1
The impact of stock market downturns on individual’s risk taking behavior and
stock market participation
Studentnr: s1708414
Name: Sietze Hylkema
2
Abstract.
--- This paper investigates if an individuals’ experience on the stock market has long-term effects on their risk taking behaviour. Using data from the Dutch Household Survey (DHS) over a 20 year period form 1996-2015 and long term stock data from the Amsterdam Exchange I calculate 2 measures of stock market crashes. I find that severe stock market crashes lead to lower stock market participation, a decrease in the fraction of assets invested in stocks and higher risk aversion. Furthermore, stock investors as a group differ from non-participants by being older, more often male and being higher educated.
---
3 I. Introduction
Portfolio theory shows that almost all households should have some assets in stocks. Assets like stocks have historically performed significantly better than relatively safe savings accounts and households could lose out a lot of wealth by not investing. However, in practice a large majority of households do not invest in risky assets, but keep most of their wealth in a savings account or in real estate. Cross-country data from the Euro system Household and Finance Consumption Survey (HFCS) show that both the willingness to take financial risks and the participation rate in stock markets (an average of 13% of the households) is relatively low in eleven countries in the Euro area (Ampudia and Ehrmann, 2014). The stock market participation in The Netherlands is in line with the average in the Euro Area: only 13% of the Dutch households invests in the stock market. This underweighting of stocks in household portfolios is often called the limited participation puzzle. This is the observation of low stock market participation despite the high historical performance of stocks. Campbell (2006) argues that most households invest adequately, but some households make serious mistakes like nonparticipation in risky asset markets, under diversification of risky portfolios and failure to exercise options to refinance mortgages. The study of stock market participation is especially relevant since a growing number of households are responsible for their own pension investments.
4
What are the main determinants of risk taking by individual investors? Do individual macroeconomic experiences influence risk preferences? What effects do extreme stock market events have on individual economic behavior? This thesis answers those questions by comparing stock market crashes against several measures of risk taking and risk attitude. The paper is structured as follows. Section 2 presents related literature and develops hypotheses. Section 3 introduces the data and the underlying methodologies. Section 4 presents the results. Section 5 discusses the conclusions and limitations.
II. Literature review
Standard economic theories assume that individuals have stable risk preferences. The business cycle leads to changes in aggregate risk taking by reducing the wealth of individuals.or by changing their beliefs about the risks they are taking. (Necker and Ziegelmeyer, 2013) The economic crisis demonstrated that the willingness to take financial risks varies over time and depends on the experiences individuals have undergone (Ampudia and Ehrmann, 2014). Moreover stock market experiences have been shown to have a long-run influence on financial risk taking independent from their own exposure. Malmendier and Nagel (2011) examined empirically whether individuals differ in their willingness to take financial risks depending on the macroeconomic history they experienced over the course of their lives. They find that households that have experienced higher real stock market returns over their lifetime tend to be more willing to take financial risks, have a higher propensity to hold stocks and hold larger amounts of stocks. Moreover, they show that more recent return experiences have stronger effects.
This finding is supported by a survey of 2000 Dutch households which examines the relationship between recent stock market price movements and expectations about future stock market The study shows that stock market expectations are influenced by the recent price movements. returns (Hurd, Van Rooij, and Winter, 2009). Previous economic and psychological literature confirms the different ways that personal experiences influence economic behavior.
5
We can find these learning effects in several strands of emperical literature.
Choi et al. (2009) report that investors who have experienced high returns in their 401(k) saving plan tend to save more in their 401(k) plan.
Kaustia and Knüpfer (2008) show that experienced IPO returns have a strong relationship with subsequent IPO participation for Finnish investors. This cannot be explained by for instance wealth effects or the IPO cycle. Personally experienced outcomes seem to be given more weight than rational.
In a study among Danish individuals who receive a sudden inheritance from their parents, a distinction is made between the impact of the wealth effect on risk taking and the impact of personal experiences. The study shows that in addition to the wealth effect, in particular first-hand personal experiences have a huge impact on risk taking (Andersen, Hanspal, and Nielsen, 2014).
However, while the learning effects of past return experiences on risk taking are well documented, the conceptual basis for these changes is still much debated.
Research to the behavior of a sample of online UK investors from September 2008 to March 2009 indicates that changes in risk taking are tied to changes in subjective expectations of risks and returns whereas risk attitude remains stable over time (Weber, Weber and Nosic, 2013). In this the paper follows a conceptual interpretation of risk taking as presented in formula1.
1. Risk taking = risk attitude + risk expectations.
Weber, Weber, and Nosic (2013) thus find that the whole difference in risk taking is due to changes in risk expectations. Their result are mirrored by (Necker and Ziegelmeyer, 2013) who find that their risk taking measure does not run via the risk attitude channel.
6
Others are even more skeptical that risk attitudes do not change.
