Australian School of Business School of Accounting. Semester 1, Idiosyncratic return volatility, earnings quality, and firm age.

Transcription

1 Australian School of Business School of Accounting School of Accounting Seminar Series Semester 1, 2013 Idiosyncratic return volatility, earnings quality, and firm age Brian Rountree Rice University Date: Friday 22 nd March 2013 Time: 3.00pm 4.30pm Venue: ASB 220

2 Idiosyncratic Return Volatility, Earnings Quality, and Firm Age Andrew B. Jackson Australian School of Business, University of New South Wales Brian R. Rountree* Jones Graduate School of Business, Rice University James P. Weston Jones Graduate School of Business, Rice University Very Preliminary and Incomplete Please do not cite without permission February 2013 ABSTRACT This study links the trend in two earnings quality metrics: 1) the standard deviation of residuals from the Dechow and Dichev (2002) model, and 2) squared residuals from the Jones (1991) model to firm age (defined using data about firm incorporation/founding dates). The results reveal the significant increasing trends in both measures during the period are related to the dramatic decrease in firm age coupled with the increasing percentage of market capitalization that young firms had during the period. Given young firms are expected to have greater variation in accruals, the temporal trends in the measures are not indicative of worse earnings quality over time. We further illustrate after accounting for firm age, there is no association between earnings quality measures and the trend in idiosyncratic return volatility during the sample period. Overall, researchers need to carefully consider the confounding effects of firm age on measures of earnings quality when conducting empirical tests. *Contact author details:

3 1. Introduction Campbell et al. (2001), Xu and Malkiel (2003), Fama and French (2004), Wei and Zhang (2006), Jin and Myers (2006), and Fink et al. (2010) all document increases in firm-specific risk over the period spanning the early 1960s to the end of the 1990s (measures of firm-specific risk used in these studies include return volatility, earnings volatility, lower profitability and survival rates). 1 Rajgopal and Venkatachalam (2011) (hereafter RV) investigates whether deterioration in earnings quality is associated with the increase in firm-specific risk after controlling for numerous other factors influencing risk. They find a statistically significant positive relation between idiosyncratic return volatility and earnings quality indicating that worse earnings quality over time is associated with the increase in idiosyncratic return volatility. This suggests that accounting has performed worse over time and could be at least partially responsible for the general decline in the value relevance of earnings over time as documented in Collins et al.(1997) and Francis and Schipper (1999). The results in RV are consistent with a general deterioration in the usefulness of financial statements over time supporting claims by both academics and practitioners that the shift in the US economy to intangible/service oriented firms has made traditional accounting suspect. It is therefore of paramount importance to understand the nature of the decreasing trend in earnings quality in order to inform the debate on accounting usefulness, as well as provide academics with insights into the time-varying behavior of earnings quality measures. We replicate the graph in RV concerning the decreasing nature of earnings quality metrics (residual standard deviation from a modified Dechow and Dichev (2002) model (DD), and squared discretionary accruals from the Jones (1991) model (ABACC 2 ). Just like in RV, we find a clear increasing trend in both the DD and ABACC 2 metrics that closely mirrors the trend in idiosyncratic return volatility, which was near its peak at the end of the RV sample period, We extend the 1 We use the terms firm-specific risk, idiosyncratic risk, idiosyncratic return volatility interchangeably throughout the paper.

