Asset Management and Affiliated Analysts Forecasts

Asset Management and Affiliated Analysts’ Forecasts

Paul Irvine Goizueta Business School

Emory University Siva Nathan School of Accountancy Robinson College of Business

Georgia State University Paul J. Simko

Darden Graduate School of Business University of Virginia

September 2003

We thank George Benston, Bruce Branson, Larry Brown, Kirsten Ely, Shehzad Mian, Kumar Sivakumar, Greg Waymire, Richard Willis, workshop participants at Emory, Florida, Georgia State, Houston, Wake Forest, and participants at the Southeast Regional Meeting of the American Accounting Association, the tenth annual Financial Economics and Accounting meetings, the Financial Management Association Meeting, and an anonymous reviewer for helpful comments. Ron Harris and Kathryn Epps provided very capable research assistance. We also thank I/B/E/S International, Inc. for use of the earnings forecast data and for releasing the identity of brokerage firms issuing earnings forecasts. This data has been provided as part of a broad academic program to encourage earnings expectations research. The Goizueta Business School provided financial support for the purchase and access of Morningstar OnDisc.

Asset Management and Affiliated Analysts’ Forecasts

Full-service brokerage firms typically engage in investment banking, broker/dealer services, and fundamental research. The activities of brokerage firms’ research departments are ostensibly

independent of other operations, but there is clear evidence that relationships exist between departments. The influence of lucrative corporate finance relationships on analysts’ opinions is

particularly well documented in both academic studies and the business press (Dugar and Nathan, 1995, Lowenstein, 1996, Lin and McNichols, 1998, Laderman, 1998, Michaely and Womack, 1999). Recent lawsuits in Massachusetts (Craig, 2002), and New York (Beck, 2002, Silverman, 2002) support the notion that relationships between sell-side research and other brokerage-firm departments adversely influence research quality. Although the business press recognizes that investment banking is not the only potential source of conflict for sell-side analysts (Sidel and Craig, 2002, Opdyke, 2002),

commentary surrounding the investment banking litigation has been particularly negative: “When internal communications reveal Merrill Lynch research analysts disparaging the very stocks they are recommending to millions of hard-working families seeking guidance on how to invest their life savings and their children’s college fund, we must ask: do their investment recommendations serve the investors to whom they owe a duty of loyalty and honesty or to their investment banking clients.” (Spitzer, 2002).

What is frequently not made clear in this debate is that positive externalities can sometimes arise from bundling different functions within a full- service broker. One example is in- house asset

management, which produces a considerable demand for high-quality financial information within the brokerage- firm. Asset management has been one of the fastest growing areas in financial service, with many brokerage firms participating through the development of mutual fund- based asset management

departments.1 To meet their demand for information, the buy-side can generate information internally, from its’ own buy-side analysts and portfolio managers, or obtain it externally from sell-side analysts. To what extent does the buy-side communicate with their own broker’s affiliated sell- side research departments? Does the potential for pooling information across the research and asset management departments have meaningful implications for sell-side analysts’ forecasts or the actions of their affiliated asset management departments?

Although there is no explicit prohibition of communication between affiliated buy-side and sell-side analysts in the Investment Company Act of 1940, we found that practitioners were divided on whether the buy-side and the sell- side within the same firm shared information. Conversations with buy-side and sell-side analysts at several full-service brokerage firms revealed polarized opinions. Some claim that such communication is rare or non-existent, while others suggest it is a common activity on the Street. To determine if there were any explicit references to intra- firm communication, we examined N-SAR item 20 that lists the top ten soft dollar brokers for each fund. N-SAR filings did not provide a definitive answer. For example, the Dean Witter Mid Cap Growth Fund’s January 1995 N-SAR lists Dean Witter as their primary supplier of soft-dollar services. In contrast, however, the Smith Barney Aggressive Growth Fund April 1995 N-SAR does not include Smith Barney among the fund’s top ten soft-dollar brokers. Our conclusion is that whether intra- firm communication has any important effects is essentially an open empirical question. This study evaluates this issue by examining the association between two variables that capture the information environments of asset management and research: the asset management departments’ level of stock ownership and the earnings forecast accuracy of affiliated sell-side analysts.

1

In contrast to the generally negative commentary surrounding the influence of the corporate finance department on sell-side research, we contend that bundling asset management and research can be benign or even improve the quality of both the sell-side analysts’ forecasts and fund managers’ investments. These benefits can arise in one of two ways. First, fund managers might closely examine stocks where their affiliated sell-side analysts have particular expertise relative to other analysts. This expertise arises from analyst talent, experience, effort, or from differences in resources across

brokerage firms, and provides a compelling reason for asset managers to follow these analysts. A complementary benefit may arise merely because of the asset manager’s investment position in a stock. A manager’s fiduciary and performance incentives impel them to closely monitor stocks that are

particularly important to the performance of their portfolio. Given sell- side analysts incentives to gather information about the stocks they cover, a clear source of information would be their own firm’s asset management department if the department owns a large position in a stock. Affiliated sell-side analysts would also have incentives to increase their focus on securities that comprise a significant share of fund manager’s investments (Sidel and Craig, 2002).

If a brokerage-firm’s sell-side analysts gather high-quality information about stocks owned by their affiliated asset management departments, either directly from their own expertise, or indirectly from buy-side analysts and portfolio managers, then we predict affiliated analysts’ earnings forecasts on these stocks will be more accurate than those of a control group of unaffiliated analysts. Further, we contend that affiliated analyst forecast accuracy, relative to unaffiliated analysts, is increasing in the significance of the fund’s investment.