Hoffman, Post, and Pennings (2011) surveyed Dutch brokerage clients on a monthly basis from April 2008 to March 2009 and report time varying risk attitude as well as subjective risk and return expectations
Guiso, Sapienza and Zingales (2013) find that changes in aggregate risk aversion are stronger than changes in wealth can explain. Furthermore, the change in risk aversion is also present in individuals without risky assets. They posit that these changes are larger than can be explained by updating of beliefs. They argue that an emotional response like fear leads to changes in risk attitude.
Overall, while some uncertainty remains over the issue of the variance of risk attitude, the literature is leaning towards stable risk attitudes.
Besides all the discussed ‘learning and experience’ effects, there is a vast collection of literature on an almost unlimited amount of other factors that influence risk taking and stock participation. Shum and Faig (2006) find that wealth, age and having sought financial advice is positively correlated with stock ownership. Van Rooij, Lusardi and Alessie (2011) show that financial literacy is a big determent of stock market participation. A basic understanding of financial concepts is almost a necessity in order to be able to invest in stocks. Luigi, Sapienza and Zingales (2008) find in Dutch and Italian micro data, as well as in cross-country data, evidence that people with higher trust invest more often in the stock market than low trust persons. Love (2010) finds evidence that family structure has profound impact on the portfolio allocation decisions.
7
Based on the discussed literature I arrive at the following hypotheses:
H1a: Stock investment decreases with market crashes.1
H1b : Risk attitude does not change with market crashes. 2
III. Data & Methodology
In order to investigate the hypothesis it is necessary to compare household portfolio data against data from stock returns. This study uses the Dutch National Bank Household Survey (DHS). The Dutch Household Survey exists of 6 simultaneous sub-surveys which are filled out sometime between April-December. On average about 2000 households fill out at least one of the surveys, but rarely all. Each year new households are introduced in the rotating panel to make up for attrition and to keep the panel representative for the Dutch population This survey provides data about sources of income, wealth, the demographic composition of the household, health and some psychological concepts. Data from 1996 until 2015 is used for a 20-year horizon.
The DHS data includes information about the independent and dependent variables. An important step is to transform several variables from personal data into household data. Some variables are personal, as for instance age, education and risk attitude. As suggested by UNECE(2011), in those cases a reference person is chosen for the household. In this paper this reference person is the person who does the financial administration3. Following Malmendier and Nagel (2011) age restrictions are set and household members under the age of 18 or over the age of 75 are excluded from the panel. Household assets and income are corrected to the Euro with the price level of 2013. Finally, households with less than 1 euro in financial wealth are excluded in line from the sample. Appendix A discusses the treatment of missing data.
1
This papers errs on the side of caution and uses two-sided tests for all regressions.
2 This papers errs on the side of caution and uses two-sided tests for all regressions. 3
8
The main explanatory variable is stock investment and calculated as a ‘virtual’ stock market return based on the age of the reference person. The measure is a simple count how often a household has experienced a stock market crash of more than 10% in a given year.
The measure is calculated using the end-of-year value of the stock market, so within-year crashes are not taken into account. Furthermore, divergent effects for single-day events versus a months-long decline do not show up in the data.
The AEX index was established in 1983, therefore this measure alone does not provide enough data for the research. I combine AEX post-1983 data with an older general Dutch stock market dataset to identify crashes in the period 1952-1983 (CBS 2015).
The literature does not provide a common definition of a stock market crash. Ampudia and Ehrmann (2014) suggest taking the yearly returns of the stock market define each downturn that is more than one standard deviation from the average return as a crash. Others use different measures and for instance Aizenman and Noy (2013) use a 10% decline from peak to trough. The Dutch stock market had an average return of 9.5% over the years with a standard deviation of 21.5%. This leads to a definition of 12% of a crash. Ampudia and Ehrmann (2014) find no evidence that the specific definition of a crash leads to different results. Taking the literature into account in this paper the definition of 10% is chosen. It is sufficiently close to the measure as suggested by Ampudia and Ehrmann (2014), while having the added benefit of slightly increasing the variance within the main explanatory variable.
In this time period the stock market had 20 years with negative growth, which is 31.7% of the total period. With the -10% growth threshold there are 13 years which can be classified as crash-years. This is 65% of the years with stock downturn and 20.6% the whole time period. 13 Years is also the max amount a household has experienced a crash in the database.
Only crashes experienced as a 18+ year old are taken into account. For example, a person born in 1966 will have experienced 2 crashes in 1996 and 4 crashes in 2005. A person born in 1975 will have experienced 0 crashes in 1996 and 2 crashes in 2005.
9
instead of taking into account learning opportunities that occurred earlier. More recent events are weighted more heavily. Taking into account the relevant literature this paper uses weights to incorporate the diminishing strength of distant events.
