, spread, skewness) of one variable for dierent values of the other. I use RA (AA) to identify standard risk (ambiguity) aversion measures from Questionnaire 2, for Barsky risk aversion measure from Questionnaire 1 I will use the name barsky to identify it. 11

CHAPTER 3. BASIC ANALYSIS

12

3.1

Risk

As we can see in the table below the relationship between uc share and standard risk and ambiguity aversion measures (barsky, RA, AA) is insignicant. In fact if we try to build regression models for uc share with those measures as independent RHS variables the resulting p-values will be much higher that 0.05. Note: the relationship relationship between risk and ambiguity aversion is negative (rank correlation also shows signicance at 2% level). According to [BGJ11] this pattern can be present among relatively wealthy individuals but any conclusive results will require a separate study. Table 3.1: Pairwise correlations for risk-related variables uc share riskav barsky RA AA uc share riskav barsky RA AA 1.0000 0.1457 1.0000 -0.0590 -0.3441 1.0000 -0.0747 -0.2191 0.3038 1.0000 0.0696 0.0843 -0.0732 -0.1027 1.0000

The results above mean that standard risk attitude measures either do not reect the real risk attitude of the subject or, more likely, are too simplistic to explain such a complex decision as the decision about the balance between the risky and riskless assets in the portfolio. On the other hand our aggregate measure riskav has a more signicant correlation with uc share and can be used as an explanatory variable. At the same time it has a very strong relationship with the standard risk aversion measures. This means that risk tolerance measures have some explanatory power on the actual behavior related to risk (such as illegal parking and attending regular health checks) but are not powerful enough to predict behavior in a complex nancial framework. Other notable eects (see table 5.5 on page 24): People who are more condent in managing their nances are more likely to invest in risky assets People with higher level of education are more likely to invest in risky assets People who are currently retired are less likely to invest in risky assets People with higher monthly income are more likely to invest in risky assets (note that there is no similar eect from the total assets measure)

CHAPTER 3. BASIC ANALYSIS

13

People with higher level of nancial literacy are more likely to invest in risky assets On the topic of dierences in the share of risky assets between dierent employment categories (see gure 5.1 on page 24): ordinary employees, students and retirees are more likely to have a lower share of risky assets in the portfolio. On the other hand craftsmen, shopkeepers, senior sta and people of liberal professions are more likely to have a higher share. I am would like to point out that the general conclusion from the results above would be the fact that the choice of the share of risky assets in the portfolio relies on the ability of the investor to account for (nancial condence and literacy, education) and control (employment, monthly income) the risk. This indicated that the choice of aggregate risk is a very rational decision.

3.2

Amount of investment

In order to measure the PM (amount of investment) I created three variables: PMA = average PM during the active time over all contracts (at least one contract with PM>0) PMM = maximal level of PM reached during the active time PMT = PMA*time measures the impact of investment more accurately than just the amount of PM (an investment that lasts 2 years has more impact than the one that lasts only 1 year) Table 3.2: Pairwise correlations for amount of investment variables pma pmm pmt pma 1 0,9455 0,9407 pmm 1 0,9177 pmt

Pairwise correlation analysis for dierent measures for amount of investment show that all of these measures are similar to the point that all the conclusions made for one of them will most likely be valid for other ones too. In this case I will use PMA for all further analysis. Notable eects (see table 5.6 on page 25):

CHAPTER 3. BASIC ANALYSIS People with higher level of education are likely to have invested more

14

People who are decision makers in their household education are likely to have invested more People who have children are more likely to have invested less People with higher amount of total assets are likely to have invested more (note that eect from higher monthly income is much smaller, a reverse situation w.r.t. uc share) People with higher level of nancial literacy are likely to invest more People who are older / retired are more likely to have invested more People who like to gamble are less likely to have invested more

Here I would like to point out the dierence in the two nancial measures: income and wealth. The total wealth measures the level of nancial situation of the investor and inuences the amount of money that is invested. Monthly income on the other hand measures the stability of nancial situation and while it also slightly inuences the amount of investment its impact on the choice of risk level is much more important. Note also the fact that education level and nancial literacy have a positive inuence on both the amount of investment and the level of risk while the correlation between the latter two is actually negative. This implies a very strong connection between education / nancial literacy and the choice of the level of risk. Interestingly, Q2 beta (time discounting factor) has no inuence on the amount of investment even though we would expect people with high factor to invest more since they are more patient in terms of their intertemporal consumption.