To test our predictions we study new security purchases by the mutual fund families of 17 full-service brokerage firms over the years 1994-2001. We examine purchases as we posit they would provide an unambiguous positive signal about the information environment within the fund family,

given portfolio managers felt sufficiently confident in these stocks to add them to their fund family’s portfolios. In our primary tests we find that affiliated analysts’ earnings forecasts become significantly more accurate (but not more optimistic), relative to unaffiliated analysts, as the percentage ownership stake in a company increases. Our results in particular show a significant relation between affiliated analyst forecast accuracy and fund family stock ownership in the highest ownership decile, a level that approximates 1 percent of ownership.

We also find a significant positive relation between affiliated analyst accuracy before the new investment and the percentage of the company’ shares outstanding acquired by the broker’s fund family. In particular, fund families make the largest purchases in stocks where their affiliated analysts were the most accurate ex-ante. This complementary result suggests that portfolio managers follow the advice of affiliated analysts’ that they have found to be particularly accurate forecasters.

Taken as a whole, our findings are consistent with prior research showing that the various departments of a full-service brokerage firm do not operate independently, but rather, that there is significant interaction among them. However, our central finding demonstrates that these externalities are not always negative and that synergies can sometimes be obtained by providing multiple services within a single brokerage firm.

Data and Methodology

Our empirical analysis requires data of two types: detailed portfolio data for mutual funds’ stock holdings and affiliated sell-side analysts’ earnings forecasts on these same securities. The objective of our sampling procedure is to identify a comprehensive sample of securities with data available from both sources. We first identify a sample of full-service brokerage firms with both research and mutual-fund-based asset management departments. Information on these firms’ portfolio composition is then

obtained from Morningstar’s OnDisc database and analyst earnings forecast data is obtained from I/B/E/S.

Mutual Fund Family Identification. From the 1994-1995 Nelson’s Directory of Investment Research we identified 310 U.S.-based brokerage firms that provided security research services. We subsequently match these firms to the mutual fund ownership information in the Morningstar Report on Equity Mutual Funds (1991-1995). Thirty- five brokerage firms have fund data available in Morningstar, and seventeen of these firms also have earnings forecasts available on I/B/E/S over the study period. These seventeen firms comprise our sample of mutual fund families owned by full-service brokerages.2

Morningstar collects mutual fund information from the SEC mandated N-SAR form. The semi-annual N-SAR filing includes information on trades, 12b-1 fees, portfolio turnover rates, sales charges and selected financial information. Every six months, in April and October, Morningstar summarizes the mutual fund stock ownership data from the N-SAR in their OnDisc database. The N-SAR filing is used in lieu of other SEC mandated filings because it has the advantage of being the most inclusive and comprehensive filing available. Our sample of mutual fund stock holdings is gathered from the October 1994 through October 2001 OnDisc reporting periods (e.g., 15 total periods).3 We end the Morningstar data collection in 2001 because, as discussed further below, to measure analyst accuracy we must allow significant time for firms to report actual earnings per share. For each OnDisc reporting period we determine the individual funds managed by the seventeen full-service brokers in our sample. For these

2

Over the sample period I/B/E/S stops collection of broker codes after mergers (e.g., Dean Witter or Salomon

Brothers). We follow I/B/E/S when identifying active brokers. When a broker code is dropped by I/B/E/S we retain the fund family in our sample but stop any further Morningstar data collection.

funds we collect the net asset value, identify the individual securities held within each fund, and calculate the market value of each equity security held by the fund. We exclude from our analysis any fund that does not hold domestic equities.

Identification of Security Investments for Each Fund Family. We focus on stock ownership at the fund- family level, as opposed to the individual fund, because this level best aggregates the importance of the ownership position to the asset management department. Furthermore, we use only new security purchases which we expect represent s an unambiguous positive signal from the fund manager. For each security we compute both the net asset value weight of the holding as a percent of all fund holdings (NAVWGHT) and the percentage of stock owned by the fund family (%OWN). More specifically, NAVWGHT for a fund family portfolio is measured, for each semi-annual Morningstar report period (t), as the summed market value (MV) of a security (i) across all k funds (f) for each fund family (F). Scaling by the fund- family’s total net asset value yields the fund family allocation attributable to security i:4 100 Value Asset Net MV = NAVWGHT t t k 1 = f t F, f, k 1 = f t F, f, i, t F, i, ×              

∑

∑

(1a)

%OWN is calculated as the total common shares held by the fund family as of the end of the OnDisc reporting period over the total common shares outstanding:

3

One limitation of the N-SAR is that it provides only semi -annual data for a given fund. Less inclusive but more frequently required is Form 13-F that must be filed quarterly by any institutional investment manager having equity assets under management greater than $100 million. In addition, Form 13-D must be filed within ten days of any transaction that results in the fund beneficially owning more than five percent of the outstanding securities of any particular stock.

100 g Outstandin Shares Owned hares S = OWN % t i, k 1 = f t F, f, i, t F, i, t ×              

∑

(1b)

Ideally we would like to be able to identify the specific date when the investment by a fund actually takes place. Unfortunately, the ability to pinpoint this date is not possible because funds report only semi-annual holdings and funds within a family have different semi-annual report dates. For instance, for a fund whose semi-annual reporting period ends November 30th, an actual trade inferred from comparing the May 31st and November 30th portfolio holdings may have occurred as early as June 1st or as late as November 30th. Because information can flow from the analyst to the portfolio manager or from the portfolio manager to the analyst, we examine earnings forecasts both before and after the dates at which we can unequivocally state that ownership exists (i.e., June 1st and November 30th in the aforementioned example).