The weighting function itself has a large influence on the measured impact of crashes. For instance more weight can be given on crashes that people experienced when they were young or more weight can be given to more recent crashes. This paper providestwo alternatives: An unweighted function, a linear weighted function where each passing year decreases the power with 1/40th. It is assumed that after 40 years a crisis loses all effects. For the stock returns I make two assumptions:
1. Only the Dutch stock market matters for investment decisions. Households are assumed not to invest in foreign assets. Research has indeed shown that individual investors have a strong home country bias and even a local bias (Coval and Moskowitz, 1999).
2. Regardless of participating in the stock market, households do experience effects from stock market fluctuations.
Consistent with the risk attitude measures of Malmendier and Nagel (2011) and Ampudia and Ehrmann (2014) I use for the dependent variables:
The first measure of risk taking is the fraction of liquid assets that households have invested in stocks. Liquid assets are defined as all financial assets, excluding retail investments and housing. Stocks are defined as the total of shares and options. Bonds are excluded from the definition because of their lower risk profile. Given the assumption that only the Dutch stock market matters, most of the bonds will be issued by the Dutch government which we can consider as risk-free. This measure aims to capture portfolio risk in a rudimentary form. Stocks and options are defined as the risky assets, while everything else is risk-free.
10
The third measure of risk taking is the absolute amount of Euro’s invested in stocks. This measure is used as an extension to the binary regression and only used on the investor- subsample. This measure aims to quantify the absolute monetary effect that stock market crashes have on risk taking.
The fourth measure is self-reported risk attitude. This dependent variable does not try to capture facets of risk taking, but rather aims to shed light on the reason behind the changes in risk taking. The literature review discussed the channels through which crashes lead to diminished risk taking. Given that risk attitude is expected to be constant all the changes in risk taking should be from changes in beliefs. A significant effect of risk attitude on risk taking is not expected.
Risk attitude is measured as follows: The survey asks several questions about risk attitude. These questions are only weakly correlated with each other, which suggests that framing, personal beliefs and interpretation differences play a large role in the answer to each question. This may make it difficult to distil an objective measure of risk attitude from the questions. In this paper the question:
“I think it is more important to have safe investments and guaranteed returns, than to take a risk to have a chance to get the highest possible returns.”
This question is answered on a scale from 1-7 where 7 is “totally agree” and 1 is “totally disagree. A low score like 1 is consistent with being risk loving, while a high score as 7 is consistent with being risk averse.
The final model has the following form:
Y𝑖t = 𝛽0 + 𝛽1Crashcountit + 𝛽2LNfinassetsit + 𝛽3LNGrossIncomeit + 𝛽4DeduHit + 𝛽5DeduLit +
𝛽6DfknowHit + 𝛽7DfknowLit + 𝛽8Dchildrenit
The subscript i refers to the household, while t refers to the time period.
Yit is the dependent variable, which is self-reported risk attitude, stock market participation or
stock allocation.
11
Depending on the dependent variables different econometric techniques are used.
Models 1A and 1B estimate the fraction of stocks on the crash count for respectively the full sample and a stockholders-only subsample. The Hausman test indicates a fixed effects model with respectively p=0.000 and p=0.0001. The Breusch-Pagan test finds heteroscedasticity at p=0.0000 and thus robust White standard errors are used. For model 1 the fixed effects linear panel model is used with robust standard errors.
Model 2A and 2B estimate fraction in shares on the weighted crash count for respectively the full sample and a stockholders-only subsample. The Hausman test indicates a fixed effects model with respectively p=0.037 and p=0.0001. The Breusch-Pagan test finds heteroscedasticity at p=0.0000 and thus robust White standard errors are used. Both regressions use a fixed effects linear panel regression model.
Model 3A & 3B estimate stock market participation in a binary outcome. Model 3A estimates it using the unweighted crashcount, while model 3B estimates it using the weighted crashcount. The Hausman test indicates fixed effects with p-values of 0.0000 for both variants. Heteroskedasticity cannot be reliable resolved in nonlinear models and thus is not tested for. A panel fixed effects logit model is used to estimate the outcomes with the unweighted crashcount and the weighted crashcount repectively.
Model 4A & 4B estimate the absolute monetary effects of the two measures of a crash. This is only done on the subsample of stockholder. The Hausman test indicates that a fixed effects model must be used with respectively p=0.5053 and p=0.1893
Model 5 tests risk attitude. The previous models all test different aspects of risk taking. This final model tests risk attitude. It is regressed using a random effects ordered logistics panel model. In this model the absolute values of the risk attitude measure are irrelevant, but the outcome is still strictly hierarchical in the sense that 3 is higher than 2 which is higher than 1. The Breusch-Pagan indicates heteroscedasticity with a p-value of 0.0000, so robust White standard errors are used.
