HOUSEHOLD PORTFOLIO DIVERSIFICATION:
THE ROLE OF OVERCONFIDENCE AND
FINANCIAL ADVICE
Stijn P.M. Broekema s2002507
Faculty of Economics & Business University of Groningen
Master Thesis MSc. Finance Supervisor: Dr. Marc M. Kramer
2
TABLE OF CONTENTS
ABSTRACT ... 3
1. INTRODUCTION ... 4
2. THEORETICAL FRAMEWORK ... 7
2.2 OVERCONFIDENCE AND INVESTMENT PERFORMANCE ... 7
2.3 FINANCIAL ADVICE AND INVESTMENT PERFORMANCE ... 10
2.4 OVERCONFIDENCE AND FINANCIAL ADVICE ... 11
2.5 RESEARCH FRAMEWORK ... 13
3. DATA AND METHODOLOGY... 13
3.1 DNB HOUSEHOLD SURVEY (DHS) ... 13 3.2 VARIABLE OPERATIONALIZATION ... 15 3.2.1 OVERCONFIDENCE ... 15 3.2.2 FINANCIAL ADVICE ... 17 3.2.3 INVESTMENT PERFORMANCE ... 17 3.2.4 CONTROL VARIABLES ... 19 3.3 METHODOLOGY ... 19 4. RESULTS ... 22
4.1 OVERCONFIDENCE AND FINANCIAL ADVICE ... 23
4.2 FINANCIAL ADVICE, OVERCONFIDENCE AND RETURN LOSS ... 26
4.3 ROBUSTNESS TESTS ... 30
4.3.1 ADDITIONAL CONTROLS ... 31
4.3.3 WEIGHTING ... 32
4.3.4 USING TWO WAVES OF DATA ... 33
4.3.5 ENDOGENEITY ... 33
4.3.6 ALTERNATIVE PROXIES FOR OVERCONFIDENCE ... 34
4.3.7 OTHER MANIFESTATIONS OF OVERCONFIDENCE ... 35
4.4 WHO IS OVERCONFIDENT? ... 37
5. DISCUSSION AND CONCLUSION ... 39
5.1 DISCUSSION ... 39
5.2 CONCLUSION ... 43
REFERENCES ... 44
APPENDICES ... 47
I. FINANCIAL LITERACY MODULE (FACTOR ANALYSIS) ... 47
II. DISTRIBUTIONS OF FINANCIAL LITERACY SCORES, SELF-ASSESSED LITERACY, AND FINANCIAL ADVICE ... 48
III. RISK AVERSION (FACTOR ANALYSIS) ... 49
3
HOUSEHOLD PORTFOLIO DIVERSIFICATION:
THE ROLE OF OVERCONFIDENCE AND
FINANCIAL ADVICE
Stijn P.M. Broekema
1University of Groningen
ABSTRACT
Portfolio underdiversification is one of the investment mistakes made by individual investors that can have severe performance consequences. A prominent driver of investment mistakes often cited in the behavioral finance literature is overconfidence. Particularly, literature shows that overconfident investors tend to hold poorly diversified portfolios, and have less demand for professional financial advice. While previous studies focus on one of these relations, it is important, from both an academic as well as a policymaker’s perspective, to explicitly look at the different parts and the interrelationships between them, to provide a more complete picture of the consequences of overconfidence and to identify potential remedies. To this end, this paper applies mediation analysis using data from the DNB Household Survey (DHS) to find that overconfident people incur higher losses from underdiversification through two distinct ways. Firstly, there is a direct effect. Secondly, overconfident people are less likely to rely on a professional financial adviser, which makes them worse of as well. The findings imply that to remedy the negative welfare consequences of overconfidence, policymakers and, for example, banks, should try to identify overconfident people (some directions are provided on how to do this) to be able to intervene in both paths that lead to higher losses from underdiversification. Next, it may be possible to steer these individuals in the right direction.
Keywords: Portfolio diversification, overconfidence, financial advice, financial literacy, individual
investors
JEL classification: G2, G11, D14
1 I would like to thank Marc Kramer for his support and helpful feedback during the process of writing
4
1. INTRODUCTION
Modern portfolio theory as introduced by Markowitz (1952) assumes that investors are mean-variance optimizers. One of the consequences is that investors (should) hold a well-diversified portfolio of assets and are, therefore, not compensated for idiosyncratic risk. Holding an underdiversified portfolio is suboptimal in this framework, since it is possible to choose a different portfolio with a higher expected return and the same level of risk.
Yet, according to Goetzmann and Kumar (2008), portfolios of most individual investors contain a maximum of five stocks. Statman (1987) shows that a well-diversified portfolio consists of at least thirty stocks. Clearly, most individual investors are underdiversified when looking at stock portfolios. The costs of this underdiversification can be quite high. For example, Goetzmann and Kumar (2008) find that, compared to the most diversified investors, the least diversified investors earn an annual risk-adjusted return that is on average 2.04% lower.
But why do investors hold underdiversified portfolios? It might be that individual investors have difficulties (for whatever reason) in creating well-diversified portfolios. It could also be some investors have an informational advantage over particular stocks, which would make it rational to hold an underdiversified portfolio.
5
important to look at the entire portfolios of investors, rather than focusing on stock portfolios only (Calvet et al., 2007).
A recent paper by Von Gaudecker (2015) studies the consequences of financial literacy and financial advice on household portfolio diversification, by looking not only at stock portfolios, but also explicitly including mutual and growth funds. He finds that investors with the lowest levels of financial numeracy and who do not rely on financial advice, have the worst investment outcomes. A possible interpretation of these outcomes is overconfidence.
Furthermore, Kramer (2014) and Gentile et al. (2016) find that overconfidence plays a role in the uptake of financial advice. This implies that financial advice might play a role in the relationship between overconfidence and the costs of underdiversification. To provide a more complete picture of the consequences of overconfidence and to identify potential remedies, it is important, from both an academic as well as a policymaker’s perspective, to explicitly take financial advice into account and look at the interrelationships between overconfidence, financial advice, and losses from underdiversification.
To this end, this paper uses mediation analysis, a methodology borrowed from the psychology and marketing literature, to identify, first of all, whether overconfidence is related to losses from underdiversification, and secondly, what the role of financial advice is in this relationship. In other words, are overconfidence and losses from underdiversification only directly related, or also indirectly through the relationship between overconfidence and financial advice, and in turn the relationship between financial advice and losses from underdiversification?
6
not only focus on whether overconfidence and losses from underdiversification are related, but also explicitly on how they are related, by investigating the role that financial advice has in this relationship. To the best of our knowledge, this is the first study to explicitly look at all these relationships simultaneously.
For policymakers, the results of the study are relevant as well. Particularly, the welfare consequences of underdiversification can be quite severe (Goetzmann and Kumar, 2008; Von Gaudecker, 2015). If policymakers want to do something about this, they should first and foremost have a clear understanding of what drives (losses from) underdiversification. Next, they should have knowledge about how to remedy the negative welfare effects of underdiversification. Consequently, they can design policies to remedy these negative welfare consequences (Campbell, 2006).