Table 1 presents descriptive data on select characteristics of the 17 fund families in our sample. Column (1) identifies the brokerage firm and columns (2) and (3) present the average number of funds and net assets under management by fund family across all Morningstar reporting periods. All fund families are represented in multiple periods but only large families have the requisite data for the full 15 reporting periods. Merrill Lynch is the largest fund family in the sample, with an average 35 billion dollars under management across an average 74 funds. Across all families the relative size and total investments vary considerably. Columns (4) and (5) of Table 1 report the composition of the sample used in the empirical tests. Specifically, we gather analyst forecast data from I/B/E/S for periods both

4

before and after a fund family’s new security purchase. The larger fund families still tend toward having more observations in the final sample, but no single family dominates. One interesting conclusion that can be readily inferred from the data in columns (4) and (5) is sell-side analysts’ tendency to initiate coverage of stocks after purchase by the affiliated portfolio manager. This conclusion is reflected in a 57 percent increase in samp le size when the measurement date of the affiliated forecast is moved from before to after the fund family investment. 5

Estimation of Analysts’ Relative Accuracy. Earnings forecasts and earnings per share (EPS) are gathered from I/B/E/S. To be included in the sample a security must (i) be actively followed by at least three unaffiliated analysts during the quarter immediately before or immediately after the end of the OnDisc reporting period, (ii) have two- year ahead earnings forecasts available for both the affiliated analyst and a control group of unaffiliated analysts, and (iii) have actual EPS data available for each earnings forecast. Each criterion is imposed because our tests rely on measures of relative forecast accuracy. The benchmark against which the level of an affiliated analyst’s earnings forecast is evaluated is a control sample of analysts without direct affiliation with the fund family.

Our first criterion is necessary to provide a reasonable comparison by mitigating the influence of outliers that may arise solely because a security is infrequently followed.6 The second sampling criterion is imposed due to the nature of the presumed investment horizon of mutual fund managers. existing funds.

5

Our predictions should also hold for the divestiture of a security by the asset manager. In this case, we should expect affiliated analysts to be relatively pessimistic. However, empirically we found that when an asset manager significantly decreases his investment in a security, about one-third of the affiliated analysts tend to drop coverage of that security. This is consistent with the results reported in McNichols and O’Brien [1997] who find that when analysts have negative information about a stock they tend to drop coverage of that stock rather than revise their earnings forecast downward or issue sell recommendations. Due to this significant drop in coverage by affiliated analysts, it will be difficult to interpret results pertaining to relative pessimism and relative forecast accuracy. Therefore, this study focuses only on new investments.

Given restrictions on short-term trading activities during our study period it is reasonable to assume that both the mutual fund manager’s private information and his investment horizon likely pertain to a time-period beyond the current fiscal year of the company.7 Thus, we use forecasts of two-year ahead earnings. Our final requirement that actual EPS be available is necessary to compute relative forecast accuracy. Thus, our sampling period is restricted to no later than October 2001.

Although our main focus is on the relation between sell-side analyst accuracy and fund family investments, our initial tests also examine forecast optimism relative to a control group of unaffiliated analysts. McNichols and O’Brien [1997] define earnings forecast relative optimism as the difference between the observed forecast and a relevant benchmark. This study adopts this approach by

specifically measuring affiliated analysts’ relative optimism (OPTIM) as:

t i, UNAFF t i, F, AFF t i, F, t i, F, FORECAST -FORECAST = OPTIM σ (2)

In Equation (2), FORECASTAFF is a two-year-ahead earnings forecast for security ‘i’ with date ‘t’ referencing the earliest date in the quarter after the end of the OnDisc reporting period. Superscript AFF refers to forecasts made by the affiliated analyst of fund family ‘F’ and UNAFF indicates the most recent forecast made by an unaffiliated analyst within ± 30 days of the affiliated analyst forecast.8 We scale the numerator by the standard deviation of all forecasts for security i to control for cross-sectional

6 _{Tightening this constraint so that five forecasts were required did not materially influence results.}

7 _{During most of our study period tax law specifies that no more than 30 percent of a fund’s net income can pertain to}

short -term trading activity if a fund desires to remain tax-exempt. Violation of this requirement renders all current year income taxable during that year.

8

Less that 3 percent of t hese analysts are also those that are “affiliated”, in terms of purchasing new securities, with other brokerage firms at some time during the sample period. Elimination of these observations has no material affect on any inference reported throughout the paper.

variation in the dispersion of analysts’ forecasts.9

Next, we examine relative forecast accuracy, a measure distinct from optimism in that the former captures the quality of forecasts made by the affiliated analyst while the latter reflects any potential bias in the forecast. We define relative accuracy for secur ity i, ACCUR, as the difference in standardized unaffiliated and affiliated analyst forecast errors (forecast less actual EPS) scaled by their respective standard deviations:

AFF i AFF t i, UNAFF i UNAFF t i, t i, F, EPS -FORECAST -EPS -FORECAST = ACCUR σ σ (3)

Table 2 summarizes the pooled distributions of the components of OPTIM and ACCUR partitioned by whether the affiliated analyst forecast was measured before or after the fund purchase. The mean forecasts of affiliated analysts (FORECASTAFF) are generally comparable to the group of unaffiliated forecasts (FORECASTUNAFF), although the former is higher at both the mean and median. Both affiliated and unaffiliated analysts’ forecasts are, on average, approximately 20 percent larger than actual reported earnings, indicating that the previously documented phenomenon of forecast optimism exists in our sample for both affiliated and unaffiliated analysts. The table further summarizes

NAVWGHT, the percentage of stock owned by the fund family (%OWN), the market value of common equity in millions (MV), and analyst following (NUMEST).10 Evident from the distribution of

NAVWGHT is that managers are not committing a significant portion of fund assets to new

investments, as its median is less than one-tenth of one percent. The mean percentage ownership by the

9

For this calculation we use all active unaffiliated analysts’ forecasts released in the 30-day window centered on the affiliated analyst forecast date. If the same analyst makes multiple forecasts during the 30-day window, the earnings forecast made nearest, in absolute days, to the affiliated analyst forecast is kept in the distribution. See Comiskey, Walkling and Weeks [1987] and Ajinkya, Atiase, and Gift [1991] for a discussion of the standardized mean earnings forecasts.

fund family, %OWN, is 0.370 for new fund purchases with a top to bottom decile range of just under one percent of firm value. The average number of analysts following the companies (NUMEST) is 12.8 and the average size of the firm, as measured with market value of equity (MV), is $14.1 billion.