12
Gross Income is a combination of 16 income components including: salary, profits and social security. A full specification can be found in appendix A. The income measure is log-transformed for normality. Given income restraints for stock market participation, this measure is expected to have a positive effect on stock market participation.
Financial assets are all financial assets excluding the cash value of insurance. This also excludes housing, retail investments and consumption assets like cars, boats etc. The amount of financial assets is expected to have a positive effect on stock market participation (Shum and Faig 2009) Financial assets are log-transformed for normality.
Education is defined as the highest level completed on a scale from 1-7 going from kindergarten to university. The high education dummy comprises university and vocational colleges. The low education dummy comprises kindergarten, no education yet, special education and elementary school. ‘Middle’ education is excluded to avoid multicollinearity. Based on Cole and Shastry (2008) higher education is expected to have a positive effect on risk taking, while low education is expected to have a negative effect.
Financial knowledge is measured using the DHS survey question: “How knowledgeable do you consider yourself with respect to financial matters? This can be answered on a scale from 1-4 with 1 being not knowledgeable at all and 4 being very knowledgeable. The low knowledge dummy is answer 1, while 4 is the high knowledge dummy. The somewhat knowledgeable dummy is excluded due to multicollinearity. Financial knowledge is expected to be positively correlated with stock participation (van Rooij, Lusardi, and Alessie, 2011).
The households in our database have between zero and six children. This is recoded in a dummy for having no children and any amount of children. Love (2010) examined the effect of having children and found large portfolio choice effects from having children. In line with his results negative effects on risk taking are expected.
Appendix B shows the correlation matrix of all the model variables.
IV. Summary statistics
13
1996. Of the full sample, 13.61% of households have stocks. In total 3.7% of the total assets from all households is invested in stocks. Note that the fractions of assets invested in stocks can be higher than 100%, The average household has experienced 6.45 stock crashes of more than 10%.
Looking at self-reported risk attitude most households seem risk averse with an average score of 5.11 out of 7. Where 7 is most risk averse.
In our sample households have on average 0.73 children, and 22% is highly educated while 24% is low educated.
Table 2 shows the summary statistics for stock participants versus non-participants. The t-score for a two-sample t-test shows if the difference between the means is significant. Values greater than (-)1.96 are statistically significant at the 5% level. Of all variables only the difference in debts between the two groups is not statistically significant.
Stock participants hold on average 27,2% of their financial assets in stocks. They have experienced more crashes than non-participants at respectively 6.985 times versus 6.2371 times. Their risk attitude is also lower at 4.529 versus 4.989. Even stock market participants are thus relatively risk averse.
On the financial side stock participant have higher income and more than 3 times as much financial assets. House value is about 50% higher. These variables together suggest that wealth is strongly correlated with stock market participation.
Mutual funds and bonds are often regarded as less-risky alternatives for stocks. However, stock investors invest on average more than 5 five times as much in bonds and mutual funds than households holding no stocks. This indicates that any investment decision riskier than the savings account has some kind of entry barrier.
14
respect to financial knowledge. Given this data is self-reported this could also be a perceived barrier, where self-doubt keeps people from entering while no requirement for financial knowledge exists.
Table 1. Summary Statistics – Full Sample
This table shows summary statistics for the most relevant variables in the sample. Data is for the time period 1996-2015 and all prices are adjusted to 2013 euro’s. The statistics are based on 21,537 observations, except otherwise noted.
Mean Median St. Dev. Min Max #Obser
15 Table 2. Summary statistics sub samples.
This table shows summary statistics based on having stocks and having no stocks. Data is for the time period 1996-2015 and all prices are adjusted to 2013 euro’s. The statics are based on 21,537 observations, except otherwise noted. The t-score refers to the result of a two-sample t-test to compare means. Values large than (-) 1.96 indicate statistically significant differences in the mean at a 5% level.
Stocks No stocks t-score
Variable Mean Std. Dev. Mean
16 IV. Results
This section presents the results of the various regressions.
Table 3 shows the regression results of models 1 and 2. These models have the fraction of shares as the dependent variable. Model 1 uses the unweighted crashcount on respectively the full sample and the subsample of only stockholders. Model 2 uses the weighted crashcount on respectively the full sample and the subsample of only stockholders.
Regression 1A shows that crash count and financial assets are statistically significant at the 1% level. The crashcount has the expexted negative sign and each experienced crash leads to a lower fraction invested in stocks of 0.91 percentage points. Given the average allocation to shares of 3.7% in the general population this appear to be a fairly large effect. Financial assets have a positive effect on the fraction of assets invested in shares. Low education is statistically significant at the 10% level and has a negative sign which conform to the expectations.
