Retail Financial Advice: Does One Size Fit All?

Transcription

1 USC FBE FINANCE SEMINAR presented by Juhani Linnainmaa FRIDAY, Oct. 3, :30 am 12:00 pm, Room: JKP-102 Retail Financial Advice: Does One Size Fit All? Stephen Foerster Juhani T. Linnainmaa Brian T. Melzer Alessandro Previtero September 2014 Abstract Using unique data on Canadian households, we assess financial advisors impact on their clients investment portfolios. We find that advisors induce their clients to take more risk, thereby raising clients expected investment returns. On the other hand, we find limited evidence that advisors personalize their recommendations: advisors direct clients into similar portfolios independent of their clients risk preferences and stage in the life cycle. An advisor s own portfolio is a good predictor of what type of portfolios his or her clients hold even after controlling for investor attributes. This one-size-fits-all advice does not come cheap. The value-weighted client portfolio lags passive benchmarks by more than 2.5% per year net of fees, which suggests that the average investor gives up all of the equity premium gained through increased risk-taking. Stephen Foerster is with the Western University, Juhani Linnainmaa is with the University of Chicago Booth School of Business and NBER, Brian Melzer is with the Northwestern University, and Alessandro Previtero is with the Western University. We thank Shlomo Benartzi, Antonio Bernardo, Chuck Grace, Luigi Guiso, Markku Kaustia (discussant), Antoinette Schoar (discussant), Barry Scholnick, and Dick Thaler for valuable comments. We are also grateful for feedback given by seminar and conference participants at Nanyang Technological University, Singapore Management University, National University of Singapore, Rice University, Yale University, University of Washington, SUNY-Buffalo, Federal Reserve Bank of Cleveland, University of Chicago, American Economic Association 2014 meetings, University of Alberta, NBER Behavioral Economics Spring 2014 meetings, 2014 Helsinki Finance Summit, and 2014 European Household Finance Conference. We are especially grateful to the Household Finance Advisory Council members for donating data and giving generously of their time by helping us to better understand the complexity of the mutual fund industry. Zhou Chen and Brian Held provided helpful research assistance. Address correspondence to Alessandro Previtero, Western University, 1255 Western Road, London, Ontario N6G 0N1, Canada ( aprevitero@ivey.ca). 1

2 1 Introduction The life-cycle consumption-asset allocation problem is complex: choosing how much to consume today and how to allocate savings across risky assets requires, among other things, an understanding of risk preferences, investment horizon, and the properties of asset returns and the labor-income process. To help solve this problem, many households turn to investment advisors. In the United States, almost 40% of households that own mutual funds made purchases through an independent financial planner, and a similar proportion made purchases through a full-service investment broker (Investment Company Institute 2013). Likewise, among Canadian retail investors nearly 60% of assets are in accounts directed by financial advisors and an additional 20% of assets are managed by full-service brokers (Canadian Securities Administrators 2012). Despite widespread use of financial advisors, relatively little is known about how advisors shape their clients investment portfolios. In this paper, we take advantage of unique data on Canadian households to explore two dimensions through which advisors may add value: first, through their impact on clients willingness to take investment risk and, second, through their ability to tailor investment risk to clients particular circumstances as opposed to delivering one-size-fits-all portfolios. Our analysis begins by using a regulatory shock to the supply of Canadian financial advisors to measure advisors impact on risk-taking. Household survey data show a strong correlation between portfolio risk and use of an advisor in Canada advised households allocate more of their portfolio to risky financial assets (at the expense of low-risk cash and bond holdings) but also indicate substantial sample selection, as advised clients are older, more educated and earn higher salaries. Thus, the disparity in portfolio risk between advised and unadvised households likely provides a biased measure of advisors impact, one that is confounded by unobserved differences in, for example, risk tolerance. To resolve this selection problem, we exploit a 2001 regulatory change that imposed financial reporting and capital requirements on Canadian financial advisors operating outside of Quebec. The goal of our empirical strategy is to identify the causal role of financial advice by isolating a shock to the supply of financial advisors that is unrelated to demand for advice. Using a differencesin-differences model to compare affected households to those in Quebec, we find that the regulatory 2