4 graph beyond the RV sample period and find the volatility of residuals from the Dechow and Dichev (2002) model essentially doubles between 2001 and 2005, whereas ABACC 2 follows the decreasing trend in idiosyncratic return volatility beginning in The divergence of the two earnings quality measures is dramatic and unprecedented over the sample period Furthermore, an extension of just 5 years to the RV sample ( ) reverses the signs between DD and idiosyncratic return volatility in the RV panel data regression, whereas the ABACC 2 result holds. In an effort to understand both the trend in earnings quality over time and the diverging nature of the earnings quality metrics, we appeal to findings in Brown and Kapadia (2007), Cao, Simin, and Zhao (2008), and Fink et al. (2010) which help to explain the increase in idiosyncratic return volatility. Brown and Kapadia (2007) illustrate a massive increase in firms undertaking IPOs during the 1990s and these firms are successively more risky than IPO cohorts in earlier periods thus helping to explain part of the increase in volatility. Cao, Simin, and Zhao (2008) find that there is no trend in idiosyncratic risk after controlling for growth options. Fink et al. (2010) expand on both of these papers by collecting information on actual founding/incorporation dates of a large sample of firms from With this data they show a massive decrease in firm age especially during the internet boom of the late 1990s. They further illustrate that the trend in idiosyncratic volatility is completely eliminated once firm age is controlled for. All of these findings indicate that younger, riskier firms with greater uncertainty regarding their growth options entered the sample at extremely high rates in the mid to late 1990s, which coincides with the greatest increase in idiosyncratic return volatility. These same firms are likely the ones with the greatest amount of accrual volatility simply because of the uncertain nature of their operations. We investigate the changing nature of the composition of firms in the sample over time to see if this helps to explain the decrease in earnings

5 quality documented in RV. Using the age data from Fink et al. (2010), we split the time-series of average earnings quality metrics across the youngest and oldest 500 firms each year in the sample. 2 For the DD measure, the youngest firms show a steady increasing trend over time followed by the dramatic rise in 2002 that peaks in The oldest firms have a slight upward trend through 2001 followed again by a relatively large increase in the period. The ABACC 2 measure is slightly different in that the youngest firms tend to follow the general sample pattern, but almost always above the sample mean, whereas the oldest firms do not exhibit a noticeable time trend at all. It is interesting to note that the two earnings quality measures have differing time-series patterns based on firm age. For young firms, the DD measure peaks in the early 2000s, while for the same firms ABACC 2 peaks in the mid-1980s and has been on a general downward trend since then returning to levels similar to the beginning of the sample period by Older firms on the other hand display no trend with ABACC 2, a small positive trend with DD up to 2001 then a large increase. The increase for old firms within the DD measure seems to be influenced by the dramatic rise in the measure for young firms. When we re-estimate the DD measure using only the 500 oldest firms each year the large increase in the period is completely eliminated, as well as shrinking the general upward trend to be modest at best. The evidence indicates that firm age plays a major role in the time-series averages of earnings quality metrics with the youngest firms having a greater influence in the determination of the sample averages. We further explore the divergence between idiosyncratic return volatility and the DD measure beginning in the 2002 period by examining the importance of young firms in the market place. Consistent with Fink et al. (2010), we note young firms (designated as operating for 20 years or less) display a massive increase in the percentage they represent of total market 2 Fink et al. (2010) has firm age data from 1926 through 2006.

6 capitalization of CRSP beginning in the 1990s, peaking in 2000, and then steadily decreasing thereafter. In other words, the period surrounding the internet bubble not only saw the greatest rise in firm IPOs in documented history, but it was accompanied by the highest cumulative market capitalization of young firms as well. Thus it is not just that a high number of firms entered the market, they also played a significant role accounting for as much as 31% of the total market capitalization of all stocks on CRSP. Although the DD measure requires 7 consecutive years to estimate (the model is estimated annually by industry requiring cash flows from t-1 to t+1 and there needs to be 5 residuals to calculate the standard deviation thus data needs to be available from t-6 to t+1), any firm with enough data in a given year is used to estimate the annual cross-sectional regression. Therefore, while the eventual earnings quality metric requires a reasonably long time-series of data to be calculated, changes in the Compustat sample over time can influence the parameter estimates and thus the residuals. Given the increasing role young firms played in the late 1990s it is not surprising that greater variation in residuals is observed in the DD metric even though many of these same firms do not have calculable DD measures because of the lack of enough time-series observations. However, the firms entering the sample during the internet bubble do eventually enter the sample in the early to mid-2000s when they have enough time-series data. Thus we observe a sort of delayed reaction to the internet bubble in the DD measure as opposed to idiosyncratic return volatility that does not require an extensive time-series to calculate. Next, we re-examine the association between earnings quality and idiosyncratic risk adopting the aggregate time-series approach in Fink et al. (2010) and Campbell et al. (2001). We construct an annual summary measure of firm-specific return volatility and regress it on a time trend and summary measures of firm age and earnings quality. A time-series association with idiosyncratic risk is evident if the variable reduces or eliminates significance of the time trend