3.3

Active / passive instruments

When managing their portfolios AXA clients can use dierent instruments which I separated into four categories according to the sign (entries/withdrawals) and type (active/passive). Active instruments are the ones that should be initiated by client every instance (e.g. opening a new contract, making a custom installment), passive are the ones that are programmed to be executed at certain times (e.g. a xed monthly payment from a bank account). As I suspected there are some dierences in the use of instruments depending on the clients personal traits but most of the variability comes from the fact that dierent instruments are convenient in dierent situations.

CHAPTER 3. BASIC ANALYSIS Notable eects for entries (see table 5.7 on page 26):

15

Older / retired people are more likely to use active investment instruments. Methodical people are more likely to use passive investment instruments. Men are more likely to use active investment instruments.

The higher use of passive investment mechanisms also corresponds to a larger amount of actions on the contract. Notable eects for withdrawals (see table 5.8 on page 26): Retired people are more likely to use passive withdrawal mechanisms.

People who participate in risky activities more actively are more likely to use active withdrawal mechanisms. Men are more likely to use active withdrawal mechanisms. The use of passive withdrawal mechanisms almost doesnt change at all with the increase of the amount of actions on the contract. Note that negligent people are more likely to use active investment instruments which means that there is no self-correction mechanism. Sophisticated negligent agents would prefer to use passive instruments in order to behave more rationally and avoid possible losses coming from their negligence. The fact that people who engage in risky activities more often are more likely to use active withdrawal instruments is quite expected. When people engage in risky activities the outcomes of their actions become less predictable implying a higher level of irregularity. Lastly I would like to point out the fact that men are more likely to use active instruments of both types. This is a peculiar gender dierence which I have no explanation for.

3.4

Duration of relationship with AXA

The variable dur axa from the AXA panel gives us the duration of the clients relationship with AXA in years that reaches values from 0 (= less than one year) to 42 years. I also studied the eects on time - active time over all contracts (at least one contract with PM>0).

CHAPTER 3. BASIC ANALYSIS

16

Apart from the obvious results such as the fact that older / retired people and wealthy people are more likely to have a longer relationship with AXA there are also some more interesting relationships. Notable eects for time and dur axa (see table 5.9 on page 26): People with good mathematical abilities are more likely to have longer contracts. Ambiguity-averse people are more likely to have longer contracts. Risk-averse people are more likely to have longer relationship with a life insurance company. People with a high share of risky support in their portfolio are less likely to have long relationship with a life insurance company. People who like to gamble are less likely to have long relationship with a life insurance company.

Chapter 4 Models
In this section I will present some of the regression models that i used to analyze the data. Most of the focus will be on the uc share variable. The results attained in this study are not very rich (the most complex model for uc share explains only 12% of the variability in the data) but nevertheless allow us to make some interesting conclusions.

4.1

Linear model for uc share

The rst model I used was the standard linear regression model (see table 5.10 on page 27). It can explain 9% of the variability in the uc share. I consider this a decent result given the complexity of uc share measure and the decisions that form its value. The variables that have signicant impact in the linear regression model are: riskav, Q1 S20 (monthly income), dur axa (length of relationship with AXA), d aa (external agents) and time. Sadly none of the character type and risk perception variables show direct impact on uc share. But as I have shown in section 2.5 they are implicitly in this equation since they inuence riskav. The fact that this model is not very informative can be primarily attributed to two reasons: non-linear data generating process and the complexity of uc share. To battle the complexity we could either use more variables and a larger sample or construct an experiment that would simulate the choice of the share of risky assets in the portfolio in a simpler setting. To battle the tness of the model we could study the data generating process and propose a non-linear model that would better t the way that the investors use to make their decisions. 17

CHAPTER 4. MODELS

18

In section 4.3 I construct a model that improves the results above by analyzing the data generating process for uc share.