Using data from the DNB Household Survey (DHS), the results show that most people appear to incur modest losses from underdiversification. Yet, overconfident investors incur significantly higher losses. Particularly, they have worse outcomes through two distinct ways. Firstly, there is a direct relationship between overconfidence and losses from underdiversification. Secondly, overconfident investors are less likely to seek financial advice from a professional, which in turn hurts their performance as well. The main implication of the findings is that policymakers should try to overconfident people, and, consequently, to make those people aware of the discrepancies between their self-assessed knowledge and actual knowledge. Furthermore, they should make them aware of the negative consequences this has on investment performance (and consequently their wealth now and in the future). If this helps in debiasing overconfident investors, this will have a positive effect on performance in two ways, namely, directly, and indirectly, through the effect on financial advice.
7
methodologies. Chapter four provides the results. The final chapter provides a discussion and conclusion.
2. THEORETICAL FRAMEWORK
This chapter contains a literature review that focuses on (the relationships between) overconfidence, investment performance, and financial advice. The derived hypotheses in this chapter will be tested in chapter four. For a quick overview of the most relevant literature, please see table 1.
2.2 OVERCONFIDENCE AND INVESTMENT PERFORMANCE
Ackert and Deaves (2010, pp. 106) define overconfidence as “the tendency for people to overestimate their knowledge, abilities, and the precision of their information, or to be overly sanguine of the future and their ability to control it”. According to Ackert and Deaves (2010, pp. 106), overconfidence can manifest itself in different ways. Firstly, overconfidence can stem from miscalibration, which means that people have the tendency to overestimate their own knowledge. Secondly, overconfidence may have to do with people finding themselves to be more competent than the average person with respect to some kind of ability (e.g., driving ability), called the better-than-average effect. Thirdly, people have the tendency to believe that the control they have over certain events is larger than it actually is, called the illusion of control. Lastly, people can be excessively optimistic, which means that they overestimate the probability of a positive outcome (like winning the lottery) and underestimate the probability of a negative outcome.
TABLE 1
LITERATURE OVERVIEW
This table provides an overview of the most relevant literature on the relationships between overconfidence, financial advice, and investment performance that are discussed in chapter 2. Authors (year) Main independent variable(s) Sample Main findings
Barber & Odean (2000) Overconfidence Household trading records from large U.S. discount brokerage firm
Households that trade the most earn the lowest returns. These households underperform the market by an average of 10.3% annually. These results are in line with an overconfidence argument: overconfident investors will put too much weight on their private information, which makes them trade too actively. Barber & Odean (2001) Overconfidence Household trading records from large U.S.
discount brokerage firm Men are more likely to be overconfident when it comes to financial matters than women, and overconfident investors tend to trade more actively. The results indicate that men trade 45% more than women.
Guiso & Jappelli (2006) Overconfidence Survey data from Italian bank Overconfidence is negatively related to information gathering. And if more overconfident people are less likely to gather sufficient information, this could reasonably be assumed to imply that overconfident investors have less demand for financial advice.
Glaser & Weber (2007) Overconfidence Survey data from German online
brokerage firm It is important to be explicit about the way in which overconfidence is measured, since the different manifestations of overconfidence may lead to different results. For instance, they find that only the better-than-average effect appears to be significantly related to trading volume.
Hackethal, Haliassos &
Jappelli (2011) Financial advice Trading records from German online brokerage firm & major commercial bank Advised investors earn higher gross returns, but lower net returns. Financial advice appears to be beneficial, but only if the incentives of the adviser are aligned with the goals of the investor. The results are stronger for bank advisers than for independent advisers.
Bhattacharya et al. (2012)
Financial advice Trading records from large German brokerage firm
A large fraction of investors does not follow the given advice. A potential explanation for this may be overconfidence. The ones who do follow the advice tend to hold more efficient portfolios.
Kramer (2012) Financial advice Individual investor data from medium-sized Dutch bank
Risk-adjusted performance is similar for investors who rely on their own judgment and investors who seek professional advice. Yet, advised investors hold better-diversified portfolios that contain less idiosyncratic risk.
Mullainathan, Noeth &
Schoar (2012) Financial advice U.S. retail advisers Advisers tend to be biased towards advising to invest in an actively managed fund, rather than an index fund. This may not be beneficial for investors, considering the fact that the associated fees may significantly reduce the net returns earned.
Kramer (2014) Overconfidence Dutch Household Survey The findings show that overconfidence is negatively related to advice-seeking: the more overconfident an investor is, the less likely he or she is to demand financial advice. Actual financial literacy is not related to advice-seeking.
Von Gaudecker (2015) Financial literacy and financial
advice Dutch Household Survey Households that rely on their own judgment when it comes to financial decision-making and are the least financial literate, have the worst investment outcomes. Gentile, Linciano &
Soccorso (2016) Overconfidence Survey data of Italian retail financial decision-makers The results indicate that overconfident investors, who may actually be in much need for financial advice, tend to rely less on financial advice. Furthermore, financial knowledge appears to be positively related to financial advice.
This study (2016) Overconfidence and financial
9
paragraphs, literature on overconfidence and its consequences for retail investors are discussed.
The paper of Glaser and Weber (2007) relates the different manifestations of overconfidence to trading volume. The main implication of their results is that it is important to be explicit about the way in which overconfidence is measured: does it have to do with miscalibration, the better-than-average effect, illusion of control or excessive optimism? In other words, the definition and operationalization of overconfidence is important in empirical studies, since the different manifestations of overconfidence may lead to different results, and because it improves the comparability of papers on overconfidence and its consequences.
Von Gaudecker (2015) is interested in the relationship between household portfolio diversification, and financial literacy and advice. The results show that investors who score low on financial literacy and do not seek (professional) advice incur the largest losses from underdiversification. According to the author, this pattern is consistent with overconfidence. In other words, the findings seem to imply that the most overconfident investors achieve the worst investment outcomes.
Another paper on the consequences of overconfidence is the paper by Barber and Odean (2000). They find that investors who trade the most, earn the lowest net returns. The authors conclude that the most likely explanation for the finding is overconfidence: “overconfident investors will overestimate the value of their private information, causing them to trade too actively and, consequently, to earn below-average returns”.
10
Therefore, men are likely to trade more than women. The findings indicate that this indeed appears to be the case: men trade 45% more than women on average.
Based on the previous discussion, the first hypothesis is: H1: Overconfidence has a negative effect on investment performance.
2.3 FINANCIAL ADVICE AND INVESTMENT PERFORMANCE
According to Bhattacharya et al. (2012), the market for financial advice is large. Estimates from the U.S. show that the market for financial planning and advice is roughly 40 billion dollars. The relatively large size of this market yields the intuition that taking financial advice can be beneficial for individual investors.