Table 3 provides descriptive evidence on the distribution of relative optimism (OPTIM) and relative accuracy (ACCUR), partitioned by whether the forecast occurred before or after fund purchase. We do not find widespread evidence that the distributions of affiliated forecasts are significantly

different from those of unaffiliated analysts, as both before and after fund purchase relative optimism and relative accuracy have medians of zero. A difference- in-means test reports that relative optimism is significantly higher both before (t=3.74) and after (t=1.86) the new investment, but these results are not confirmed after the investment using a non-parametric sign-rank test. Consistent with the descriptive evidence in Table 2, relative accuracy show a slight deterioration after the fund purchase, with affiliated analysts initiating coverage after the fund’s purchase, on average, forecasting somewhat worse than affiliated analysts that cover the firm prior to the fund’s purchase. As a whole, at this full sample univariate level the only inference that is consistent in both parametric and non-parametric tests at conventional significance levels is that affiliated analysts are more optimistic than unaffiliated analysts before fund purchase. In our tests that follow we explore whether conditioning on the analyst fund families percentage ownership of the stock affects the relative accuracy of affiliated analysts’ forecasts.

Main Results – Ownership Level and Relative Forecast Accuracy

We report our main results in two stages. We first examine the cross-sectional association between the relative accuracy of affiliated analysts’ forecasts and the fund family’s ownership level in

10

MV is price per common share multiplied by total common shares outstanding. Price and share data are gathered from CRSP. For each OnDisc reporting period the descriptive statistics for all variables are similar to those reported in

the stock. We then estimate multivariate regression models that control for other factors that could affect the relation between accuracy and ownership. After establishing these main results we provide a summary of robustness checks.

Relative Forecast Optimism and Accuracy Conditional on Percentage Ownership. Our primary empirical tests examine whether affiliated analysts’ forecasts exhibit a systematic association with the fund family’s percentage ownership of a stock, %OWN. Our preliminary examination of this issue involves the use of portfolios and is presented in Table 4. We form ten portfolios by ranking all observations based on the values of %OWN, then examine the coincident values of affiliated analyst relative accuracy for each of these deciles. While our primary focus is on the quality of affiliated analyst output as reflected by relative accuracy, we also include relative optimism in the table for comparison. To provide a more clear and meaningful measure of relative optimism and accuracy across the ownership deciles, we report the percentage of affiliated forecasts that are above the sample-wide median of the respective metric. Thus, under the null hypothesis that there is no relation between affiliated forecasts and percentage ownership (%OWN), the percentage of relatively optimistic

(%OPTIM) and relatively accurate (%ACCUR) forecasts in any given decile is an expected 50 percent. Note that by construction when affiliated forecasts are more accurate for one portfolio they must be, on average, less accurate for another.

Table 4 presents relative accuracy (Panel A) and relative optimism (Panel B) by ownership decile for forecasts made after the fund purchase. The results reported in Table 4 are quite striking. We find no consistent pattern between relative optimism and percentage ownership, but the percentage of

aggregate.

relative affiliated analyst above the median (i.e., > 50 percent) rises dramatically to 61.5 percent in the tenth decile of security ownership. Using a likelihood ratio test appropria te for percentages, (Greene, 1997, p. 886) the ?2-statistic against the null of no difference in forecast accuracy is 21.50 (p-value < 0.01).While we find that the higher is percentage ownership the more accurate the affiliated analysts forecast, across the entire sample there appears to be little systematic relation between ownership and accuracy; no other significant test statistics are observed across ownership deciles. Figure 1 provides a histogram of the results in Table 4. In the figure it is readily apparent that the top decile of %OWN, where ownership approaches 1 percent of the companies shares outstanding (see Table 2), yields the highest percentage of relatively superior forecasts, with a monotonic increase also observable beginning in decile eight. Taken together, the results reported in Table 4 are consistent with the overall

distribution of optimism and accuracy presented in Table 3, as for most deciles there are no sample wide differences between affiliated and unaffiliated analysts.

Untabulated results report a similar distribution in relative accuracy before the fund purchase, with 60.9 percent of affiliated analysts more accurate than their peers in the largest decile of percentage ownership. Thus, fund managers are making the largest new investments in the stocks where their affiliated analysts appear to have the highest quality information, as measured by accuracy of their earnings forecasts before fund purchases.

Multivariate Regression Analysis. The portfolio-based results presented in Table 4 support our hypothesis that affiliated forecast accuracy is related to percentage ownership. However, the tests in Table 4 do not control for other factors that could impact relative accuracy. To evaluate the robustness of our conclusions, we cond uct two rank-transform regression analyses. The specification of each model depends on whether we are examining forecasts made before or after the new investment. The

regressions differ because we expect that the direction of influence differs across the two periods. Before purchase, asset managers would be more inclined to use information from affiliated sell- side analysts that they believe is relatively more accurate. After the purchase, we expect affiliated sell-side forecasts to benefit from the externalities generated by the assetmanagement ownership position. Thus, the models test for two causal relations between ownership levels and relative accuracy: presented first for the after purchase period where relative accuracy depends on the asset managers’ new investment position, and second for the before purchase period where ownership is dependent on the quality of sell-side analyst information, measured by the relative accuracy before the fund purchase. We view these models as complementary in that each is designed to evaluate the association between fund ownership and sell-side analyst output. The following regressions are estimated (security subscripts omitted):

t ε γ γ γ γ γ γ γ FAMILY ) R(MV ) R(ASSETS ) R(NUMEST ) R(NAVWGHT ) R(%OWN ) R(ACCUR F 16 1 F 5 F t 5 t 4 t 3 t 2 t 1 0 t + + + + + + + =