3 change reduced households likelihood of using an advisor by roughly 10%. Exploiting this variation within an instrumental variables model, we estimate that financial advisors increase the marginal households risky asset share by 40 percentage points. This estimate of advisors influence on risky asset shares exceeds that obtained from least squares regressions that control for observable differences. This finding suggests that advisors facilitate greater financial market participation and risk-taking, perhaps by reducing households uncertainty about future returns or by relieving households anxiety when taking financial risk. Next, we delve deeper into advisors impact on risk-taking by examining the portfolios held by advised households. Taking advantage of uniquely detailed and extensive data furnished by three large Canadian financial institutions, we ask whether advisors tailor their advice on portfolio risk. The data include transaction-level records on nearly 10,000 financial advisors and these advisors 750,000 clients, along with demographic information on both investors and advisors. Most importantly, the data include many variables such as risk tolerance, age, investment horizon, income and financial knowledge that one would expect to be of first-order importance in determining the appropriate allocation to risky assets. What determines cross-sectional variation in investors exposure to risk? In neoclassical portfolio theory, differences in risk aversion account entirely for variation in risky shares. In richer classes of models, many other factors also shape investors optimal risk exposures: an investor s stage in the life cycle, volatility of labor income, correlation between labor income and the stock market, differences in beliefs about expected return, return predictability and risk, and so forth. A more risk-averse investor should invest less in risky assets than a less risk-averse investor; young investors should hold riskier portfolios (in most models); and investors whose labor income varies in tandem with the stock market should invest less in risky assets. Studying cross-sectional variation in risky shares across investors, we test whether advisors adjust portfolios in response to such factors. We find that advisors tailor portfolios based on client characteristics, with an investor s risk tolerance and point in the life cycle being particularly important. As one would expect, more risk-tolerant clients hold riskier portfolios: the least risk tolerant allocate, on average, 50% of their portfolio to risky assets, while the most risk tolerant allocate 75%. Controlling for risk tolerance, the risky share also declines with age, peaking before age 40 and declining as retirement approaches. While risk-taking peaks at the same age as in a life-cycle fund, the risky share of advised clients 3

4 otherwise differs substantially from the pattern in a life-cycle fund: among advised clients, the risky share peaks at 75% among investors aged 35 to 39 and declines by 5 to 10 percentage points as investors reach retirement age, whereas in life-cycle funds, the risky share peaks at 90% and declines by 40 to 50 percentage points for those of retirement age. The most striking finding from our analysis of portfolio allocations, however, is that clients observable characteristics jointly explain only 11% of the variation in risky share in the crosssection of investors. That is, although differences in risk tolerance and age translate into significant differences in average risky shares, a remarkable amount of variation in portfolio risk remains unexplained. In contrast, we find that advisor fixed effects have substantial explanatory power. Advisor fixed effects raise the model s R 2 threefold, from 11% to 31%, meaning that advisor fixed effects explain almost twice as much variation in risky share as explained by the full set of client characteristics. The advisor effects are also economically significant: controlling for investor attributes, a one standard deviation movement in the advisor distribution corresponds to a 17-percentage point change in the risky share. One interpretation of this finding is that, instead of customizing, advisors build very similar portfolios for all of their clients. Another interpretation is that matching between investors and advisors leads to common variation in risky share among investors of the same advisor; in that case, advisor fixed effects stand in for omitted client characteristics that are common across investors of the same advisor. We find little support for the latter hypothesis. Our data include investor identifiers that allow us to track investors that switch advisors. For this subset of investors (and their associated advisors) we can estimate both advisor fixed effects and investor fixed effects, the latter controlling flexibly for any unobserved persistent differences across investors and the former capturing advisor-specific input. In these models, advisor fixed effects continue to be pivotal, as they explain an equally large share of the variation in risky share as do investor fixed effects. If advisors do not base their advice on investor characteristics, what explains variation in portfolios across advisors? We find that advisors may project their own preferences and beliefs onto their clients. A unique feature of our data is that we observe the portfolio allocations for advisors who maintain investment portfolios at their own firm (nearly 60% of advisors in our sample do so). For these advisors, we find that their own risk-taking is far and away the strongest predictor of the 4

5 risk taken in their clients portfolios even after controlling for advisor and client characteristics. The picture that emerges here is that no matter what a client looks like, the advisor views the client as sharing his/her preferences and life situation. We find, in addition, that older advisors recommend substantially riskier portfolios. Although there may be other explanations for this pattern, one explanation is that agency conflicts can worsen as career horizons shorten. Given that advisors do not provide completely customized advice, the puzzle in this market is the high cost of advice. Including all management fees and loads paid to advisors and mutual funds, we find that the average client pays at least 2.5% per year. Since advisors do not add value through superior investment recommendations there is no evidence of skill in the distribution of gross alphas investors net underperformance equals the fees they pay. Accounting for an equity premium of, say, 6% per year and our earlier finding that advisors raise their clients allocation to risky assets by 40 percentage points, we estimate that households gain 2.4% per year, on average, from using an advisor. That is, the average investor gives up her entire incremental investment return through fees paid for advice and money management. This finding is reminiscent of Berk and Green s (2004) competitive mutual fund industry, in which capital flows in and out mutual funds exactly to such an extent that even though managers have skill, in expectation none of that skill gets passed on to investors in the form of positive alphas. To be clear, our result does not imply that clients do not benefit from using financial advisors, only that investment advice alone does not seem to justify the fees paid to advisors. In particular, our analysis does not account for value provided through broader financial planning, including retirement, tax and estate planning. Our paper contributes to the empirical literature on the quality of financial advice. Bergstresser, Chalmers, and Tufano (2009) provide indirect evidence that advised clients earn poor returns by documenting substantial underperformance of mutual funds sold exclusively by brokers or advisors. Yet their study lacks specific data on financial advisory portfolios, which prevents them from precisely quantifying the underperformance of advised clients and examining the extent to which advisors shape their clients portfolios in response to factors such as risk tolerance and stage in the lifecycle. Mullainathan, Noeth, and Schoar (2012) evaluate the quality of advisors recommendations in the context of a field experiment, and find evidence of poor advice; advisors encourage return chasing and also direct clients toward higher cost, actively managed funds. Finally, Chalmers and Reuter (2013) find that Oregon University retirement plan participants opting for 5