Bhattacharya et al. (2012) study whether financial advice makes investors to hold more efficient portfolios. They find that a large fraction of the investors does not follow the given advice. A potential reason for this may be overconfidence: the more overconfident an investor, the less likely he or she is to take financial advice (Kramer, 2014). Yet, the ones who do follow the advice, hold more efficient portfolios. Furthermore, it appears that the investors who can be most benefitted by taking professional advice, are least likely to actually do this. The findings imply that financial advice is “a necessary but not sufficient condition for benefiting retail customers” (Bhattacharya et al., 2012).
Hackethal et al. (2011) find that, even though advised investors obtain larger gross returns, the net returns are lower than those of unadvised investors. The findings imply that financial advice is beneficial, but only if the incentives of the adviser are aligned with the goals of the investor. Otherwise, the adviser may advice the investor to make relatively many purchases (so that the adviser obtains commission fees) which increases the transaction costs of the investor, which may lower the net returns earned by the investor.
11
tend to be biased towards advising to invest in an actively managed fund, rather than an index fund. This may not be beneficial for investors, considering the finding of Hackethal et al. (2011) that the associated fees may significantly reduce the net returns earned.
Kramer (2012) finds that, firstly, when it comes to returns and timing skills, advised and unadvised investors perform similarly. It might be that, in line with the suggestion of Hackethal et al. (2011), conflicts of interest may make an advised investor not better off than an unadvised investor on average. Secondly, it appears that advised investors perform better with regards to portfolio diversification. This also means that advised portfolios contain less idiosyncratic risk. Furthermore, investors who switch to financial advice, hold better-diversified portfolios thereafter. The findings show that financial advice adds value for individual investors, but mainly in promoting well-diversified portfolios.
Furthermore, Von Gaudecker (2015) finds that households that rely on some kind of financial advice (professional or through private contacts) perform relatively well in terms of investment outcomes. On the other hand, households that rely on their own judgment have worse investment outcomes.
Despite some mixed evidence, the following is hypothesized: H2: Financial advice has a positive effect on investment performance.
2.4 OVERCONFIDENCE AND FINANCIAL ADVICE
As discussed previously, Bhattacharya et al. (2012) find that a large fraction of advised investors does not actually follow the given advice. One of the likely explanations for this finding is overconfidence, since overconfident people think they know more (or have better abilities) than they actually do.
12
gather sufficient information, this could reasonably be assumed to imply that overconfident investors have less demand for financial advice.
Furthermore, Kramer (2014) and Gentile et al. (2016) find that overconfidence is negatively related to seeking (professional) financial advice: the more overconfident an investor is, the less demand an investor has for financial advice.
Based on the foregoing, the following hypothesis is derived:
H3: Overconfidence has a negative effect on the propensity to seek professional financial advice. When looking at hypothesis 2 and 3, it appears that financial advice could play an important role in the relationship between overconfidence and investment performance. In other words, overconfidence may be directly related to investment performance, but also indirectly through its effect on financial advice (i.e., financial advice may “mediate” the relationship between overconfidence and performance). For instance, it might be that the relationship between overconfidence and performance goes entirely via financial advice. This would mean that overconfident people achieve worse investment outcomes solely because they are less willing to rely on a professional financial adviser. If this indeed appears to be the case, policymakers should find ways in which overconfident people can be made more willing to take financial advice (or could be made less overconfident). Yet, it might as well be that there are both a direct and an indirect relationship.
Therefore, the last hypothesis is:
13
2.5 RESEARCH FRAMEWORK
The derived hypotheses are graphically presented in figure 1.
3. DATA AND METHODOLOGY
This chapter discusses, firstly, the data from the DNB Household Survey. Next, the operationalizations of the variables will be discussed, followed by the methodology.
3.1 DNB HOUSEHOLD SURVEY (DHS)
Use is made of data from the Dutch Central Bank (DNB) Household Survey (DHS),
which is part of the CentERpanel.2 The survey is administered through the Internet
(DNB Household Survey Codebook 2005). To prevent a selection bias, households that do not have Internet access are provided with access by CentERdata. The data is considered to be of high quality and a close representation of the Dutch population. The DHS data is used for two primary reasons: firstly, the survey does not only ask respondents to disclose whether they invest in a certain asset class (e.g., stocks), but also to disclose in what specific assets they invest and for what amounts. This yields the opportunity to analyze the risk and return properties of households’ portfolios. Secondly, the data contains a large variety of variables that can be used as control variables.
2 Access to the data can be obtained via www.centerdata.nl.
Figure 1 Research framework
14
More specifically, date from the 2005 wave of the DHS is used. The main reason for using data from this wave is that an extra module was added to the survey in this year. The added module was designed to measure actual (instead of self-assessed) financial literacy. Please see Van Rooij et al. (2011) for a detailed discussion on the module. Following Kramer (2014), this extra module can be used to derive a proxy for overconfidence (i.e., self-assessed literacy can be compared to actual financial literacy). The extra module on financial literacy was send out to 2,028 households, of which 1,508 actually responded. Households that responded to the additional module, but for which no information on their main source of advice is available, are excluded from the sample. This yields a sample of 1,277 observations. Of these 1,277 people, 350 are investors (i.e., people who invest in at least one of the following: shares, funds, bonds
or options). At least one item is reported by 328 of the 350 investors.3 Of these 328
portfolios, 276 remain by excluding the portfolios for which no item can be matched to a return series. Furthermore, in all return loss analyses, observations of which less than 30% of the risky financial portfolio could be recovered are excluded, after which
257 portfolios remain.4
Table 2 provides descriptives of the full sample the subsample of investors, and for the subsample of investors for which detailed portfolio characteristics can be derived. The main variables in table 2 (Panel A) will be discussed in more detail in the following
sections.5
3 Only items are reported for mutual funds and shares. The characteristics of the bonds and option parts
of the portfolios are thus unknown, and are assumed to behave similarly to the other parts of the portfolio. Fortunately, the bond and option holdings are small (less than 5%), so it is unlikely that this has severe consequences for the results.
4 Even when only considering portfolios of which almost 100% can be recovered, the results remain
similar. Clearly, due to the smaller sample sizes, significance is not always reached (even though the results are in basically all cases significant or close to significance).
5 The focus will be on the full sample and the subsample of investors. This is to avoid repetition, since
15
TABLE 2
DESCRIPTIVE STATISTICS
This table provides descriptive statistics for the main variables (Panel A) and the control variables (Panel B) for the full sample, the subsample of investors, and the subsample of investors with detailed portfolios. The numbers are based on the 2005 wave of the DNB Household Survey. Standard deviations are provided in parentheses, except for dummy and categorical variables. The full sample consists of 1,277 observations, the subsample of investors of 350, and the subsample of detailed portfolios of 257.