∑

= + (4) and t 16 1 F F 5 F t 5 t 4 t 3 1 -t 2 1 -t 1 0 t FAMILY ) R(MV ) R(ASSETS ) R(NUMEST ) R(OPTIM ) R(ACCUR ) R(%OWN υ δ δ δ δ δ δ δ + + + + + + + =

∑

= + (5)

where R(X) represents the standardized ranked value of variable X and sixteen fund family fixed effects coefficients (FAMILY) included to control for any potential effects across fund families. We specify Equations (4) and (5) as rank-transforms to largely overcome potential violations of the OLS linearity

assumption that exist in our data.11 The primary coefficients of interest are those on %OWN_t-1 in Equation (4) and ACCUR_t-1 in Equation (5). We expect %OWN_t-1 to be positively related to accuracy ACCUR_t in Equation (4), and lagged accuracy, ACCUR_t-1, to be positively related to subsequent ownership (%OWN_t) in Equation (5).

We expect the coefficient on %OWN_t-1 in Equation (4) will be positive because we believe that the direct and indirect channels of information on which the accuracy of the affiliated analyst’s forecast relies will be stronger the greater the percentage of the stock owned by the fund. An alternative

hypothesis is that the results in Table 4 could be explained if analysts exert greater effort forecasting earnings for stocks that are important to their brokers’ asset management department. To control for this alternative, we include the NAVWGHT, the net new investment’s relative importance to the fund portfolio. We posit that the higher the net asset value weight, the greater incentives for the analyst to forecast accurately, therefore we expect a positive coefficient on NAVWGHT. Since the dependent variable in the regression is the forecast accuracy of the affiliated analyst relative to that of an

unaffiliated analyst following the same company, company specific variables affecting analyst forecast accuracy need not be included in the regression (see Clement, 1999). However, characteristics

explaining cross-sectional differences in a stock’s information environment could be important controls. For this reason we include the number of analysts forecasting earnings, NUMEST, and the market value of the firm, MV. In addition, we control for potential resource differences across fund

11

Rank transformation gives us the ability to correct any non-linearities in the data. Evident from Table 4 and Figure 1 is that in the right tail of %OWN there is a clustering of more relatively accurate forecasts. In addition, very small deflators in the calculations of ACCUR and OPTIM would lead to very extreme observations. The estimation procedure used in this rank transform regression is to replace the continuous variables with their correspondin g ranks. Specifically, when there are ‘n’ observations for a sample variable, a rank of 1 is assigned to a variable’s smallest observation and a rank of ‘n’ to the largest. It is not necessary to take logs before the ranking procedure because another monotonic transformation will not change the assignment of ranks. Average ranks are used in case of ties. Ordinary least squares regression is performed on the ranked data. See Iman and Conover [1979] for further details.

families with the variable ASSETS, which represents the total assets under management.

Equation (5) presents a model where the level of ownership is dependent upon the sell- side analyst. Our expectation here is that the coefficient on ACCUR_t-1, d1, will be positive, because asset

managers are, on the margin, more likely to invest in securities in which the firm’s sell-side analysts have revealed their expertise by producing more accurate forecasts than their competitors. As control variables in Equation (5) we include ranked OPTIM to control for the possibility that fund managers would invest in securities for which the firm’s affiliated analysts have higher expectations tha n other analysts. We also include NUMEST, MV, and ASSETS, in Equation (5) to control for the effects outlined above.

Tables 5 and 6 report the results from estimating restricted and full specifications of Equations (4) and (5), respectively.12 The results in Table 5 reinforce our inferences from the portfolio analysis reported in Table 4. Consistent with the general pattern of increasing accuracy in Table 4, in the restricted regression (1) %OWN is positive and significant (?₁=0.038, t=2.60). In the unrestricted regression (2) these results are even stronger (?₁=0.110, t=4.89). The more of a particular security the affiliated fund family owns, the more accurate the affiliated analyst forecasts subsequent to the purchase. In the restricted regression, the slope coefficient on NAVWGHT is also positive and significant (?₂=0.073, t=5.04), indicating that larger stocks in the fund family are associated with accurate forecasts. However, this result is insignificant in the unrestricted regression. Thus, the positive coefficient on NAVWGHT in the restricted regression appears to be explained by the fact that larger,

12 _{Coefficients for the intercept and fund family fixed effects are not tabulated for clarity. We estimate restricted and}

unrestricted regressions to ensure that the significant effects we document are not caused by a particular regression specification. For each regression we also estimate variance inflation factors to evaluate the influence of

multicollinearity. While multicollinearity does not bias reported coefficients, its presence will potentially result in an incorrect inference of insignificance. No undue influence of collinearity across regressors was noted.

heavily- covered firms, which also tend to have a high NAVWGHT, tend to have higher relative forecast accuracy. Also, all else equal, larger fund families tend to be associated with more accurate affiliated analysts.