6 financial advice underperform both passive investment benchmarks and the returns of self-directed plan participants. Qualitatively, our findings are similar to these studies, but an important contribution of our analysis is to provide evidence on a large and representative group of advised clients. The rest of the paper proceeds as follows. Section 2 provides background on the Canadian retail investment industry. In Section 3 we use survey data to investigate advisors effect on their clients risk-taking. Sections 4 and 5 describe our administrative data on client accounts, and present analysis of risk-taking, investment performance and the cost of advice. Section 6 concludes. 2 The Retail Investment Industry in Canada Canadian households purchase investment products and services through five main channels, three of which involve financial advice and two of which are client-directed. By far the most common choice is to invest with the help of an advisor: out of $876 billion of retail investment assets as of year-end 2010, roughly 80% are in accounts directed by an advisor (Canadian Securities Administrators 2012). 1 Non-bank financial advisors, which are the subject of our study, account for the largest portion of retail assets $390 billion, or 44% of total assets. 2 Among advisors the services vary, but the core services offered by all advisors are financial planning and investment advice. As part of financial planning, advisors help clients formulate retirement and education savings plans, first and foremost, but also arrange mortgage loans and provide insurance and estate planning in some cases. Within the scope of investment advice, advisors offer guidance on asset allocation and investment selection, and execute trades on their clients behalf. For the accounts in our sample, discretionary trading by the advisor is not permitted; each trade must be initiated or approved by the client. The range of investment products sold by an advisor depends on his securities licenses. Our analysis focuses exclusively on advisors that are licensed as mutual fund dealers, a designation which permits sale of mutual fund and deposit products, but precludes sale of individual securities 1 The two non-advisory channels, bank branch sales and self-directed accounts (including discount brokerage and mutual fund direct sales), account for 11% and 7% of retail investment assets, respectively. 2 The other two types of advisors, for which we do not have data, are full-service brokers, who oversee 20% of retail assets, and financial advisors within bank branches, who oversee 17% of retail assets. 6

7 and derivatives. 3 In addition to being licensed to sell mutual funds, some financial advisors in our sample also have licenses to sell segregated funds, labor funds, and principal protected notes. 4 Financial advisors who are licensed to distribute mutual funds in Canada do so through one of two self-regulatory organizations. The first, Mutual Fund Dealers Association (MFDA), supervised 80,132 advisors at the year-end 2010 and these advisors had a combined $271 billion in assets under advisement. The second regulator, Investment Industry Regulatory Organization of Canada (IIROC), supervised 28,598 advisors. The combined number of financial advisors in Canada licensed to distribute mutual funds is thus 108,730. Under Canadian securities legislation, advisors have a duty to make suitable investment recommendations, based on their clients investment goals and risk tolerance. To that end, advisors are required to conduct Know Your Client surveys with each client at account origination and annually thereafter. The extent of advisors fiduciary duty, however, is a gray area; it is not clear that they are required to put the client s interests before their own, though they are legally mandated to deal fairly, honestly and in good faith with their clients (Canadian Securities Administrators, 2012). This gray area is important, given the potential for agency conflicts between advisors and their clients. Agency conflicts are a concern due to the compensation scheme for advisors. Most commonly, clients pay no direct compensation to advisors for their services. Rather, the advisor earns commissions from the investment funds in which his client invests, raising the possibility that their investment recommendations are biased toward funds that pay larger commissions without providing clients better investment returns. The size and source of these commission payments vary depending on the asset class and load structure of the mutual fund purchased by the client. Commission payments are lowest on money market funds and highest on balanced funds and equity funds, which potentially skews advisor recommendations toward riskier funds. Across load structures, commissions are lower, on average, for no-load funds and higher for load funds. For no-load funds, which are so-named because the investor pays no explicit commission on purchases and redemptions, the advisor still collects a 3 Full-service brokers, who are not represented in our sample, offer access to the widest range of investment products individual securities as well as mutual funds and deposit products by virtue of being licensed as investment dealers and mutual fund dealers. 4 Segregated funds are variable life insurance contracts that reimburse capital upon death. Labor funds are funds that direct (venture capital) investments to small non-public firms. 7