Panel A. Main variables Full sample Investors only Detailed portfolios
Professional financial adviser 27.1% 27.7% 24.1%
Return loss Mean (S.D.) 0.57 (0.99) Min 0.01 Median 0.28 Max 8.77 Overconfidence Mean (S.D.) 0.00 (1.00) 0.19 (1.00) 0.23 (0.96) Min -1.93 -1.93 -1.93 Median -0.25 -0.27 -0.28 Max 3.26 2.76 2.76 Measured literacy Mean (S.D.) 0.04 (1.01) 0.51 (0.67) 0.60 (0.50) Min -2.54 -2.54 -2.54 Median 0.47 0.73 0.76 Max 0.93 0.93 0.93 Self-assessed literacy Mean (S.D.) 2.15 (0.69) 2.35 (0.69) 2.40 (0.67) Min 1 1 1 Median 2 2 2 Max 4 4 4
Panel B. Control variables Full sample Investors only Detailed portfolios
Age 50.3 (15.0) 53.7 (14.3) 55.1 (13.8)
Male 57.0% 71.1% 76.3%
Total wealth €161,035 (€188,394) €257,059 (€239,817) €270,279 (€240,990)
Gross income €29,912 (€22,861) €37,254 (€25,026) €39,810 (€27,120)
Primary education Higher vocational or university, 39.2% Higher vocational or university, 52.5% Higher vocational or university, 52.3%
Primary occupation Employee, 50.8% Employee, 54.3% Employee, 50.6%
Partner 67.7% 72.9% 74.3%
Kids 34.8% 30.3% 28.4%
3.2 VARIABLE OPERATIONALIZATION
3.2.1 OVERCONFIDENCE
16
measure is corrected for the part that reflects actual financial literacy. This operationalization most closely represents the ‘miscalibration’ manifestation of overconfidence (i.e., the tendency of people to overestimate their knowledge).
The actual financial literacy is derived from the financial literacy module.6 More
specifically, following Von Gaudecker (2015), Van Rooij et al. (2011), and Kramer (2014), a factor analysis is carried out. To this end, two dummies are created for all the eleven advanced questions in the financial literacy module (please see Van Rooij et al (2011) for the exact wording of the questions): one dummy indicating whether the respondent answered the question correctly, and a second dummy to indicate whether the respondent chose the “do not know” option, since many “do not know” answers were given for the financial knowledge questions. It is important to take this behavior explicitly into account (Van Rooij et al., 2011). Consequently, the factor analysis is carried out on 22 dummies. Bartlett’s test of sphericity (p < .01) and the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy (KMO = 0.92) both indicate that a factor analysis is appropriate. The factor scores are obtained using Bartlett’s method (Bartlett,
1937).7 Please see table 9 of Appendix I for some descriptives on the literacy questions
and the corresponding factor loadings.
Of the eleven questions, the average number of correct answers equals 6.4 (8.2) for the full sample (subsample of investors). Only about 7% correctly answers all questions for the full sample, while this is 17% for the subsample of investors. Clearly, investors tend to score better on average. Please see Appendix II, table 10 panel A for the full distribution of correct scores.
Self-assessed literacy can be immediately obtained from the DHS data using the responses on the following question: “How knowledgeable do you consider yourself with
6 I would like to thank Prof. Alessie of the University of Groningen for providing the data belonging to
the financial literacy module.
7 The results are robust to using different factor analysis techniques and different methods to obtain the
17
respect to financial matters?”. Most people rate themselves as being “more or less knowledgeable” (59.3%), followed by “knowledgeable” (23.2%) and “not knowledgeable” (14.3%). Only about 3% of the respondents rate themselves as “very knowledgeable” with respect to financial matters. In the subsample of investors, relatively more people rate themselves in the top two categories (32.6% “knowledgeable”, and 5.1% “very knowledgeable”). Please see Appendix II, table 10 panel B for the complete distributions.
Next, the residuals (standardized, for ease of interpretability) from a regression of self-assessed financial literacy on measured financial literacy are taken to reflect the degree of overconfidence (please note that the variable thus also measures underconfidence) It appears that investors are more overconfident on average (0.19 versus 0.00 for the full sample). The distributions are right-skewed, as shown by the medians being lower than the means.
3.2.2 FINANCIAL ADVICE
Financial advice is obtained from the following question: “What is your most important source of advice when you have to make important financial decisions for the household?”. Most people rely on a “professional financial adviser” (27.1%), closely followed by “parents, friends or acquaintances” (22.9%), and “financial information from the Internet” (20.8%). The fewest people rely on “financial computer programs” (less than 1%). Most investors rely on a professional adviser as well (27.7%). Please see Appendix II, table 10 panel C for the complete distributions.
To operationalize (professional) financial advice, a dummy is created that is set equal to one if the respondent answered “professional financial adviser”, and zero otherwise.
3.2.3 INVESTMENT PERFORMANCE
18
how much expected return an investor “loses” by not choosing a portfolio, with equivalent risk as the household’s portfolio, on the efficient frontier. In other words, it quantifies the losses resulting from underdiversification.
More specifically, the return loss is defined as !"# = %&∙ (#∙ )#∙ *+*,*
-- , (1)
where !"# is the return loss for household h, %& equals the expected excess return of
the market portfolio, (# represents the weight in risky assets of household h, .& equals
the Sharpe ratio of the market portfolio, and .# represents the Sharpe ratio of the
portfolio of household h.8
The average return loss is 57 bps. Clearly, the distribution is right-skewed, since the median return loss is just 28 bps, which also shows that the median household performs relatively well. The minimum return loss is only about 1 bps, while the maximum return loss equals almost 900 bps.
Use is made of the data and code provided by Von Gaudecker (2015) to derive the
relevant measures.9 To derive the return loss, the following steps have to be taken:
1. Portfolio items are matched to return series, the betas of the individual assets are estimated by imposing the CAPM, and the variance-covariance matrix is estimated;
2. the household beta is estimated by using the vector of weights and the asset betas. The standard deviation of the household’s portfolio is estimated using the vector of weights, and the variance-covariance matrix;
3. using the household beta, the expected return and, consequently, the Sharpe ratio of the household’s portfolio are calculated;
8 *+,*
-*- measures the diversification loss, in line with Von Gaudecker (2015). Furthermore, to limit the
influence of outliers, the return loss is winsorized at the 99th percentile. Robustness is checked for
different percentiles.
19
4. plug all variables in equation (1) to find the return loss.
Following Von Gaudecker (2015), the risk-free rate is proxied for by the one-month Euribor rate, and the efficient market portfolio is proxied for by the MSCI Europe Index. The results are robust to using the MSCI World Index or AEX Index instead. 3.2.4 CONTROL VARIABLES
With respect to the control variables, Von Gaudecker (2015), Calvet et al. (2007), and Kramer (2014) are followed. Therefore, the following control variables are included (please see table 2 panel B for some descriptives):
• Age;
• gender (with female as the reference category);
• education (with higher vocational or university as the reference category);
• the natural logarithm of gross income10;
• primary occupation (with people that are neither an employee, self-employed nor retired as the reference category);
• having a partner in the household; • having kids in the household;
• and the natural logarithm of total wealth.