Model 1B uses the same model for the subsample of stockholders. Within the group of stockholders the crashcount is statistically significant at the 1% level. The effect of experienced crashes becomes stronger when looking only at investors and they reduce their fraction of assets invested in stocks with 2.80% point per experienced crash. Interestingly, financial assets have a negative effect on relative stock holdings within the group of investors. Financial assets thus seem to decrease risk taking. Combined with the positive effect in model 1A, this suggest that some financial assets are necessary before a household can invest, but once this threshold is reached more assets do not lead to more risk taking. The final interesting variable is the dummy for having children in the household. They have a large and statistically significant negative effect. This is in line with (Love, 2010) who also finds negative effects on the risk allocation of portfolio’s in household with children.
17
allocation to stocks decreases by 1.18 percentage points. In this interpretation 1962 contributes 0.15 to the weight of the crash and 1990 contributes 0.85 for a total of 1 crash. Furthermore, financial assets are also significant at the 1% level, with a positive sign. The coefficient interpretation of financial assets is the same as in model 1A.
Similar to model 1A, low education has a statistically significant negative effect at 5% level on stock allocation. In the full sample this leads to a lower ratio of 1.8 percentage points. This is in line with the effects as found by Van Rooij, Lusardi and Alessie (2011) and suggests an entry barrier if not having a basic understanding of financial concepts.
Model 2B Compares the weighted crash count against the investor-only subsample. The effect of crashes become more pronounced and are statistically significant at the 1% level. A crash leads to a decrease in stock allocation of 5.14 percentage points if the crash happens in the current year. Given that the average stockholder has 27.2% of his assets in stocks this is a quite large effect.
18 Table 3. Fraction of shares and crash count.
This table shows the fixed effects unbalanced panel regression based on model 1 & 2. Dependent variable is the fraction of assets allocated to shares.
Model 1A Model 1B Model 2A Model 2B
Crashcount Weighted Crashcount
Full sample Stock holders Full sample Stock holders VARIABLES Shares / Assets Shares / Assets Shares / Assets Shares/ Assets
Crashcount -0.0091 *** -0.028 *** (0.00172) (0.00876) Weighted Crashcount -0.0118 *** -0.0514 *** (0.00257) (0.0142) Ln Income -0.00148 0.00404 -0.000992 0.00531 (0.00160) (0.00956) (0.00159) (0.00943) Ln Financial assets 0.00974 *** -0.0335 ** 0.00927 *** -0.0394 *** (0.00162) (0.0134) (0.00160) (0.0130) D. High education 0.00502 0.00544 0.00872 0.00615 (0.00792) (0.0312) (0.00807) (0.0322) D. Low education -0.0157 * -0.00369 -0.018 ** -0.00322 (0.00878) (0.0295) (0.00905) (0.0283) D. High financial knowledge 0.00496 0.00611 0.00421 0.000610 (0.00815) (0.0204) (0.00816) (0.0204) D. Low financial knowledge -0.000617 -0.0247 -8.60e-05 -0.0244 (0.00286) (0.0350) (0.00287) (0.0361) D. children -0.00737 -0.0732 ** -0.00173 -0.0569 * (0.00637) (0.0330) (0.00603) (0.0322) Constant 0.0227 0.816 *** -0.00336 0.828 *** (0.0238) (0.149) (0.0229) (0.157) Observations 20,621 2,854 20,621 2,854 R-squared 0.018 0.043 0.015 0.045 Number of nohhold 5,814 1,002 5,814 1,002
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
Table 4 shows model 3 which looks at participation in the stock market and model 4 for the absolute amount of stocks given the decision to participate.
19
of 0.564. One extra crash increases the odds of holding stock by a factor of 0.564. Note that factors under 1 lower the eventual odds that a family is holding stock.
Financial assets s statistically significant at the 1% level increases the chance of holding stocks. Intererestingly, this is the only factor that has a positive effect on the log- odds of holding stocks. Both education dummies are significant at the 10% and 5% level for respectively the high education and low education dummy. Both high and low education lower the log-odds of holding stock, where the effect of the latter is greater. The coefficients of -0.386 and -0.558 correspond with odds ratios of respectively 0.68 and 0.57. This is compared to middle education, which is not present in the model. The odds of holding stock when having low education is 0.57 times the odds of having middle education, holding the other variables fixed.
Model 3B uses the same regression but with the weighted crash count as explanatory variable. The weighted crashcount, financial assets and low education are are all statistically significant at the 1% level. Each crash lowers the log-odds of having stocks by 0.575. This is very similar to the unweighted coefficient. Financial assets increase the log-odds on stock market participation with 0.928 Low education has a negative effect and lowers the log-odds with 0.827.
Models 1-3 have conclusively shown that experienced crashes lead to lower participation in both the amount of households who invest in stocks and also leads to a lower fraction invested in stocks for those who are already active on the stock market. Model 4 aims to quantify the amount of Euro’s that a household subtracts from their stock investments after a market downturn by looking at the monetary effects within the group of investors.
In regression 4A only financial assets are significant at a 1% level. This is expected because stock holdings are a part of the financial assets. The crashcount cannot significantly explain monetary withdraws.