8 trailing commission of up to 1% per year from the mutual fund as long as the client remains invested. For funds with a back-end load, the investor pays a fee to the mutual fund company at the time of redemption typically, the redemption fee declines with horizon, starting at 6% within one year of purchase and declining to zero after 5 to 7 years. The advisor, in turn, is paid by the mutual fund company in the form of a payment at the time of purchase (typically 5% of the purchase amount) as well as a trailing commission (typically 0.5% annualized) as long as the client remains invested. Finally, on purchases of front-end load funds, the advisor receives an upfront sales commission directly from the client (up to 5% of the purchase amount, but negotiable between the investor and advisor), along with a trailing commission paid by the fund company (up to 1% per year while the client remains invested). While the exact source of the commissions varies, ultimately these payments are funded by clients, whether directly or indirectly through management and operating expenses deducted from their fund investments. After summing up these commission payments and deducting the typical share of commissions (20%) that go to their employer, the average advisor in our sample earns revenue of $80,000 to $120,000 per year. 3 The Effect of Financial Advisors on Risk-taking: Evidence from Survey Data In this section, we use the Canadian Financial Monitor (CFM), a household survey covering both advised and unadvised households, to evaluate the impact of financial advisors on their clients risk-taking. Ipsos-Reid, a survey and market research firm, designed the CFM survey and collected the data through monthly interviews of approximately 1,000 households per month between January 1999 and June In addition to providing a wealth of demographic information, each interview measures households stock and mutual fund holdings, and their asset allocation and savings decisions. Most importantly for our analysis, the survey collects also information on the use of financial advisors. Table 1 displays descriptive statistics for Canadian households, stratified by use of a financial 5 The data are structured as a repeated cross-section, but some households do participate repeatedly, so the total number of observations (175,000) exceeds the number of unique households (79,600). 8

9 advisor. Advised households are on average four years older (56.3 vs. 52.4), 7.5 percentage points more likely to be retired (34.7 vs. 27.2), and 14 percentage points more likely to have either a college or graduate degree (69.0% vs. 54.9%). From a financial standpoint, advised household also have higher average incomes (CND $74,900 vs. 52,400), substantially higher net worth (CND $472,300 vs. 212,500) and more financial assets (CND $221,100 vs. 71,600). Last, households that use financial advisors invest more in equity (15% vs. 6.9% of financial assets), more in mutual funds (39.7% vs. 12.7%) and less in fixed income products (45.2% vs. 80.4%). These summary statistics indicate that advised households shift their portfolio allocation away from safer cash and fixed income assets to riskier equity and mutual fund assets. Given the substantial differences in other characteristics, such as income and wealth, however, it is unclear whether these differences arise due to client preference or due to advisor input. This is a fundamental challenge in measuring the impact of financial advisors: the demand for advisory services will depend on the outcomes of interest, such as the participation in risky asset markets. For example, if riskaverse individuals perceive less benefit from investment advice on mutual funds because they prefer not to own risky assets, they will be less likely to work with an advisor. This pattern in advisor selection would cause a downward bias when estimating advisors impact on risky asset shares by comparing advised and unadvised households. We address this identification issue by using a regulatory change in the early 2000s that reduced the supply of financial advisors. Specifically, as of February 2001 mutual fund dealers and their agents, such as financial advisors, were required to register with the Mutual Fund Dealers Association of Canada (MFDA) and follow the rules and regulations of the MFDA. The introduction of this registration requirement meant that dealers who wished to remain in business were now subject to more stringent regulatory standards, including minimum capital levels as well as audit and financial reporting requirements. For the underlying advisors, the registration requirement also mandated securities training and established a basic standard of conduct. 6 The draft rules and bylaws were originally posted for comment on June 16, An overview of public comments given by dealers and advisors in response to the draft proposal reveals particular concern about costs imposed by the requirement, including compliance costs associated with financial reporting 6 The standard of conduct is quite broad, prescribing that advisors deal fairly, honestly and in good faith with clients, observe high standards of ethics in their business transactions and not engage in conduct detrimental to the public interest. 9

10 and capital costs created by minimum capital standards. To the extent that these changes reduced the supply of advisors, they are useful in identifying a change in households use of advisors that is unrelated to their demand for advisory services. Importantly, the regulatory change did not apply to dealers and advisors in the province of Quebec, allowing us to use Quebec residents as a baseline from which to measure the impact of the registration requirement over time. We assess the impact of the registration requirement through the following differences-indifferences model: y ipt = α + βregister p Post t + γregister p + δpost t + θx it + ε ipt, (1) in which subscripts i, p, and t index households, provinces, and months between January 1999 and January 2004, respectively. The variable Post is an indicator that takes the value of one for dates after June 2000, when the registration requirement was announced and draft rules were published for comment. Register is an indicator variable that takes the value of one for households located in provinces that faced the registration requirement. Through β, the coefficient on the interaction of Register and Post, we measure the impact of the registration requirement over time, taking changes in Quebec as a baseline from which to measure this effect. The vector X it contains household-level controls for income, education, age and retirement status, each of which is predictive of household demand for advisory services. 7 In some versions of the model we include province and month fixed effects to control more flexibly for differences over time and across provinces. To estimate the model we use weighted least squares, incorporating survey weights from the CFM to provide regression estimates that reflect a nationally representative sample, and cluster the observations by province in calculating Huber-White standard errors. First, we estimate the impact of the registration requirement on households use of financial advisors. Table 2 Panel A reports the regression estimates from three models in which the dependent variable is an indicator variable that takes the value of one for households who use a financial advisor at the time of the survey. The baseline probability of using an advisor in these surveys is 7 Ipsos-Reid codes household income as a categorical variable, and we use indicator variables that represent these categories as controls. We control flexibly for the age of the head of household with indicator variables for 16 fiveyear age bins covering ages 20 to 100. We code education based on the maximum level of education of the head of household and spouse, and include indicators for each of four categories: high school diploma or less, some college, college degree, and graduate degree. 10