3.3 METHODOLOGY
Looking at the research framework in figure 1, it becomes clear that two analyses have to be carried out. A first one relating overconfidence to financial advice, and a second one explaining the return loss by overconfidence and financial advice. Yet, as part of the mediation analysis (which will be discussed in more detail later), the relationship
10 There appear to be relatively many missing values for gross income (201 out of 1,277). To increase the
20
between the return loss and overconfidence (i.e., without including financial advice) should be estimated as well. In other words, three main estimations will be done. Firstly, to test the effect of overconfidence on financial advice, the following linear
probability model11 is estimated
/01234# = 56+ )6813#+ 938:;!8"#+ <#, (2)
where /01234# is a dummy variable that equals one for households that rely on a
professional financial adviser and zero otherwise, 813# represents the degree of
overconfidence for household h, and 38:;!8"# represents a set of control variables
for household h.
Next, for the relationship between overconfidence and the return loss, the following model is estimated
!"# = 5=+ )=813#+ >38:;!8"#+ <#, (3)
where the variables are as defined previously.
Lastly, to estimate the simultaneous effects of financial advice and overconfidence on the return loss, the final model corresponds to
!"# = 5?+ )?/01234#+ )@813#+ ∆38:;!8"#+ <#, (4)
where all variables are as defined previously.
From a policy perspective, it is interesting to know whether overconfidence influences the return loss, but it would be even more interesting, if overconfidence indeed appears to influence investment outcomes, to know how this process works (Hayes, 2009). To this end, mediation analysis is appropriate, where financial advice corresponds to the mediator variable. In other words, overconfidence may have a direct effect on the return loss, but this effect may actually go through (i.e., be mediated
11 Clearly, a logit or probit type of model would be more suitable for this situation. Yet, mediation
21
by) financial advice: overconfident people are less likely to demand financial advice, and are therefore likely to achieve worse investment outcomes.
Mediation analysis is not used frequently in the finance literature, but it is common in, amongst others, the marketing and psychology literature. Since behavioral finance has a strong psychological basis, it may actually be relevant to focus more on mediation type of relationships, since it not only provides insights into whether a certain relationship exists, but also through which intermediate variables this relationship exists.
To test whether the relationship between overconfidence and return loss is mediated by financial advice, two approaches are followed.
The first approach is the methodology put forward by Baron and Kenny (1986). The first step is to identify whether a significant relationship exists between overconfidence and the return loss (see equation (3)), which corresponds to c in figure 2. Secondly, there should be a significant relationship between overconfidence and financial advice (equation (2) and a in figure 2).
Lastly, the effect of overconfidence on the return loss should become insignificant or the effect should become weaker if financial advice is included in the model as an additional control variable (equation (4) and c’ in figure 2). If the relationship between overconfidence and return loss becomes insignificant in this step, there appears to be full mediation: overconfidence is related to the return loss, but only through its relationship with financial advice. If the relationship becomes weaker, there is so-called partial mediation: overconfidence is directly related to the return loss, but also indirectly via financial advice. Furthermore, financial advice should show a significant relationship with the return loss in this last step (b in figure 2).
22
Yet, the Baron and Kenny (1986) approach has been criticized for different reasons (Hayes, 2009). For example, it appears that the methodology has relatively low power. Furthermore, it does not immediately test the significance of the indirect effect ab (rather, it assumes that if both a and b are significantly different from zero, the indirect effect ab is statistically significant as well).
Therefore, the significance of the indirect effect ab will be directly tested for by using a bootstrapping procedure (Hayes, 2009). The idea is to resample with replacement from the original sample k times to come up with a distribution of the indirect effect ab. The corresponding confidence interval shows whether the effect is significant or not.
4. RESULTS
This chapter provides the results of the different analyses. It starts by presenting the results of the relationship between overconfidence and financial advice, followed by the effect of overconfidence and financial advice on the return loss, and the mediation analysis. Several robustness checks will be discussed as well.
Overconfidence Return loss
Financial advice c’
a b
Overconfidence c Return loss
23
4.1 OVERCONFIDENCE AND FINANCIAL ADVICE
The results of the estimation of equation (2) can be found in table 3. Interestingly, financial literacy shows no relation with financial advice for both the full sample and the subsample of investors, while self-assessed financial literacy shows a significantly negative relationship with financial advice. This implies that how people rate themselves matters more for the propensity to seek financial advice than their actual knowledge. Clearly, this leaves room for overconfidence (Kramer, 2014). Indeed, the
final two columns of table 3show that overconfidence appears to have a significantly
negative relationship with financial advice: a one standard deviation increase in overconfidence corresponds on average to a three percentage points decrease in the propensity to seek financial advice for the full sample, and four percentage points for the subsample of investors. These findings are in line with hypothesis 3.
Looking at the control variables, a 10 percent increase in wealth corresponds to an increase in the propensity to seek financial advice by approximately 0.1 (0.3) percentage points for the full sample (subsample of investors).
Furthermore, there is some evidence that people who finished at most pre-university education, are self-employed, have a partner, or have kids, tend to rely more on financial advice, and retired people less.
24
TABLE 3
FINANCIAL ADVICE AND OVERCONFIDENCE, MULTIVARIATE RESULTS
This table provides estimation results of various linear probability models. In all cases, the dependent variable is financial advice, which is a dummy that equals one if the respondent relies on a "professional financial adviser", and zero otherwise. The results are provided for the full sample, and for the subsample of investors. The results are based on the 2005 wave of the DNB Household Survey. The first two columns of results have financial literacy as main independent variable, the next two columns self-assessed financial literacy, and the last two columns have overconfidence as main independent variable. The coefficients are reported along with the corresponding robust standard errors in parentheses. * p < .10, ** p < .05, and *** p < .01 (one-sided for overconfidence).
Financial advice Financial advice Financial advice
25
categories, except “Other”).12 The results are reported in table 4. The reference category
is “self-decider” in panel A, and “external advice” in panel B.
It appears that more overconfident people are less likely to rely on “parents, friends or acquaintances”, and “professional financial advisers”. For example, a one standard deviation increase in overconfidence, leads to a decrease in the odds of choosing a “professional financial adviser” (over being a self-decider) by a factor of 0.79 for the full sample. The results are similar for the subsample of investors.
TABLE 4
FINANCIAL ADVICE (MULTIPLE CATEGORIES) AND OVERCONFIDENCE
This table provides results of a multinomial logistic regression. The results are provided for the full sample, and for the subsample of investors. The results are based on the 2005 wave of the DNB Household Survey. The dependent variable is financial advice, which consists of the categories that are reported in the table. The reference category is self-decider in panel A, and external advice in panel B. The exponentiated coefficients (i.e., >1 means a positive effect and <1 means a negative effect) are reported along with the corresponding standard errors in parentheses. * p < .10, ** p < .05, and *** p < .01
Parents, friends or acquaintances Professional financial adviser
Panel A. Reference category:
self-decider Full sample Investors only Full sample Investors only
Overconfidence 0.72*** (0.06) 0.56*** (0.11) 0.79*** (0.06) 0.73** (0.10)
Controls included (see table 3) Yes Yes Yes Yes
Nr. of observations 1,228 347 1,228 347
Media Digital
Panel B. Reference category: external
advice Full sample Investors only Full sample Investors only
Overconfidence 1.28*** (0.10) 1.56*** (0.23) 1.24** (0.12) (0.23) 1.18
Controls included (see table 3) Yes Yes Yes Yes
Nr. of observations 1,119 319 1,119 319
Looking at panel B, it becomes clear that overconfident people are more likely to rely on “media”, and on “digital”. Investors appear to rely mainly on “media” instead of “external advice”. Clearly, all results point in the same direction: more overconfident
12 This second way of collapsing categories is chosen since Barber and Odean (2002) find that
26
people tend to rely less on external advice (i.e., professional or friends and family), and more on themselves.