20 Table 4 Effects on holding stocks and monetary change
This table shows the results of model 3 & 4. Model 3 is a fixed-effects panel logit regression where the dummy for having shares is the dependent variable. The results are expressed in log-odds coefficients. Model 4 Is a random effects panel regression where the monetary amount invested in stocks is the dependent variable. This regression is based on the subsample of people who do have shares.
Model 3A Model 3B Model 4A Model 4B
Fixed effects panel logit regression Random effects panel linear regression Crashcount Weighted crashcount Crashcount Weighted crashcount
VARIABLES D. Shares D. Shares Household Shares Household Shares
Crashcount -0.573 *** -904.4 (-0.0505) (877.0) Weighted Crashcount -0.575 *** -9,691 *** (-0.0838) (2,534) LN income -0.137 -0.0811 3,233 3,075 (-0.0847) (-0.0826) (2,344) (2,302) LN financial assets 1.006 *** 0.928 *** 29,974 *** 30,172 *** (-0.0628) (-0.0602) (3,820) (3,790) D. High education -0.386 * -0.0646 6,168 2,539 (-0.219) (-0.215) (4,218) (4,371) D. Low education -0.558 ** -0.827 *** -3,733 -2,674 (-0.251) (-0.248) (6,463) (6,479) D. High knowledge -0.126 -0.14 -6,222 -6,688 (-0.253) (-0.248) (6,551) (6,695) D. Low knowledge -0.218 -0.28 7,104 6,986 (-0.25) (-0.247) (6,442) (6,329) D. Children -0.13 0.269 -2,062 -2,601 (-0.209) (-0.202) (5,920) (5,610) Constant -319,912 *** -296,015 *** (48,468) (44,717) Observations 3,629 3,629 2,854 2,854 R-squared - - 0.103 0.109 Number of households 497 497 1,002 1,002
21
Models 1-4 have all provided evidence that risk taking indeed changes based on lifetime experiences. Model 5 investigates if risk attitude is indeed time invariant. Table 5. shows the final model. The first row shows the primary regression results. Row 2-10 shows ancillary parameters that are used to order model outcomes into one of the 7 ordered values of risk aversion. For instance, the coefficient of the first cut can be interpreted as the threshold value of an underlying latent variable that differentiates low risk aversion (score=1) from medium and high risk aversion (score = 6-7).
The primary regression results show that crashcount is significant at the 1% level and has a positive sign. A higher value of risk attitude means being more risk-averse. This results suggests that risk attitude can change and crashes lead to individuals becoming more risk averse. Other explanatory variables are also statistically significant, such as financial assets at 1% level. More financial assets increase risk aversion.
High education is statistically significant at 5% level and has a negative sign. This means that education can lead to individuals becoming more risk loving. Also, high financial knowledge is statistically significant at 1% level and leads to households becoming less risk averse.
22 Table 5 – Ordered logistic regression
This table shows the result of model 5. Risk attitude is measured on a 7-point scale where 1 means risk loving and 7 means risk averse. The model is estimated using a ordered logistic regression model with robust standard errors.
VARIABLES
Risk
attitude cut1 cut2 cut3 cut4 cut5 cut6 cut7 cut8 cut9 sigma2_u
Crashcount 0.0753 *** (0.0118) Ln Income -0.0155 (0.0300) Ln Financial assets 0.0756 *** (0.0166) D. High education -0.16 ** (0.0676) D Low education -0.130 (0.0841) DknowH -0.354 *** (0.101) DknowL -0.00579 (0.0727) Dkids -0.0223 (0.0606) Constant -4.615*** -2.790*** -2.134*** -1.299*** -0.202 1.055*** 2.967*** 8.193*** 8.527*** 3.549*** (0.361) (0.355) (0.349) (0.344) (0.343) (0.343) (0.345) (0.381) (0.381) (0.192) Observations 15,383 15,383 15,383 15,383 15,383 15,383 15,383 15,383 15,383 15,383 15,383 Number of nohhold 4,575 4,575 4,575 4,575 4,575 4,575 4,575 4,575 4,575 4,575 4,575
23 V. Robustness checks
Up until now the models have assumed no time-fixed effects. This is allowable because limited variation in outcome variables means adding time-fixed effects increases the risk of overfitting the model. The robustness check does add these time fixed effects to model 1A in table 6.
The significant effects of the crashcount and low education disappear when adding time fixed-effects. Financial assets are still highly significant at a 1% level. Adding time-fixed effects also increases the R-squared from 0.018 to 0.027. The overall model is thus not robust against adding time-fixed effects.
24 Table 6. Model 1A including time fixed effects.
Model 1A has the fraction of assets invested in shares as the main dependent variable. It uses a fixed effects linear panel regression with White standard errors. This model adds time-fixed effects.