11 0.38. The estimates in the three models, which differ in terms of the inclusion of household controls and fixed effects, suggest that the registration requirement had both a statistically and economically significant effect on the use of financial advisors. The point estimates in the three models place the marginal effect between and 0.040, which translate into a proportional decrease of approximately 11% in the use of financial advisors. In the first model, which excludes household controls, the coefficient on the registration-requirement indicator is positive and marginally significant at the 10% level, indicating that before the law change residents of Quebec are less likely to use advisors than their counterparts in provinces subject to the registration requirement. However, this disparity is entirely explained by differences in income and demographics; the coefficient on Register is very close to zero once household-level controls are added to the model. This evidence helps support our premise that, after controlling for observable differences, Quebec residents can serve as an reasonable baseline from which to measure the change in advisor usage. The substantial increase in R 2 induced by the inclusion of these controls shows that income, education, age, and retirement status indeed substantially correlate with the demand for advisory services. Next, we use the variation documented above to estimate the effect that financial advisors on households financial choices in a two-stage least squares model: Use Advisor ipt = α + βregister p Post t + η p + Ψ t + θx it + ε ipt, (2) y ipt = α + β Use Advisor ipt + η p + Ψ t + θ X it + ε ipt. (3) Each regression includes both household-level controls as well as province and month fixed effects. The first stage provides an estimate of each household s predicted probability of using an advisor ( Use Advisor ipt ), allowing for variation due to the Register-Post instrumental variable, and the second stage uses this predicted probability to provide an estimate of advisors impact on risktaking and, in a income. The estimates from this instrument variables analysis, which are shown in Table 2 Panel B, are consistent with financial advisors affecting risk-taking. A household s likelihood of owning any risky assets (stocks and mutual funds) increases by 0.67, or 67 percentage points, with the use of an advisor, and the proportion of risky assets in the portfolio increases by In each case, the IV estimate exceeds the OLS estimate, which suggests downward bias in the OLS estimate, perhaps 11

12 because individuals that are comfortable holding risky assets are less likely to solicit an advisor s input. To provide validation that the registration requirement captures a supply shock that is unrelated to demand for advisors, we also test for a correlation between household income and use of an advisor. OLS analysis reveals that high-income households are significantly more likely to use financial advisors: there is a large positive OLS coefficient in regression of log income on Use Advisor). 8 Since there is no obvious channel through which financial advisors should causally influence household earnings, this correlation likely stems from differences in demand for advisors. Once we instrument for use of an advisor with the registration requirement, we indeed find no significant relationship between log-income and households use of financial advisors, which provides further comfort that the registration requirement leads to changes the supply of advisors while leaving key demand-side factors unchanged. 4 Analysis of Dealer Data: Portfolio Choices 4.1 Description of the data In the balance of the paper we use detailed, transaction-level data on the portfolios of advised clients to measure the extent to which advisors shape their clients portfolio choices over and above investor demographics, and to measure the costs of financial advice. Three large Canadian financial advisory firms supplied the data for our study. Each firm provided a full history of client transactions over a 10-year period, from 2001 to 2010, along with background information on clients and advisors. The total value of assets under advice at the end of this period was $30.9 billion, representing 11% of the assets of Mutual Fund Dealers (MFDs). Key summary statistics of these data are provided in Table 3. Table 3 Panel A describes the investor side of the sample and shows that our data cover a broad swath of different types of investors both in terms of their demographics, the length of the investor-advisor relationship, risk tolerance, financial knowledge, income, and wealth. Across the entire sample, we have data on 748,287 investors with 1.5 million accounts; 86% of these investors were active as of year-end Men and women are equally presented in the data. The median 8 This specification excludes the income controls. 12

13 investor in the data is 49 years old, and the 10th and 90th percentiles of the age distribution are 32 and 68 years. The data display considerable heterogeneity with respect to how long an investor has known his or her advisor. At the end of the sample period, over 10 percent of investors had been in the client-advisor relation for less than a year (row investor known since ); and at the end other end of the spectrum, investors in the 90th percentile of the distribution had known their advisors for at least 7 years. The first two blocks at the bottom of Panel A detail the distributions of account types and investment horizons. One-fifth of the accounts are unrestricted general-purpose accounts; 71% of accounts are classified as either retirement savings or retirement income accounts that receive favorable tax treatment comparable to the 401(k) plans in the U.S.; 4% of the accounts are education savings plans; and the remaining 4% are tax-exempt accounts that face restrictions on how much money can funds can be invested and withdrawn. The data include both open and closed accounts. As of the end-year 2010, 44% of the accounts were active; the others were either inactive or had been closed. Panel C s bottom block tabulates the self-reported time horizons of the accounts. The typical investment horizon reported for 63% of the accounts for which this information is supplied is six to ten years. One-fifth of accounts are associated with reported investment horizons greater than ten years, and the remaining 15% of accounts have shorter investment horizons. There are some very-short-term accounts as well. Some 3% of the accounts or just over 30,000 accounts are associated with an investment horizon shorter than a year. The remaining blocks at the bottom of Panel A describe investors responses to questions about their risk tolerance, financial knowledge, net worth, and income from Know Your Client forms. Financial advisors collect this information at the start of the advisor-client relationship and at the time they create new accounts for their existing clients. These investor attributes display considerable heterogeneity. The most common level of risk tolerance is the second highest at 61%. Just 13% of investors tolerate low or very low levels of risk. Two-fifths of investors report low financial knowledge and only 6% of investors report having high financial knowledge. More than half of investors report a net worth of over $200k while 20% report a net worth of $50k or less. Annual salaries display dispersion similar to that observed in net worth: 27% of investors report an annual salary of less than $30k, and 10% of investors earn more than $100k per year. Table 3 Panel B shows summary statistics for the advisors in our sample. The median advisor s 13