4.2 FINANCIAL ADVICE, OVERCONFIDENCE AND RETURN LOSS
Panel A of table 5 provides univariate results for the relationship between the return loss and overconfidence. Clearly, the return loss rises steadily over the first three quartiles (from 46 bps to 54 bps), after which it increases sharply. It averages at 77 bps in the top overconfidence quartile, which is significantly higher than the lowest quartile. This supports the hypothesis that overconfident people tend to incur higher losses from underdiversification.
Similarly, panel B of table 5 reports univariate results for the relationship between the return loss and financial advice. In line with hypothesis 2, people that rely on a professional financial adviser earn a return loss that is on average 24 bps lower than people who rely on other types of advice.
TABLE 5
RETURN LOSS, FINANCIAL ADVICE AND OVERCONFIDENCE, UNIVARIATE RESULTS
This table provides the return loss by overconfidence quartile, and the return loss by type of financial advice (professional or other type of advice). The numbers are based on the 2005 wave of the DNB Household Survey. For each overconfidence quartile, and for each group of financial advice, the means and the corresponding standard errors (in parentheses) are provided. The differences between the groups along with the corresponding standard errors (in parentheses) are provided as well. The total number of observations equals 257. * p < .10, ** p < .05, and *** p < .01
Panel A. Overconfidence Return loss Panel B. Financial Advice Return loss
Lowest (0.06) 0.46 Professional financial adviser (0.06) 0.38
Second 0.51 Other type of advice 0.62
(0.11) (0.08) Third (0.13) 0.54 Difference -0.24** (0.14) Highest (0.17) 0.77
27
Clearly, these results do not provide enough evidence in favor of the hypotheses, since the results may depend on several other variables. Therefore, the multivariate estimation results of equation (3) and (4) can be found in table 6. The table provides results for three models relating to equation (3): the first one with financial literacy, the second one with self-assessed financial literacy, and the third with overconfidence as main independent variable, in line with table 3. The last column shows the estimation results of equation (4).
Actual financial literacy is not significantly related to the return loss, whereas self-assessed literacy appears to be significantly and positively related to the return loss (i.e., more confident people tend to incur a higher return loss). This leaves room for overconfidence, since people’s confidence in their own knowledge matters more for the return loss than their actual knowledge (Kramer, 2014).
Indeed, in line with hypothesis 1, overconfidence appears to be significantly and positively related to the return loss. More specifically, a one standard deviation increase in overconfidence, ceteris paribus, corresponds to a 15 bps increase in the return loss (which appears to be quite a strong effect, considering the average return loss of 57 bps in the sample). The finding implies that more overconfident people tend to incur higher losses from underdiversification.
Since it appears that overconfidence is related not only to the return loss, but also to financial advice, it is important to figure out how these effects relate to each other. To this end, as explained in section 3.2, a mediation analysis will be carried out.
According to the methodology put forward by Baron and Kenny (1986), equation (4) has to be estimated to determine whether the relationship between overconfidence and
return loss is mediated by financial advice. Please see the last column of table 6for the
28
TABLE 6
RETURN LOSS, OVERCONFIDENCE AND FINANCIAL ADVICE, MULTIVARIATE RESULTS
This table provides estimation results of various OLS regressions. In all cases, the dependent variable is the return loss. The results are based on the 2005 wave of the DNB Household Survey. The first column of results has financial literacy as main independent variable, the next column self-assessed financial literacy, the third column of results has overconfidence as main independent variable, and the last column has overconfidence and financial advice as main independent variables. The coefficients are reported along with the corresponding robust standard errors in parentheses. * p < .10, ** p < .05, and *** p < .01 (one-sided for overconfidence and financial advice).
Return loss Return loss Return loss Return loss
Financial literacy -0.20
(0.18)
Self-assessed financial literacy 0.20** (0.09)
Overconfidence 0.15** 0.14** (0.06) (0.06) Financial advice -0.20** (0.10) Age 0.02** 0.02** 0.02** 0.02** (0.01) (0.01) (0.01) (0.01) Gender (0.15) 0.26* (0.15) 0.13 (0.15) 0.13 (0.14) 0.10
Primary or pre-vocational education 0.00 0.07 0.07 0.06
(0.12) (0.13) (0.13) (0.13)
Senior vocational education (0.19) 0.14 (0.18) 0.24 (0.18) 0.25 (0.18) 0.26
Pre-university education -0.24 -0.24 -0.24 -0.25
(0.18) (0.18) (0.18) (0.18)
29
Interestingly, the effect of overconfidence remains significant when financial advice is added to the model. This indicates that there does not appear to be full mediation: there is a direct effect of overconfidence on the return loss, which is potentially accompanied by an indirect effect through the effect of overconfidence on financial advice (and the effect of advice on the return loss).
When comparing the final two columns of table 6, it appears that the effect of overconfidence decreases from 15 bps to 14 bps when financial advice is included in the model as well. This again indicates that there is no full mediation. Yet, there appears to be partial mediation.
Using a bootstrapping procedure based on Hayes (2009), the significance of the indirect effect is assessed directly. Using 5,000 bootstrap samples (with replacement), the bias-corrected confidence interval (see e.g., MacKinnon et al., 2004; Mallinckrodt et al., 2006; Hayes, 2013: 111) of the indirect effect shows that it is significant at the five percent level. The degree of mediation is equal to a little under 10 percent (calculated
by dividing the estimated indirect effect of 1.2 bps13 by the total effect of 15 bps). A one
standard deviation increase in overconfidence corresponds to a 15 bps higher return loss, of which roughly 1.2 bps can be explained by the effect of overconfidence on financial advice and in turn the effect of financial advice on the return loss (i.e., the indirect effect), while the rest of the total effect (roughly 13.8 bps) can be interpreted as the direct effect of overconfidence on the return loss. In other words, more overconfident people tend to have a higher return loss because of a direct effect, but also because they are less likely to rely on financial advice (which in turn corresponds to a higher return loss as well). This result is in line with hypothesis 4.