4.2

Robustness check

In order to verify the results obtained by the regression in section 4.1 I extracted yearly average values of uc share (from 2002 to 2011). First I calculated the correlation between uc share and its yearly values (see table 5.11 on page 27) and found out that all of the yearly values are closely related to the average uc share (the lowest correlation coecient is 0.85). Then I ran same linear regression model on every yearly value and extracted the p-values to look at the signicance of dierent explanatory variables at dierent time frames (see table 5.12 on page 27). I found out that Q1 S20 (monthly income) and dur axa (length of relationship with AXA) perform extremely well at all time frames, riskav & d aa (external agents) are signicant in majority of time frames and time is the least robust variable (signicant in only 30% of yearly regressions).

4.3

Hurdle model for uc share

A hurdle model is a modied count model in which there are two processes, one generating the zeros and one generating the positive values. In our case the variable of interest in uc share. Hurdle model assumes that investors make two decisions: rst they decide whether to invest any money into risky assets or not, then they decide exactly how much to invest into risky assets.

Figure 4.1: The hurdle model for uc share

CHAPTER 4. MODELS

19

This model makes sense in the context of decisions on purchasing risky assets because before the investor chooses what amount of risky assets to buy he rst decides whether he wants to have any risky assets or not. The results of separate correlation analysis (see tables 5.13 and 5.14 on page 28) show us that the two separate decision are inuence by two separate sets of variables. This means that hurdle model is indeed better tted to describe the decision process in our case. The decision to buy risky assets is inuenced by mostly behavioral traits such as Q1 Q1e (impulsiveness), nancial literacy, cognitive abilities and education level. The choice of the amount of risky assets to purchase on the other hand in inuenced by mostly socioeconomic and demographic factors such as Q1 S20 (monthly income) and age. It is also heavily inuenced by the clients relationship with AXA: dur axa (duration of the relationship), d aa (use of external agents) and the amount of investment. We can see that the hurdle model is a denite improvement. Personality traits like impulsiveness aect the initial choice to purchase risky assets introducing some level of irrationality at this stage. At the second stage investors choose the amount of risky assets that they want to purchase by relying on socioeconomic and demographic factors making this decision quite rational. For the results of applying hurdle model to uc share via regression analysis see tables 5.15 and 5.16 on page 29. Here I would like to note that using the two-step model increased the explanatory power of the linear model by more than 30%. The new linear model (uc share>0) explains 12% of the variability in the uc share.

Chapter 5 Conclusions
Comments on the results: In the end I believe that the most important result of my study was nding the connection between the aversion to risk / personality traits and the choice of the share of risky assets in the portfolio by introducing an intermediary measure of degree of engagement in risky activities. This measure is quite simple and easy to implement. It is based on questions that are easy for the participants to understand as opposed for example to the risk tolerance measure. Plus the fact that this measure is closely connected with the personality traits makes us more certain in its practical applicability. The fact that the choice of aggregate level of risk in the portfolio is a mostly rational decision is also quite important. As evident in the section 4.3 only irrationality comes from the inuence of personality trait factors at the rst stage of the decision process when the investors decide whether to purchase any risky assets or not. This means that in the absence of the inuence of behavioral biases in the decisions related to certain risky assets the choice of risky assets would be quite rational in general. It also means that by consciously forcing themselves to maintain a diversied portfolio of many dierent risky assets investors could get rid of the irrationality. Remarks for further research: With regards to the direction of future studies and possible improvements I would like to pay attention to the fact that in this study I only count the risk coming from having risky support in the portfolio. This does not take into account the fact that a person with steady income can allow himself to have a riskier portfolio and a person with an unsteady income will likely choose to make only safe investments. 20

CHAPTER 5. CONCLUSIONS

21

There can also be other factors of nancial risk that can aect the decision. So in further studies I would propose to include the measure for overall safety of the investors own nancial situation as an explanatory variable for the share of risky assets in the portfolio. Also one of the possible space for improvements would be using more objective measures for the personality trait variables instead of self-reported ones. There are a lot of psychological tests that contain questions used to identify the character of the subject and using such questions could bring us better data. The problem here is the fact that collecting good psychological data requires the use of long and quite tiresome questionnaires making it harder to connect this data with the information on actual investment decisions. In my opinion it would be interesting to conduct an experiment where the decision process for the participant would be a simplied variant of uc share decision process. The participants would ll in the questionnaires and then participate in a dynamic game where they would have as information their nancial situation (wealth, income, employment stability) and would be required to choose their investment strategies using the pool of dierent investment opportunities and the information on the returns in the previous period. If the game would be run for 36-60 steps (3-5 years of monthly decisions) I believe it would be able to closely simulate the actual decision process of individual investors.