In Table 6, we document a significant positive relation between ownership and the relative accuracy of affiliated analysts before the fund purchase. The coefficient on relative accuracy is

significantly positive in both the restricted (d1=0.195, t=15.63) and unrestricted regressions (d1=0.103,

t=8.49). This result is notably strong and potentially indicative of more accurate forecasts (i.e., analyst expertise) influencing fund mana gers decisions to purchase larger positions. Unlike accuracy, the results for optimism are not robust to the inclusion of the control variables. Although significant in the restricted regression (1), optimism is insignificantly positive in the unrestricted regression (2). Not surprisingly, because %OWN represents funds committed to a particular stock, this variable is

positively related to the size of the fund family (ASSETS), and negatively related to market value and the number of analysts covering the stock.13

Taken as a whole, the results in Tables 4, 5 and 6 support our contention that a positive externality exists between fund managers’ investment decisions and affiliated sell-side analysts’ forecasts. Fund family percentage ownership, which measures incentives driving the buy-side’s

information collection, is positively related to the relative accuracy of their affiliated analyst’s forecasts, both before and after the initial fund investment. The relation appears much stronger, however, when ownership is measured subsequent to the forecast. We infer, from the regressions results in Tables 5 and 6, that the greater the ownership in the stock and the relative accuracy of the affiliated analyst, the greater the overall information quality within the brokerage firm.

Robustness Checks. A number of untabulated robustness checks were performed to ensure reported results were not an artifact of our research design. Our tests focus on whether relative forecast accuracy improves as ownership increases. We examined whether this result holds for forecast accuracy, without netting out the control group of unaffiliated analysts. Replication of Equation (4) using analyst absolute forecast errors as the dependent variable yields results consistent with those reported in Table 5. Other alternative specifications were performed which include the following: (i) computation of OPTIM and ACCUR with the consensus forecast as an alternative deflator, (ii) use of the consensus forecast as our control rather than a single unaffiliated forecast, and (iii) inclusion of reporting period intercepts in equations (4) and (5). Each alternative research design produced inferences that were qualitatively the same as those reported above.

Finally, we tested whether reputation and investment banking relationships affected our results in an analysis of the 1995-96 subperiod. Specifically, Stickel (1992) shows that analysts who are members of the Institutional Investor All- American Research Team (IIAA) supply more accurate forecasts than other analysts who follow the same company. We included, as a control in the regression, a variable if an affiliated analyst had All-American research status during the 12- month window

surrounding the forecast date. The slope coefficients on IIAA were positive, consistent with Stickel (1992), but our results were not statistically significant. In addition, the papers on the influence of investment banking on analyst activity that are cited in the introduction suggest that underwriting activity could bias analyst’s forecasts and result in poorer accuracy. We controlled for investment banking activity by including a dummy variable in the regression if the brokerage-firm engaged in any

13

We also estimated Equation (5) using NAVWGHT as the dependent variable because this variable might capture the importance of the stock to the fund family better than %OWN. However, this model produces similar conclusions to

investment banking activity, including underwriting stock issues and mergers and acquisition advisory services on the stock in question in the year prior to forecast issue. Overall, only 7% of the forecasts were identified with investment banking activity in the SDC database. The investment banking dummy variable was not significant in any regression specification. Because reputation and investment banking relationships were insignificant in our estimation of Equation (4) in this subperiod, coupled with the costs of data collection, we did not expand this robustness test to the full sample.

Conclusion

This study examines the relation between holdings of mutual funds operated by full-service brokerage firms and the relative earnings forecast accuracy of affiliated sell-side analysts. The purpose is to provide evidence on the interaction between the asset management and research department of a full-service brokerage firm. We measure information quality using relative earnings forecast accuracy, defined as the difference in absolute earnings forecast errors between affiliated analysts and a control group of unaffiliated analysts. Relative accuracy thus captures expertise differences across analysts arising from such sources as analyst talent, experience, effort, or from differences in resources across brokerage firms. Using this measure, we test whether: (i) portfolio managers tend to follow affiliated analysts’ who are relatively more accurate, and (ii) whether affiliated analyst forecast accuracy improves in the percentage ownership of a security after the fund family’s investment.

We find that a fund families largest new investments are significantly associated with affiliated analysts’ forecast accuracy before the fund purchase decision, with the relation particularly strong when ownership is measured subsequent to the forecast. This suggests that manager’s can identify affiliated

those reported in the paper, so these results are not tabulated.

analysts’ that have demonstrated expertise in a particular stock and that, to some degree; they base their new investment decisions on their beliefs in these analysts’ expertise. After the fund family’s purchase of a new stock, affiliated analyst coverage increases and affiliated analyst forecast accuracy is

significantly positively related to the percentage of the stock owned by the fund family.

Our findings contribute to ongoing research regarding the properties of analysts’ earnings forecasts. We document that affiliated analysts’ earnings forecasts vary from others in a predictable fashion and these differences are related to the actions of affiliated asset management departments. However, in contrast to the research emphasizing the negative influence of underwriting on analysts’ forecasts, our results suggest that positive externalities can exist between the different departments of a full-service brokerage firm.

The recent introduction of Regulation FD could have significant impact on our results. If Regulation FD effectively curtails private communications between company management and analysts, affiliated sell-side analysts might lose any informational advantage they obtain from communications with affiliated buy-side analysts or from access to management that accrues to

institutions who own large positions. On the other hand, since private communication with all, affiliated and unaffiliated, sell- side analysts could be restricted by Regulation FD, the presence of a second team of analysts examining the same stock could continue to help affiliated sell-side analysts to provide more accurate forecasts. Preliminary analysis based on our post Regulation FD (August 2000) sample

suggests both affects are present. The relation between affiliated forecast accuracy before the purchase and the size of the fund purchase is less positive than the remaining sample, suggesting that portfolio managers could be relying less on affiliated analysts post Regulation FD. On the other hand, the relation between ownership and accuracy after the fund purchase is more positive post Regulation FD, which suggests that affiliated analysts and portfolio managers may be effectively sharing more

information now that both parties access to management forecasts has been restricted. We suggest that the long run impact of Regulation FD on the relationships we have documented in the paper is an interesting topic for future research.