14 age equals that of the median investor at 50 years, and the 10th and 90th percentiles are 36 years and 63 years. Advisors differ significantly from each other in terms of their experience, the number of clients they advise, and how many and what types of licenses they have. Over 10% of advisors have been in the job for less than a year, and another 10% have at least 8 years of experience. While the median advisor advices 18 clients with a total of 29 accounts, these numbers are very different at the 10th and 90th percentiles: advisors in the bottom decile have just one client with two accounts; those in the top decile have over 200 clients with more than 400 accounts. Over a quarter of advisors have just one license either the mutual fund or dual (mutual fund and life insurance) license and just over 10% have three or more licenses. The most common non-mutual fund license is the segregated-funds license. This license allows the advisor to sell variable life insurance contracts. 9 Just every tenth and every fifth advisor have licenses to sell principal protected notes and laborsponsored funds, the latter of which are mutual fund-style vehicles investing primarily in private companies. 4.2 Portfolio choice, investor attributes, and advisor fixed effects In this section we study cross-sectional variation in investors portfolios and measure the extent to which advisors adjust their clients portfolios in response to these factors. The dealer data contain information on many variables that one would expect to be of first-order importance in explaining cross-sectional differences in investors willingness to assume equity risk. For example, although many factors may influence portfolio decisions, we would still expect risk tolerance to play a dominant role. Our analysis here proceeds in three stages. We first estimate regressions that explain cross-sectional variation in investors portfolios with investor attributes and advisor fixed effects. We then focus on a subset of investors who move across advisors to estimate models with both investor and advisor fixed effects. These regressions speak to the possible role of any omitted investor attributes as well as to the nature of the mechanism that matches investors and advisors. After establishing that advisor fixed effects are important determinants of portfolio choice over and above all investor-specific effects, we ask whether advisor attributes such as age and gender explain why some advisors give their clients far riskier or safer portfolios than others. 9 The name segregated fund comes from the requirement in the Canadian law that the funds invested in these vehicles to be separated from the company s general investment funds. 14

15 Table 4 presents estimates from regressions that explain cross-sectional variation in investors risky shares (Panel A) and home bias (Panel B). Risky share is the fraction of wealth invested in equity and home bias is the fraction of equity invested in Canadian companies. 10 The first column reports estimates from a panel regression against year fixed effects and a large swath of demographic variables summarized in Table 3. The unit of observation in these regressions is an investor-year and we cluster errors by advisor to account for arbitrary correlations in errors over time and between investors who share an advisor. The intercept of this regression, 43.2%, is the average risky share of such an investor at the year-end of 1999 who is in the lowest (omitted) category for every variable that is, an investor between 21 and 24 years of age, male, very low risk tolerance, and so forth. We exclude from the analysis clients who are advisors themselves we describe and utilize this information in Section 4.5. Risk tolerance stands out in the first regression for its statistical and economic significance in explaining cross-sectional variation in risky shares. Investors in the second lowest risk-tolerance category invest 10 percentage points more in equities than those in the lowest category, and those in the top two categories hold between 24 and 27 percentage points higher risky shares. Many other regressors are also statistically highly significant with signs typical to the literature. The age profile, for example, is hump-shaped, with investors aged between 35 and 39 having the highest risky shares. Figure 1 illustrates the age and risk tolerance profiles in the data. The thick line shows, for reference, the age profile used in Vanguard s target-date funds but all firms reference target-date portfolios are very similar. These target-date funds invest 90% in equities for investors up to the 40-year mark and then approximately linearly decrease equity exposure so that it falls to around 50% around the expected retirement date of 65. The risky-share profiles in the dealer data are very different from those of target-date funds. All investors, independent of their risk tolerance, assume too little equity exposure relative to the Vanguard benchmark when they are young and too much when they are old. The remaining regressors in Table 4 show that women s risky shares controlling for other demographics such as risk tolerance are on average 9 percentage points below those of men. 10 We assume that an all-equity fund invests 100% in equities, a balanced fund invests 50% in equities, and a fixed-income fund invests nothing in equities. We compute each investor s risky share and home bias by taking the market value-weighted average of the funds the investor holds. We set the home-bias measure to missing for those observations in which the investor has no equity exposure. 15