CHAPTER 5. CONCLUSIONS

22

Appendix: Tables and gures


Table 5.1: Legend for variables with ambiguous naming Q1 Q14 Q1 Q18a Q1 Q1c Q1 Q1d Q1 Q1e Q1 Q1f Q1 Q25m Q1 Q4a Q1 Q4c Q1 Q4e Q1 Q17 Q1 S12 Q1 S13 Q1 S18 Q1 S1A Q1 S1R Q1 S20 Q1 S6 Q2 Q1 Q2 Q15 n Q2 Q2 Q2 Q5 Q2 Q6 Q2 Q9 Q2 beta cogn d aa dur axa qual axa time Level of mathematical abilities Attitude to managing your nances: 1 Condent - Suspicious 7 Character type: 1 Methodical/orderly - Negligent 7 Character type: 1 Shy - Self-condent 7 Character type: 1 Thoughtful - Impulsive 7 Character type: 1 Very emotional - Not emotional 7 Barsky risk aversion Financial risks perception: 1 Uncontrollable - Controllable 7 Financial risks perception: 1 Anxiety - Enjoyment 7 Financial risks perception: 1 Loss - Gain 7 Amount of risky games (poker, lotto, etc.) played Educational level Are you the decision-maker in your household? How many children do you have? Are you retired? Employment type What is the level of the net monthly income of your household? What is the range of your total assets? Risk aversion on gains Financial literacy Risk premium Ambiguity aversion on gains Ambiguity aversion on losses Risk aversion on losses Time discounting factor cognitive abilities dummy for reseau AA (contracts sold by external agents) duration of relationship with AXA quality of the client (by AXA) observed duration of relationships with AXA

CHAPTER 5. CONCLUSIONS

23

Table 5.2: Q1 S1R: Employment types 1 2 3 4 5 6 7 8 9 10 farmers craftsmen, shopkeepers heads of businesses liberal professions senior management (senior professionals) middle management (intermediate professionals) employees laborers retired students

Table 5.3: Principal component analysis for riskav components Component Comp1 Comp2 Comp3 Comp4 Comp5 Comp6 Comp7 Eigenvalue Dierence 1,58707 0,477907 1,10916 0,0930667 1,0161 0,0770343 0,939062 0,00238717 0,936674 0,153219 0,783456 0,154976 0,628479 0 Proportion 0,2267 0,1585 0,1452 0,1342 0,1338 0,1119 0,0898 Cumulative 0,2267 0,3852 0,5303 0,6645 0,7983 0,9102 1

Table 5.4: Correlation analysis for riskav vs. personality variables Variable Correlation Q1 Q1d 0,1709 Q1 Q1e 0,0744 Q1 Q1f 0,1726 Q1 Q4c 0,1334 Q1 Q4e 0,0950 Q1 S13 0,0765 cogn 0,1508 sex -0,1770 Rank correlation 0.1540 (p = 0.0007) 0.0760 (p = 0.0978) 0.1813 (p = 0.0001) 0.1244 (p = 0.0066) 0.0846 (p = 0.0648) 0.0772 (p = 0.0921) 0.1614 (p = 0.0004) -0.1770 (p = 0.0002)