VARIABLES Shares/ Assets
Time-fixed
effects Shares/ Assets Crashcount -0.00286 1997.Year 0.0308 *** (0.00215) (0.00545) LNinc -0.000608 1998.Year 0.0408 *** (0.00166) (0.00637) LNfinass 0.01 *** 1999.Year 0.0503 *** (0.00169) (0.00762) DeduH 0.000696 2000.Year 0.0204 ** (0.00834) (0.00846) DeduL -0.0141 2001.Year 0.0236 *** (0.00871) (0.00661) DknowH 0.00461 2002.Year 0.00925 * (0.00810) (0.00534) DknowL -0.00150 2003.Year 0.0152 *** (0.00290) (0.00528) Dchildren -0.00872 2004.Year 0.0162 *** (0.00641) (0.00497) 2005.Year 0.0137 *** (0.00519) 2006.Year 0.00901 * (0.00496) 2007.Year 0.00567 (0.00507) 2008.Year 0.00583 (0.00423) 2009.Year 0.00100 (0.00344) 2010.Year 0.00308 (0.00364) 2011.Year 0.00330 (0.00328) 2012.Year 0.00295 (0.00327) 2013.Year 0.00369 (0.00328) 2014.Year 0.00350 (0.00232) 2015o.Year - Constant -0.0405 (0.0275) Observations 20,621 R-squared 0.027 Number of nohhold 5,814
25 V. Discussion and Conclusion
This paper investigated the effect of experienced stock market crashes on stock market participation. The paper contributes to the existing literature by extending the effects of a crash on stock allocation from a simple count to a broader view which includes weighted crashes, entry decisions and monetary allocation.
The paper combines long-term return of the Dutch stock market with the DHS for the period 1996-2015. This panel approach allows to simultaneously consider the effects of crises’ and changes in income or education.
The first result of this paper is that stock market participants as a group differ from the total population by age, gender and most importantly income and wealth. Stock market participants are on average 3 years older than non-participants and overwhelmingly male. Furthermore, stock participants are more often higher educated and less often lower educated. Investors have more twice as often high financial knowledge as non-investor when looking at self-reported measures. The difference for low-knowledge is even more extreme and non-investors are almost 4 times as likely to report they have low-knowledge.
The second result is that negative downturns in the stock market indeed have a lasting effect on the participation decision. Stocks crashes drive people away from stock market participation. Both unweighted and weighted measures of stock crashes have statistically significant negative effects on the participation decision (holding shares), on the fraction of assets invest in stocks and on the monetary value of those investments. The models suggests that there’s a threshold value of financial assets before a household can participate in the stock market. Once this threshold is breached there’s a negative effect of financial assets on risk taking as measured by the fraction of assets allocated in stocks.
A third result is that contrary to classic theory risk attitude seems to change based on experiences. Self-reported financial risk-attitude is statistically significantly related to experienced crashes, financial assets, education and financial knowledge.
26
year time-period does not allow a more fitting definition of a crisis. The analysis would benefit from adding more frequent data on stock market returns and the related portfolio allocations. Furthermore, the review of the literature did not produce a common definition of extreme events in the stock market.
This paper can be further extended by looking at large positive swings. Earlier research by Malmendier and Nagel (2011) has found effects for average growth over a longer time frame. This paper and (Ampudia & Ehrmann, 2014) both find negative effect on participation as a result of extreme crashes. It would be interesting to see if positive swings have an equally strong and lasting effect. The diverging results on risk attitude between this paper and (Malmendier & Nagel, 2011) suggests that only large shocks can make an individual switch preferences. Further research can try to verify if such an effect indeed exist.
This paper has important policy implications. A large number of people in the Netherlands are responsible for their own pension investments. The current jobs market shows a trend with growing numbers of freelancers. This makes the issue of optimal portfolio allocation for households of growing importance.. An nationwide tendency to save pension funds in saving accounts can hurt demand and lead to lower overall growth. Given the current negative real returns on savings accounts it is unlikely that freelancers are able to save enough for their pension. Historically stock investments have performed better than saving accounts.
27 Bibliography
UNECE, U. N. (2011). Canberra Group Handbook on Household Income Statistics., Second Edition. United Nations.
Ampudia, M., & Ehrmann, M. (2014). Macroeconomic Experiences and Risk Taking of Euro Area Households. Working Paper.
Andersen, S., Hanspal, T., & Nielsen, K. (2014). Once bitten, twice shy: Do personal experience or wealth changes affect risk taking. Working paper.
Aizenman, J. & Noy, I. (2013) Saving and the long shadow of macroeconomic shocks, NBER working paper No. 19067
Camerer, C. (1989). An experimental test of several generalized utility theories. Journal of Risk and Unceratainty, 2 61-104.
Camerer, C., & Ho, T. (1999). Experienced-weighted attraction learning in normal for games. Econometrica, 827-874.