Looking at the control variables (focusing on the full model with overconfidence and financial advice as main independent variables), it appears that age is positively related to the return loss, and gross income negatively. Interestingly, people with a
30
partner in the household tend to earn a lower return loss on average, which is highly significant. This appears to be in line with Barber and Odean (2001), who find that, particularly for men, being married can help in achieving better performance, since the partner may influence the decisions made, and since women tend to be less overconfident than men. In fact, if the current sample is split on gender, the result appears to be entirely driven by the subsample of men.
To shed some first light on the robustness of the results, one would expect that if overconfident investors incur higher losses from underdiversification, they are probably less efficient risk-takers. In fact, when looking at table 7, overconfident investors tend to have riskier portfolios, driven by more efficient risk-taking (higher systematic risk) combined with more uncompensated risk-taking (higher idiosyncratic risk). This is in line with Odean (1998). Furthermore, relying on a professional financial adviser appears to be negatively related to idiosyncratic risk, which is in line with Kramer (2012).
TABLE 7
PORTFOLIO RISK, OVERCONFIDENCE AND FINANCIAL ADVICE
This table provides results of various OLS regressions. The dependent variables are total risk, systematic risk, and idiosyncratic risk. The results are based on the 2005 wave of the DNB Household Survey. The main independent variables are overconfidence and financial advice. The coefficients are reported along with the corresponding robust standard errors in parentheses. * p < .10, ** p < .05, and *** p < .01 Total risk Systematic risk Idiosyncratic risk Overconfidence 0.02** (0.01) 0.01** (0.00) 0.01** (0.01) Financial advice -0.04** -0.01 -0.04*** (0.02) (0.01) (0.01)
Controls included (see table 3) Yes Yes Yes
Nr. of observations 255 255 255
R2 0.073 0.046 0.092
4.3 ROBUSTNESS TESTS
31
weighting, using two waves of data, endogeneity, and other proxies for overconfidence are addressed. To not overdo the reader with tables, the main results of the robustness tests are reported in the following paragraphs.
4.3.1 ADDITIONAL CONTROLS
Firstly, following Kramer (2014), cognitive abilities is included in all models. A priori, it is unclear what the effect of cognitive abilities might be on both financial advice and the return loss. For instance, people with more cognitive abilities may actually be more overconfident, but it might as well be that higher cognitive abilities make processing all the information related to investing (efficiently) easier.
To proxy for cognitive abilities, a factor analysis is carried out on the basic questions 1 through 5 of the financial literacy module. The factor analysis is appropriate according to Bartlett’s test of sphericity (p < .01) and the KMO measure of sampling adequacy
(KMO = 0.85). Please see Appendix I table 9 for some descriptives and the factor
loadings.
Furthermore, it might be that the results are driven by people’s risk aversion: less risk averse people may be more overconfident, and more likely to rely on a financial adviser. Therefore, risk aversion is added to all specifications as well. Risk aversion is operationalized by performing a factor analysis on six questions relating to risk aversion. The factor analysis is appropriate according to Bartlett’s test (p < .01) and the KMO measure of sampling adequacy (KMO = 0.67). Please see Appendix III table 11 for some descriptives and the factor loadings.
32
Investment experience is operationalized in two different ways. Firstly, the number of years are counted in which a certain household is included in the CentERdata panel (before 2005) and holds risky assets. Secondly, to account for the fact that households are likely to be included in the panel for a limited time, the number of years in which a household holds risky assets is divided by the total number of years that the household has been observed in the panel before 2005.
Most importantly, the reported main results are robust to including these additional controls. In fact, the results become even stronger: financial advice is now significantly related to the return loss at the one percent level, and the degree of mediation increases to roughly 15%. Furthermore, all additional controls are negatively related to the return loss (and marginally significant): people with higher cognitive abilities, that are more risk averse, or have more investment experience tend to incur lower return losses. No such effects are found for the relationship between financial advice and overconfidence.
These findings imply that the additional controls appear to be important variables to consider: when controlling for cognitive abilities, risk aversion, and investment experience (next to the other controls), overconfidence and financial advice are even better able to explain the incurred return losses.
4.3.2 MUTUAL FUND FEES
The return loss regressions are re-estimated by subtracting mutual fund fees from the gross returns, based on Von Gaudecker (2015). This has negligible effects on the results.
4.3.3 WEIGHTING
33
For the part on overconfidence and financial advice, the main results are similar when sample weights are used. For the part relating the return loss to overconfidence and financial advice, the weights are scaled up by the ratio of the sum over all respondents of risky assets to the number of respondents who listed portfolio items (following Von Gaudecker (2015)), to take into account that portfolio items are missing only for those with risky assets. Again, the main conclusions remain intact.
4.3.4 USING TWO WAVES OF DATA
Clearly, the sample size is relatively limited, primarily for the return loss regressions. Therefore, it might be that the results are driven by certain households that performed particularly bad in 2005. Yet, it is unlikely that this has severe implications for the average results, since other households may have actually performed particularly well in 2005.
Still, to shed some light on this issue, all models are re-estimated by combining data from two waves (2005 and 2006) and clustering standard errors at the household level (since many households are included in both waves). The results are qualitatively and quantitatively similar, which again highlights the robustness of the results.
4.3.5 ENDOGENEITY
Endogeneity is basically always a concern in empirical research, which is difficult to overcome. Yet, to shed some light on this issue, a couple of steps are taken.
Firstly, to overcome omitted variable bias, a relatively large set of covariates is included in all specifications. Furthermore, additional controls are included as robustness checks as discussed previously.
34
sample is split on gender. The findings show that the results appear to be mainly driven by men for both the propensity to seek financial advice and the return loss. This provides support for the causality argument (since gender is clearly exogenous), but it does not rule out the possibility of reversed causality.
4.3.6 ALTERNATIVE PROXIES FOR OVERCONFIDENCE
Firstly, as alternative proxy for the miscalibration type of overconfidence, data on the respondents’ estimates of inflation are used. More specifically, a dummy is created which equals one for people whose highest and lowest inflation estimates differ by less than one percentage point, and zero otherwise. It might be that those people actually have superior knowledge and are capable of forecasting inflation with relatively high certainty. However, in case of inflation estimates, people tend to have relatively poor abilities in forecasting inflation (Mankiw et al., 2003), which implies that the dummy variable may actually reflect overconfidence. Since more knowledgeable people may be more certain about the inflation estimate, measured literacy is included as a control variable.
The inflation dummy does not appear to be significantly related to the propensity to seek financial advice for both samples. Yet, the inflation dummy shows a significantly positive relationship with the return loss.
Furthermore, a dummy is created that equals one if people never give a “do not know” answer in the financial literacy module. Controlling for their measured financial literacy, this may also indicate overconfidence in one’s own knowledge.
The effect of the dummy is only significant on financial advice for the full sample, and not for the subsample of investors. The dummy shows no significant relationship with the return loss.