CHAPTER 5. CONCLUSIONS Table 5.5: Correlation table for uc share Variable Q1 Q18a Q1 S12 Q1 S1a Q1 S20 Q2 Q1 Q2 Q15 n d aa dur axa nbacts PMA riskav Correlation -0,0924 0,1079 -0,1183 0,1783 -0,0747 0,0665 0,1661 -0,1599 -0,0169 -0,0993 0,1258 Rank correlation -0.0805 (p = 0.0795) 0.1183 (p = 0.0111) -0.1241 (p = 0.0066) 0.1445 (p = 0.0019) -0.0783 (p = 0.0877) 0.0834 (p = 0.0688) 0.1297 (p = 0.0046) -0.1558 (p = 0.0006) 0.1052 (p = 0.0215) -0.1539 (p = 0.0007) 0.1298 (p = 0.0045)

24

Figure 5.1: Boxplot of uc share over the employment type

CHAPTER 5. CONCLUSIONS Table 5.6: Correlation table for PMA Variable Q1 S12 Q1 S13 Q1 S18 Q1 S1A Q1 S20 Q1 S6 Q2 Q9 Q2 Q15 n age cogn d aa dur axa qual axa time uc share Q1 Q17 Correlation 0,1665 0,1089 -0,0977 0,1604 0,0748 0,2385 -0,0764 0,1007 0,2014 0,0382 -0,0887 0,228 0,3471 0,1956 -0,0993 -0,1399 Rank correlation 0.2050 (p = 0.0000) 0.1252 (p = 0.0062) -0.0921 (p = 0.0446) 0.2161 (p = 0.0000) 0.1516 (p = 0.0011) 0.3377 (p = 0.0000) -0.0996 (p = 0.0296) 0.1486 (p = 0.0011) 0.2834 (p = 0.0000) 0.1175 (p = 0.0102) -0.1671 (p = 0.0002) 0.3870 (p = 0.0000) 0.3956 (p = 0.0000) 0.3687 (p = 0.0000) -0.1539 ( p = 0.0007) -0.1362 (p = 0.0029)

25

Figure 5.2: Boxplot of riskav over the employment type

CHAPTER 5. CONCLUSIONS

26

Table 5.7: Correlation table for shAE Variable Q1 S1A Q1 Q1C age nbacts pma sex Correlation Rank correlation 0,1126 0.0833 (p = 0.0725) 0,0736 0.0890 (p = 0.0549) 0,1135 0.0918 (p = 0.0477) -0,3085 -0.5703 (p = 0.0000) 0,2741 0.1370 (p = 0.0030) -0,0933 -0.0820 (p = 0.0769)

Table 5.8: Correlation table for shAW Variable Correlation Rank correlation Q1 S1A -0,081 -0.1291 (p = 0.0259) nbacts -0,0563 -0.1290 (p = 0.0260) pma -0,2156 -0.2856 (p = 0.0000) riskav 0,0993 0.1205 (p = 0.0375) sex -0,0702 -0.1141 (p = 0.0490) Q1 Q12 0,0909 0.1341 (p = 0.0205)

Table 5.9: Correlation tables for time and dur axa time Q1 Q14 Q2 Q5 dur axa Q1 Q25m uc share Q1 Q17 Correlation 0,0841 0,0943 Rank correlation 0.0949 (p = 0.0383) 0.1065 (p = 0.0200)

Correlation Rank correlation 0,0939 0.1160 (p = 0.0113) -0,1599 -0.1558 (p = 0.0006) -0,1139 -0.1063 (p = 0.0202)

CHAPTER 5. CONCLUSIONS Table 5.10: Estimation results for linear regression of uc share Variable q1 s20 d aa dur axa time riskav Intercept Coecient 0.023 0.059 -0.007 0.016 0.009 0.075 (Std. Err.) (0.007) (0.025) (0.002) (0.006) (0.004) (0.050)

27

Table 5.11: Correlation table for yearly uc share values uc uc uc uc uc uc uc uc uc uc Variable share 2002 share 2003 share 2004 share 2005 share 2006 share 2007 share 2008 share 2009 share 2010 share 2011 Correlation 0,8782 0,8797 0,9129 0,9258 0,9272 0,9152 0,9083 0,9094 0,8692 0,8518