References

Ajinkya, B., R. Atiase, and M. Gift, “Volume of Trading and the Dispersion in Financial Analysts’ Earnings Forecasts.” The Accounting Review. 1991, Vol. 66, No.2: 389-401.

Beck R. Changes are only one step towards analyst objectivity. Associated Press, September 5, 2002. Clement, M. “Analyst Forecast Accuracy: Do Ability, Resources and Portfolio Complexity Matter?”

Journal of Accounting and Economics. 1999, Vo. 27, No. 3: 285-303.

Comiskey, E., R. Walkling, and M. Weeks, “Dispersion of Expectations and Trading Volume,” Journal

of Business Finance and Accounting, 1987, Vol. 14: 229-237.

Craig S. Massachusetts claims CSFB stock reports led investors astray. Wall Street Journal, October 22, 2002, C1.

Dugar, A., and S. Nathan, “The Effect of Investment Banking Relationships on Financial Analysts’ Earnings Forecasts and Investment Recommendations,” Contemporary Accounting Research, 1995, Vol. 12,1: 131-160.

El-Gazzar, S., “Predisclosure Information and Institutional Ownership: A Cross-Sectional Examination of Market Revaluations During Earnings Announcement Periods,” The Accounting Review, 1998, Vol. 73, 119-129.

Greene, W. Econometric Ana lysis. 3rd Edition, 1997. Prentice Hall, Upper Saddle River, New Jersey. Iman, R., and W. Conover, “The Use of the Rank Transform Regression” Technometrics, 1979, Vol. 21,

No.4: 499-509.

Laderman, J., “Wall Street’s Spin Game: Stock Analysts Often Have a Hidden Agenda,” Business Week, October 5, 1998, pp. 148-156.

Lin, H., and M. McNichols, “Underwriting Relationships, Analysts’ Earnings Forecasts and Investment Recommendations,” Journal of Accounting and Economics, 1998 Vol. 25, No.1: 101-127 Lowenstein, R., “Today’s Analyst Often Wears Two Hats,” The Wall Street Journal, August 30, 1996:

C1.

Nelson’s Directory of Investment Research, Nelson Publications, Inc., Port Chester, N.Y., 1994. McNichols, M., and P. O’Brien, “Self-Selection and Analyst Coverage,” Journal of Accounting

Michaely, R., and K.Womack, “Conflict of Interest and the Credibility of Underwriter Analyst Recommendations,” Review of Financial Studies, 1999, Vol. 12 No. 4, 653-686. Opdyke, J. “Stock Ad vice You Can Trust?” Wall Street Journal, October 31, 2002, D1.

Sidel R. and S. Craig. “Does Independent Research Mean Better Stock Picks?” Wall Street Journal,

October 31, 2002, C1.

Silverman, G. “Investors Positive About Spitzer’s Citigroup action.” Financial Times, October 2, 2002. Spitzer, E. “The Crisis of Accountability.” Law Day Remarks, Office of the New York State Attorney

General, May 1, 2002.

Stickel, S.,"Reputation and Performance Among Security Analysts,” Journal of Finance, 1992, Vol. 47, No. 5, 1817-1836.

Table 1

Fund Family Characteristics and Sample Composition Morningstar OnDisc Semi-Annual Reporting Periods: October 1994 through October 2001

Final Sample: New Investments with I/B/E/S Data Available

FUND FAMILY

Average No. of funds

Average NAV (in $ mill.)

Before Fund Purchase # (% of Securities Held)

After Fund Purchase # (% of Securities Held) BLAIR WILLIAM 3.6 739 85 (5) 147 (9) BURNHAM GROUP 2.4 145 19 (3) 23 (3) DEAN WITTER 34.4 28,853 148 (2) 189 (2) GOLDMAN SACHS 17.2 4,971 566 (7) 992 (12) KEMPER 42.4 14,469 79 (1) 97 (1) LEGG MASON 8.3 8,727 44 (1) 74 (1) MERRILL LYNCH 73.5 35,152 994 (4) 1,565 (7) MONTGOMERY 14.1 2,724 268 (6) 444 (10) MORGAN KEEGAN 1.0 56 43 (6) 56 (8) MORGAN STANLEY 52.6 29,247 856 (4) 1,459 (7) OPPENHEIMER 49.3 32,589 494 (3) 807 (5) PAINEWEBBER 23.2 3,294 503 (3) 680 (5) PIPER JAFFRAY 6.3 760 47 (2) 66 (3) PRUDENTIA L 46.6 14,456 710 (4) 1,042 (6) ROBERTSON STEPHENS 10.2 2,012 223 (8) 311 (12) SALOMON BROTHERS 7.9 1,037 97 (4) 126 (5) SMITH BARNEY 38.8 14,698 727 (6) 1,189 (10) TOTAL 5,903 9,267

Table 2

Distribution Characteristics for Components of Relative Earnings Forecast Optimism and Accuracy, the Weight of a Security Purchase to Fund Family NAV, Ownership Percentage of Firm by the Mutual

Fund Family, Analyst Following, and Security Size

Variable Mean Std. Dev. 90th Decile Median 10th Decile Before Fund Purchase (n=5,903):

FORECASTAFF 1.650 1.215 3.200 1.480 0.370

FORECASTUAF 1.630 1.206 3.169 1.470 0.370

σ 0.144 0.208 0.330 0.077 0.018

After Fund Purchase (n=9,267):

FORECASTAFF 1.642 1.199 3.800 1.520 0.350 FORECASTUAF 1.639 1.190 3.800 1.520 0.355 σ 0.143 0.205 0.322 0.077 0.019 NAVWGHT (%) 0.171 0.464 0.415 0.038 0.001 NUMEST 12.8 8.6 24 11 4 EPS 1.237 1.336 2.960 1.190 -0.220 %OWN 0.370 1.108 0.905 0.088 0.003 MV ($000’s) 14,134.0 32,698.9 33,063.3 3,830.0 526.2

Table presents distributions for variables partitioned by measurement period. Variables measured before fund purchase relate I/B/E/S analyst forecast data gathered during the quarter preceding the fund family’s N-SAR reporting date. Variables measured after fund purchase are for those measured as of the end of the reporting period, or, in the case of forecast data, during the quarter immediately after the reporting period.