16 Investors with longer investment horizons assume substantially more equity risk than those with short horizons. Investors who report higher levels of financial knowledge hold slightly higher risky shares between 3 and 4 percentage more than low-knowledge investors. Investors with higher levels of income and wealth have higher risky shares relative to investors in the lowest categories, but these effects are not economically as impressive as those on the other regressors. That is, after accounting for all other investor attributes, wealth and income contribute only modestly in explaining cross-sectional variation in risky shares. The most striking fact about the first regression is that all the regressors in the model there are 28 variables in the model in all if not counting the year fixed effects jointly explain just onetenth of the cross-sectional variation in risky shares. That is, although differences in risk tolerance translate to significant differences in average risky shares, a remarkable amount of variation remains unexplained. The low explanatory power is even sharper in Panel B s home-bias regressions. The same set of regressors yields an adjusted R 2 of just 1.7% and, although some coefficients are statistically significant in isolation, no clear age, risk-tolerance, or investment-horizon patterns are apparent in the data. The lack of explanatory power in this regression is perhaps not surprising. Unlike the optimal risky share, the optimal mix of domestic and international equities should be invariant to investor-level attributes. 11 beliefs, transaction costs, or other frictions. Any variation in home bias must arise from differences in The second regression in Table 4 modifies the first by adding advisor fixed effects for the 4,984 distinct advisors who serve the 765,483 investors in the data. The inclusion of these fixed effects addresses the possibility that advisors have an influence on investors portfolio choices over and above the variation induced by heterogeneity in (observable) investor attributes. That is, this regression measures the extent to which some advisors give systematically higher risky shares to all their clients while other advisors give their clients lower risky shares. The data reveal remarkably powerful advisor effects. The adjusted R 2 in Panel A nearly triples from 11.0% to 30.8% as we add the advisor fixed effects. In Panel B s home-bias regression the adjusted R 2 increases from 1.7% to 26.4%! 11 In a model in which labor income correlates with asset returns, the optimal mix of domestic and international equities would vary in the cross-section of investors if there is variation in how investors labor incomes correlate with domestic and international equities. The analysis in section 4.4 addresses the role of any omitted variables such as this correlation. 16

17 Figure 2 plots the distributions of the advisor fixed effects from the regressions presented in column 2 of Panels A and B. The standard deviations of these distributions are 17.1% and 24.6%. These distributions illustrate that, in addition to being statistically very important in explaining cross-sectional variation in portfolio choices, the advisor fixed effects are economically important. To put this result into perspective, for the share of risky assets we observe that a one-standard deviation movement in the advisor distribution is approximately equal to changing an investor s risk tolerance from low to high! It is important to emphasize that the fixed-effect estimates are orthogonal to the investor attributes of column 2. That is, they measure differences in risky share and home bias after accounting for variation originating from differences in age, gender, risk tolerance, and so forth. 4.3 Interpreting advisor fixed effects How should we interpret our finding that advisor fixed effects explain cross-sectional variation in portfolio choices? We can delineate two potential explanations. First, advisors may have idiosyncratic tastes in portfolio allocation. These tastes may reflect advisors personal beliefs for example, equities are relatively safe in the long run and offer a very attractive return-to-risk trade-off or they may arise from agency conflicts some advisors may respond more to financial incentives, recommending higher-commission equity funds over cheaper fixed-income funds. Second, advisor fixed effects may appear to be important because of matching between advisors and investors. If investors match with advisors who share their beliefs and preferences, then advisor fixed effects will capture common variation in portfolio choices due to shared beliefs among clients rather than the advisor s common influence across clients. We test directly for the importance of omitted investor attributes in section 4.4. Before describing that analysis, however, we should first observe that the results in Table 4 cast some doubt on the matching explanation. First, although it is possible that we are missing out on some determinants of optimal portfolio choice we, for example, lack data on the correlation between labor income and market returns we measure and control for a number of important attributes. If some investor attributes are to explain differences in equity allocation, we would expect risk tolerance, financial knowledge, age, and wealth to be at the top of the list. Nevertheless, these variables jointly explain just one-tenth 17

18 of the variation in risk shares and less than 2% of the variation in home bias. We would need to find variables that are important determinants of portfolio choice and that are also substantial drivers of the advisor-investor match. Second, when we include advisor fixed effects, moving from the first regression to the second in Table 4, we estimate similar coefficients on the investor attributes and we estimate those coefficients with markedly more precision. These findings imply little colinearity between investor attributes and advisor fixed effects, which means that if investors and advisors are matched by shared attributes that determine portfolio allocations, these attributes must be largely unrelated to age, gender, risk tolerance, and financial knowledge. If the matching relates, at least in part, to the variables included in the model, then the advisor fixed effects perfect proxies for the shared link would kill the statistical significance of the imperfect empirical proxy such as age or gender. This argument is intuitive if we think of running the regression in two stages. Suppose that we first clean the data by regressing risky share only on advisor fixed effects. Column 2 s estimates show that if we now collect the residuals from such a first-stage regression and run them against investor attributes, many attributes are statistically more significant in the residual data relative to the raw data. That is, the variation in risky shares that emanates from advisor fixed effects is mostly noise when studied from the vantage point of investor attributes. The attenuation of the income and net worth dummies when moving from the first regression to the second offers some limited evidence that these variables may influence both portfolio choices and advisor-investor matching the advisory market appears to be segmented based on client wealth. Overall, however, the slope estimates are remarkably similar between the regressions in columns 1 and 2. Figure 3 illustrates this point by plotting the marginal effects associated with the age and risk-tolerance categories with and without advisor fixed effects. Although these results do not rule out the possibility of important omitted variables that drive both the portfolio choice and the investor-advisor match, they substantially narrow down the set of potential variables that could be at work. Third, the last two regressions in Table 4 show that advisor fixed effects are equally important whether an advisor serves a diverse or an homogeneous group of clients. We divide advisors into high- and low-dispersion groups based on the estimated client-base heterogeneity. We measure heterogeneity each year by recording the predicted values from the first column s regression and 18