35
average, and the second dummy (called high financial literacy dummy, HFL) equals one if people score above the sample median with respect to financial literacy. The second step involves creating two more variables: the first variable is the difference between HSA and HFL (i.e., the resulting outcomes are -1 for underconfident people, 0 for well-calibrated people, and 1 for overconfident people), and the second variable is a dummy that equals one if the difference between HSA and HFL is equal to one (i.e., people who assess themselves highly, while scoring low on financial literacy), and zero otherwise. These last two variables will both be used as a proxy for overconfidence (i.e., the first one most closely measures overconfidence as a scale from -1 to 1, and the second one only looks at overconfident people).
For the full sample, the difference between HSA and HFL does not have a significant effect at the ten percent level (albeit close, and in the expected direction) on financial advice. Furthermore, the overconfidence dummy appears to have a significant effect at the ten percent level. For the subsample of investors, the effect of the difference between HSA and HFL on financial advice is significant at the five percent level, which is also true for the effect of the overconfidence dummy.
The difference between HSA and HFL is significantly positively related to the return loss, while the effect is not significant for the overconfidence dummy, even though it is in the expected direction. This most likely stems from the fact that only nine people in the remaining sample get a one for the overconfidence dummy. Yet, compared to well-calibrated people, underconfident people (i.e., when the difference between HSA and HFL equals -1) appear to incur a lower return loss on average.
4.3.7 OTHER MANIFESTATIONS OF OVERCONFIDENCE
36
definitions, as to extend the analysis to also explicitly look at the different manifestations of overconfidence. The previously discussed proxies are most closely linked to miscalibration.
Unfortunately, the DHS data does not provide an opportunity to proxy for the better-than-average effect. Yet, the data can be used to derive proxies for illusion of control and excessive optimism.
More specifically, the survey includes several “locus of control” questions, which are used in a factor analysis (see Appendix IV table 12 for the exact wording of the questions and descriptives). Again, factor analysis is appropriate according to Bartlett’s test and the KMO measure of sampling adequacy (p < .01, and KMO = 0.76). Yet, it is difficult to determine when there is an illusion of control. Therefore, a dummy is created that equals one for people who score in the top decile, and zero otherwise. The results show that illusion of control is not significantly related to both the propensity to seek financial advice and the return loss.
Lastly, to proxy for excessive optimism, the following question is used: “How likely is it that you will attain (at least) the age of 80?”. A dummy is created that equals one if respondents think they have a chance of ten (on a scale from one to ten) to reach that age, and zero otherwise. Furthermore, it might be that those people are relatively healthy as compared to others, which would not make them “excessively” optimistic. Therefore, the Body Mass Index of the respondents is calculated and included as a control variable in the estimation.
37
4.4 WHO IS OVERCONFIDENT?
Since it appears that overconfidence is negatively related to the return loss, it is important from a policy perspective to be able to identify what kind of people tend be overconfident.
Please see table 8 for the estimation results of regressing overconfidence on a set of covariates. Please note that this regression is predictive in nature. The covariates are based on the variables that are used throughout this paper. The table provides results for both the full sample and the subsample of investors.
It appears that age and cognitive abilities are both negatively related to overconfidence: younger people, and people with lower cognitive abilities, tend to be more overconfident. Furthermore, people who have finished at most senior vocational education (compared to people who have finished at most higher vocational education or university), tend to be less overconfident. This is in line with, for example, Bhandari and Deaves (2006), who find that higher-educated individuals are likely to be more overconfident.
Furthermore, men, people that have a partner or children in the household, that have more investment experience, or a high locus of control, tend to be overconfident. Turning to the subsample investors, they tend to be overconfident when they are male, younger, have finished higher vocational education or university, have a partner in the household, and are less risk averse.
38
TABLE 8
EXPLAINING OVERCONFIDENCE
This table provides estimation results of two OLS regressions. In both cases, the dependent variable is overconfidence. The results are provided for the full sample, and for the subsample of investors. The results are based on the 2005 wave of the DNB Household Survey. The coefficients are reported along with the corresponding robust standard errors in parentheses. * p < .10, ** p < .05, and *** p < .01
Overconfidence
Full sample Investors only
Age -0.01*** -0.02**
(0.00) (0.01)
Gender 0.23*** 0.31**
(0.07) (0.16)
Primary or pre-vocational education -0.02 -0.18
(0.09) (0.15)
Senior vocational education -0.19** -0.33**
(0.09) (0.16)
Pre-university education -0.02 -0.03
(0.10) (0.16)
Log gross income 0.01 -0.01
39
5. DISCUSSION AND CONCLUSION
This chapter starts with a discussion of the results, followed by the limitations and future research directions. It ends with a conclusion.
5.1 DISCUSSION
Literature shows that households tend to incur relatively large losses from investment mistakes (Goetzmann and Kumar, 2008; Von Gaudecker, 2015). Yet, it is difficult to exactly figure out what the main drivers are of this finding. More and more evidence is supporting the view that certain psychological biases can explain this behavior. For example, Von Gaudecker (2015) finds that the least financially literate, who rely on their own judgment, tend to incur the highest losses from underdiversification. As a possible interpretation of this finding, the author points to overconfidence: “the individuals with the highest risk of incurring large return losses trust their own capabilities more than those of others, and they appear to overestimate the former” (Von Gaudecker, 2015).
Furthermore, Kramer (2014), and Gentile et al. (2016) show that overconfidence is negatively related to the propensity to seek financial advice. This suggests that financial advice may play an important role in the relationship between overconfidence and investment performance.
40
Furthermore, the results support the notion put forward by Bhattacharya et al. (2012) that investors who can benefit most from financial advice (i.e., those who incur high return losses, being overconfident people) are least likely to actually do this.
Yet, the study by Glaser and Weber (2007) shows that it is important to be explicit about how overconfidence is operationalized and to what manifestation of overconfidence it belongs, since the results may differ. The results highlight the importance of the notion put forward by Glaser and Weber (2007): the miscalibration type of overconfidence appears to be significantly related to the return loss and the propensity to seek financial advice, but this is much less the case for the other manifestations of overconfidence.
The findings have several implications for policymakers.14 Firstly, stimulating the
uptake of financial advice seems like a worthwhile strategy, since investors who rely on a professional financial adviser achieve better performance. In fact, the upcoming introduction of MiFID II, legislation aimed at protecting retail investors in Europe, actually tries to stimulate the uptake of financial advice, by limiting the number of
financial products available to self-deciders.15 However, this may not help in making
overconfident investors better off, since they are found to be less willing to take financial advice. Yet, since overconfident investors are found to hold riskier portfolios, and since under the new legislation risky and complex products can only be accessed through a professional adviser, it may actually be beneficial for overconfident
investors as well.16
Secondly, policymakers could intervene in the direct relationship between overconfidence and losses from underdiversification by identifying overconfident
14 Abstracting from self-selection issues and causality concerns.
15 MiFID II: What will be its impact on the investment fund distribution landscape? Available at:
www2.deloitte.com/content/dam/Deloitte/lu/Documents/financial-services/IM/lu-en-wp-mifid2-crossborder-29062015.pdf.
16 Clearly, this depends on whether overconfident investors willingly hold riskier portfolios, or whether