Campbell, J. Y. (2006). Householde finance. The Journal of Finance, 61.4 1553-1604. Centraal Bureau Statistiek. (2015).
http://statline.cbs.nl/Statweb/publication/?DM=SLNL&PA=70936NED&D1=0&D2=12,25,3 8,51,64,77,90,103,116,129,142,155,168,181,194,207,220,233,246,259,272,285,298,311,324,3 37,350,363,376,389,402,415,428,441,454,467,480,493,506,519,532,545,558,571,584,597,610 ,62. Opgeroepen op Januari 1, 2016
Choi, J., Laibson, D., Madrian, B., & Metrick, A. (2009). Reinforcement Learning and Savings Behavior. Journal of Finance, 64 2515-2534.
Cole, S. A., & Shastry, G. K. (2008). If You are So Smart, why Aren't You Rich?: The Effects of Education, Financial Literacy and Cognitive Ability on Financial Market Participation. Harvard Business School.
Coval, J. D., & Moskowitz, T. J. (1999). Home bias at home: Local equity preferences in domestic portfolios. Journal of finance, 2045-2073.
Deutsches Aktieninstitut. (2008). In turbulenten Zeiten–Aktionaszahlen rucklaufig. DAIKurzstudie, 2 1/2.
DNB Household Survey. (1993-2015). Tilburg: CentERdata .
Dohmen, T. , Falk, A. , Huffman, D. , Sude, U. , Schupp, J. , Wagner, G. G. (2011) Individual risk attitudes: Measurement, Determinants, and Behavioral Consequences. Journal of the European Economic Association, Vol 9.3 522-550
Erev, I., & Roth, A. (1998). Predicating How people Play Games: Reinforcement Learning in
Experimental Games with Unique, Mixed Strategy Equilibria. American Economic Review, no 88 848-881.
28
Hoffman, A., Post, T., & Pennings, J. (2011). How Sever Was the Impact of the Financial Crisis on Individual Investor Perceptions and Behavior>. Net spar Discussion Paper.
Hurd, M., Van Rooij, M., & Winter, J. (2009). Stock Market Expectations of Investors. Working paper no 228.
Kaustia, M., & Knüpfer, S. (2008). Do investors overweight personal experience? Evidence from IPO subscriptions. The Journal of Finance, 2679-2702.
Love, D.A. (2010). The effects of marital status and children on savings and portfolio choice, International Economic Review, 19 -253-259
Luigi, G., Sapienza, P., & Zingales, L. (2008). Trusting the stock market. the Journal of Finance, 2557-2600.
Malmendier, U., & Nagel, S. (2011). Depression Babies: Do Macroeconomic Experiences Affect Risk. Quarterly Journal of Economics, 126(1), 373-416.
Millward Brown. (2015). Retail Investor.
Nakov, A. & Nuno, G. (2015) Learning from experience in the stock market. Journal of Economic Dynamics & Contro. , 52 224-239
Necker, S., & Ziegelmeyer, M. (2013). Risk taking after the financial crisis. Mimeo, University of Freiburg.
Shum, P., & Faig, M. (2006). What explains household stock holdings? Journal of Banking & Finance, 30.9 2579-2597.
van Rooij, M., Lusardi, A., & Alessie, R. (2011). Financial literacy and stock market participation. Jounnal of Financial Economics , 101.2 449-472.
Vissing-Jorgenson, A. (1999). Limited stock market participation and the equity premium puzzle. Working paper.
29
Appendices
A. Dataset & Missing values
The Dutch Household Survey exists of 6 simultaneous sub-surveys which are filled out sometime between April-December. On average about 2000 households fill out at least one of the surveys, but rarely all.
Table 1 –Response rates for the Dutch Household survey 2014
Subject no. of persons no. of households
1. general information on the household 5120 2072
2. household and work 2320 1722
3. accommodation and mortgages 1520 1520
4. health and income 2228 1729
5. assets and liabilities 2186 1696
6. economic and psychological concepts 2439 1867
7. aggregated data on income 2227 1729
8. aggregated data on assets, 2341 1826
liabilities and mortgages
Using only fully completed survey results leads to a very small number of observations. This group in itself won’t be representative of the general population given their abnormal response rate. Furthermore, in the dataset it is impossible to see if people didn’t fill in a survey results because they didn’t have to or because they didn’t want to. For instance, only one person in generally needs to fill in wealth data, but this is not indicated in the dataset. Therefore individuals are only removed if they did not even started with a survey. If individuals only partly complete a survey it is assumed that the value is zero for the household total. In case nobody from the household filled in a variable, the household total is still classified as missing.
30 B. Correlation matrix. Crash count Wcrash
count Fracshares hhShares Risk att. Dshares LN inc LN
finass DeduH DeduL DknowH DknowL Dchildren