Table 5.12: Table of p-values for original model and all yearly models uc share share 2002 share 2003 share 2004 share 2005 share 2006 share 2007 share 2008 share 2009 share 2010 share 2011 # obs 477 93 276 310 332 361 385 417 458 473 465 q1 s20 0.001 0.178 0.048 0.275 0.113 0.060 0.014 0.001 0.001 0.001 0.001 d aa 0.017 0.017 0.027 0.002 0.023 0.040 0.185 0.288 0.186 0.525 0.295 dur axa 0.000 0.751 0.002 0.002 0.001 0.001 0.002 0.000 0.000 0.000 0.000 time 0.008 0.751 0.922 0.143 0.126 0.331 0.564 0.097 0.002 0.004 0.008 riskav 0.031 0.639 0.956 0.439 0.146 0.045 0.020 0.065 0.014 0.003 0.006 const 0.128 0.876 0.628 0.009 0.002 0.003 0.004 0.261 0.614 0.250 0.258

uc uc uc uc uc uc uc uc uc uc

CHAPTER 5. CONCLUSIONS Table 5.13: Correlation table for d ifuc Variable Q1 Q1e Q1 S12 Q1 S13 Q1 S1A Q1 S20 Q2 Q15 n cogn nbacts qual axa riskav Correlation Rank correlation 0,1158 0.1243 (p = 0.0066) 0,1115 0.1126 (p = 0.0157) 0,1007 0.1030 (p = 0.0245) -0,1063 -0.1063 (p = 0.0203) 0,0976 0.0913 (p = 0.0500) 0,0835 0.0948 (p = 0.0385) 0,0854 0.0940 (p = 0.0406) 0,1495 0.2159 (p = 0.0000) 0,1195 0.1153 (p = 0.0183) 0,088 0.0925 (p = 0.0436)

28

Table 5.14: Correlation table for uc share>0 Variable Q1 Q18a Q1 S20 age d aa dur axa nbacts pma riskav Correlation -0,1116 0,1624 -0,0981 0,2325 -0,2175 -0,1485 -0,2050 0,1340 Rank correlation -0.1100 (p = 0.0386) 0.1379 (p = 0.0093) -0.1033 (p = 0.0518) 0.2171 (p = 0.0000) -0.2529 (p = 0.0000) -0.1071 (p = 0.0437) -0.3516 (p = 0.0000) 0.1208 (p = 0.0228)

Note: marks the variables that have signicant correlations with uc share in original correlation analysis.

CHAPTER 5. CONCLUSIONS Table 5.15: Estimation results for probit regression of d ifuc Variable q1 q1e q1 s13 q1 s1a q1 s20 qual axa Intercept Coecient 0.129 0.310 -0.354 0.120 0.106 -0.581 (Std. Err.) (0.043) (0.107) (0.152) (0.039) (0.040) (0.270)

29

Table 5.16: Estimation results for linear regression of uc share (uc share>0) Variable q1 s20 cogn d aa dur axa riskav Intercept Coecient 0.018 -0.032 0.089 -0.005 0.008 0.334 (Std. Err.) (0.007) (0.012) (0.026) (0.001) (0.005) (0.053)

Bibliography
[BCGR08] Stephen V. Burks, Jerey P. Carpenter, Lorenz Goette, and Aldo Rustichini. Cognitive skills explain economic preferences, strategic behavior, and job attachment. IZA Discussion Papers 3609, Institute for the Study of Labor (IZA), July 2008. [Bea97] Robert B. Barsky et al. Preference parameters and behavioral heterogeneity: An experimental approach in the health and retirement study. The Quarterly Journal of Economics, 112(2):53779, May 1997. [BGJ11] Jerey V. Butler, Luigi Guiso, and Tullio Jappelli. The role of intuition and reasoning in driving aversion to risk and ambiguity. CEPR Discussion Papers 8334, C.E.P.R. Discussion Papers, April 2011. [BO11] Brad M. Barber and Terrance Odean. The behavior of individual investors. Working paper, Social Science Research Network, September 2011. [Eps99] Larry G. Epstein. A denition of uncertainty aversion. The Review of Economic Studies, 66(3):579608, 1999. [KT02] A. Kapteyn and F. Teppa. Subjective measures of risk aversion and portfolio choice. Discussion Paper 2002-11, Tilburg University, Center for Economic Research, 2002.

30

You might also like