19 then computing within-advisor standard deviations of these predicted values. Advisors in the lowdispersion group have homogeneous client bases. By construction their clients have such observable attributes that the first column s model predicts similar portfolio allocations. Advisors in the highdispersion group, by contrast, have more heterogeneous client bases for which the first column s model predicts substantial variation among portfolios. If the increase in adjusted R 2 when we add the advisor fixed effects is due to omitted variables, we would expect advisor fixed effects to play a far smaller role in the sample of high-dispersion advisors. In the data, however, the overall explanatory power of the model is largely insensitive to this grouping; the change in the adjusted R 2 when adding advisor fixed effects is similar for both homogeneous and diverse client groups. 4.4 Selection of new funds with two-way fixed effects The results of Table 4 suggest that advisors have approximately twice as large an effect on the risky share as the set of investor characteristics. In terms of the home bias, the effect is even larger: the explanatory power of the model increases from close to zero to one-quarter when we add advisor fixed effects. Moreover, we find little evidence that advisor fixed effects relflect clientadvisor matching on observable characteristics. The possibility remains, however, that the model omits an important characteristic. To address concerns about such unobservable characteristics, we use a subset of the data to control for unobserved heterogeneity among investors and thereby disentangle investor effects from advisor effects. We estimate panel regressions of the form, y iat = µ i + µ a + µ t + ε iat, (4) in which y iat is investor i s risky share or home bias in year t and µ i, µ a, and µ t represent investor, advisor, and year fixed effects. To identify separate investor and advisor fixed effects, we must observe portfolio choices for investors who use multiple advisors during the sample period. If each investor uses a single advisor during the sample period, such an investor alone cannot be used to identify separately the investor and advisor fixed effects. But if we have a sufficient number of investors who move across advisors, we can use these movers to identify fixed effects in both dimensions. The presence of such movers, however, makes also non-movers useful: we can make 19

20 inferences about a non-mover s fixed effect if that investor is paired with an advisor who is associated with at least one mover. This estimation approach originates in Abowd, Kramarz, and Margolis (1999), who use it to disentangle the firm and employee effects on wages. Graham, Li, and Qiu (2012) bring it to the finance literature to disentangle the roles that firm and manager effects play in executive compensation. 12 We construct the sample for estimating regression (4) by first identifying investors who change advisors at least once during the sample period. We exclude cases in which the investor initiates the switch and instead focus on the subset of switches caused by the initial advisor s disappearance from our sample due to retirement, death, or withdrawal from the advisory business. We infer these disappearances by recording an investor s move from advisor A to advisor B only if advisor A stops advising all of his or her clients within one year of the date of the move. That is, we require that advisor A has to disappear from the data completely. After generating a list of investors who complete at least one move from advisor to another, we create another list of all advisors who are ever associated with these investors. In the final stage of sample construction we collect data also on those non-movers who are, at any point, associated with any of the advisors on this second list. Instead of studying portfolio-level risky share and home bias within this sample as we did in Table 4 we study the average risky share and home bias for the flow of new investments made while paired with the current advisor. The concern with the portfolio-level measures is that the stock of assets changes slowly: when an investor moves from one advisor to another, that advisor may not reset the investor s portfolio overnight. An investor, for example, may be locked in to some investments through back-end loads on redemptions within seven years of purchase. Focusing instead on the flow of new investments allows us to judge more clearly the current advisor s input to the portfolio. After computing the average risky share and home bias of new funds for each investor-advisor pair, we regress those outcomes on investor and advisor fixed effects. The first two columns in Table 5 replicate the regressions from Table 4 using this alternative sample. The coefficient patterns are similar, which reassures us that this subset of investors does not differ from the main sample. The decrease in sample size, of course, reduces the precision of 12 The early research on models with high-dimensional fixed effects, such as Abowd, Kramarz, and Margolis (1999), relied on approximate solutions due to constraints imposed by substantial memory requirements. Modern techniques for solving these models use memory-saving iterative techniques that can be iterated arbitrarily close to the exact solution. These techniques are now pre-packaged for the major statistical softwares. In Stata, for example, felsdvreg and reg2hdfe are available for efficiently solving models with two high-dimensional fixed effects. 20