FORECASTAFF = the fund family analyst’s two -year-ahead earnings forecast as defined by I/B/E/S,

FORECASTUAF = the most recent two-year-ahead earnings forecast made by an unaffiliated analyst but within ± 30 calendar days of FORECASTAFF,

σ = the standard deviation of all forecasts for security i released within ± 45 calendar days of FORECASTAFF,

NAVWGHT = the weight in security i for fund family F, as of the end of the six month fund family reporting period, NUMEST = the number of analysts following security i,

EPS = the actual two-year-ahead earnings per share as reported by I/B/E/S,

%OWN = the percentage owned of a given security by a fund family, computed as common shares owned over common shares outstanding, and

Table 3

Affiliated Analyst Relative Earnings Forecast Optimism and Accuracy Measured Before and After Fund Family Security Purchase

Variable N % Pos. % Neg. Median

Pr (Sign

Rank) Mean Std. Dev. t-stat. Before Fund Purchase:

OPTIM 5,903 46.8 44.5 0.000 0.021 a 0.079 1.622 3.74 ACCUR 5,903 46.7 44.5 0.000 0.600 a -0.028 1.500 -1.41

After Fund Purchase:

OPTIM 9,267 45.0 44.7 0.000 0.651 a 0.030 1.553 1.86 ACCUR 9,267 44.4 45.1 0.000 0.072b -0.039 1.449 -2.55

Table presents affiliated analyst forecast relative optimism (OPTIM) and relative forecast accuracy (ACCUR) based on affiliated analyst forecasts measured before and after a security purchase. OPTIM is measured as FORECASTAFF less FORECASTUAF, scaled by σ. ACCUR is the difference between the absolute standardized unaffiliated analyst forecast error and the absolute standardized unaffiliated analyst forecast error.

a

denotes a positive sign rank for the distribution

b

Table 4

Percentage of Observations Above the Median Level of Relative Forecast Optimism and Accuracy for each of Ten Percentage Ownership (%OWN) Deciles: Each Measured After Fund Family Security Purchase

%OWNt Decile 1 2 3 4 5 6 7 8 9 10

Panel A: Relative Accuracy

% ACCURt above median 49.8 51.5 52.2 50.3 48.3 49.9 49.7 52.4 52.8 61.5

# Observations 930 930 915 930 930 929 920 930 928 925

Likelihood ratio ?2: 0.004 0.37 0.79 0.01 0.45 0.001 0.01 0.90 1.31 21.50

p-value (.95) (.54) (.37) (.91) (.50) (.98) (.91) (.34) (.25) (.00)

Panel B: Relative Optimism

% OPTIMt above median 50.0 49.6 50.3 51.0 51.5 49.7 51.4 48.6 49.0 50.2

# Observations 930 930 915 930 930 929 920 930 928 925

Likelihood ratio ?2: 0.001 0.03 0.01 0.17 0.37 0.02 0.32 0.32 0.15 0.01

p-value (.98) (.86) (.91) (.68) (.54) (.90) (.57) (.57) (.70) (.93)

Table presents the percentage of observations that fall above the median level of relative forecast optimism (OPTIMt) and accuracy

(ACCURt), for deciles formed on ranked percentage ownership (%OWNt). The expectation for this statistic would be 50% in the absence

of an association between ownership level and the respective relative forecast measure. The number of observations in each cell, and the ?2 statistic that tests each percentage against the population mean of 50% is also reported.

Table 5

Rank Regression of Affiliated Analysts’ Relative Forecast Accuracy After Fund Family Purchase on Percentage of Security Owned by the Fund Family, Change in Security Weight within the Fund

Family, Analyst Following, Firm Size and Fund Family Size

R(%OWN t) R(NAVWGHT t) R(NUMEST t) R(ASSETS t) R(MV t) F-Stat

(1) 0.038 (2.60) 0.073 (5.04) 11.63 (2) 0.110 (4.89) -0.041 (-1.64) 0.042 (3.17) 0.008 (4.16) 0.092 (4.77) 15.15

R(X) represents the standardized ranked value of variable X, where the variable is ranked in ascending order by reporting period with 1 assigned to the lowest observation and “n” to the next highest observation (Iman and Conover (1979)). ASSETS is the total net asset value under control by fund family F. See Tables 2 and 3 for definitions of other variables. The total number of observations is 9,267.

Table 6

Rank Regression of Percentage of Security Owned by the Fund Family on Affiliated Analysts’ Relative Forecast Accuracy and Relative Optimism Before Fund Purchase, Analyst Following, Firm

Size and Fund Family Size

R(ACCURt-1) R(OPTIM t-1) R(NUMEST t) R(ASSETS t) R(MV t) F-Stat

(1) 0.195 (15.63) 0.003 (2.83) 48.84 (2) 0.103 (8.49) 0.002 (1.56) -0.004 (-3.05) 0.593 (32.77) -0.132 (-9.28) 105.19

Figure 1

Affiliated Analyst Forecast Optimism and Accuracy by Ranked Percentage Ownership Analysts Forecasts After Fund Purchase

0.40 0.45 0.50 0.55 0.60 0.65 1 2 3 4 5 6 7 8 9 10

Portfolios of Ranked Percentage Ownership 1=low, 10=high

Relative to Median Optimism or Accuracy