7 variable. The results reveal that the earnings quality metrics are generally associated with idiosyncratic risk prior to controlling for firm age (the lone exception is the DD measure for the period). However, once age is accounted for neither earnings quality variable is associated with idiosyncratic risk. Overall, the evidence in this study sheds light on the nature of the time-series variation in earnings quality metrics and the influence of firm age on the behavior. The results show that young firms are primarily responsible for the time variation in the earnings quality metrics. These same firms are expected to have the highest degree of variation in the metrics thus it is premature to conclude that accounting has necessarily performed worse over time using these metrics. Our results further show that annual cross-sectional regressions by industry of the Dechow and Dichev (2002) model are particularly influenced by the increase in young firms in the sample even influencing the estimation of earnings quality for the oldest firms in the sample. Controlling for innate accruals quality as in Francis et al. (2006) and using the discretionary piece which by design has no time trend would help avoid the confounding effects of firm age. Regardless, researchers need to carefully consider how to appropriately control for the expected variation in accrual activity based on firm age in their research designs, otherwise results are difficult to interpret. 2. Sample, Data, and Methodology The sample from this study focuses on the period The beginning date of the sample mirrors Rajgopal and Venkatachalam (2011), which begins in 1962 primarily because of the need to gather the necessary data to calculate the Dechow and Dichev (2002) earnings quality measure. The ending date of our sample is 2006 because that is the last year for which we have the necessary information to accurately calculate firm age. To construct this variable, we use data collected by Jovanic and Rousseau (2001) on date of incorporation/original founding for firms

8 between In addition we use founding dates for IPOs from Loughran and Ritter (2004). Finally, we supplement these data with additional incorporation/founding dates from Fink et al. (2010), which examines Mergent Manuals for dates. Fink et al. (2010) use the same data and report that the sample of firms with age data represents over 94% of the total market capitalization of the NYSE, AMEX, and NASDAQ firms, with the majority of firms missing data being small over-thecounter (OTC) stocks. The earliest available date is used to determine firm age. Consistent with Fink et al. (2010), we note that this measure is often quite different than traditional mechanisms for measuring firm age such as using the first date listed on CRSP because of a dramatic drop in the typical age of a firm at IPO over the last 40 years. A prominent example of the bias induced by using CRSP to calculate age are the firms included in CRSP as a result of the creation of the NASDAQ. Many of these firms existed for many years, but CRSP only includes data on them since the early 1970s when NASDAQ was founded. We measure idiosyncratic stock return volatility over time using the method in Campbell et al. (2001) that does not require calculations of firm specific betas from a particular asset pricing model. Instead, it estimates aggregate firm-specific risk of the "typical" firm by averaging deviations from market returns first within industries and then over industries using market value of equity to weight the measures. The measure provides a convenient way to examine the time-series behavior of idiosyncratic risk over a large series of firms placing the fewest restrictions on the data. Figure 1 documents the changing nature of idiosyncratic return volatility and the influence of young firms in terms of market capitalization over time. Both plots replicate findings in Fink et al. (2010) and illustrate that idiosyncratic return volatility has indeed been increasing over the sample period, but much of the increase is focused around the late 1990s, a period which is often referred to as the internet bubble. Using monthly data from Fink et al. (2010) notes that

9 there has been a relatively minor upward slope in idiosyncratic return volatility absent the large spike in the late 1990s and early 2000s. Also, noteworthy is the steady decrease in idiosyncratic return volatility after the collapse of the internet bubble. Trends in mean and median firm age are displayed in Table 1 over the nine five-year subperiods in our sample. Median (mean) age drops from a high of 55 (57) in to a low of 18 (32) in , the period often characterized as the internet bubble. Clearly, firm age has shifted dramatically over time. Accompanying the plot of idiosyncratic return volatility in Figure 1 is the percentage of total CRSP market capitalization of firms aged 20 years or less (young firms). 3 There is a clear upward trend over most of the sample period culminating in a similar spike to idiosyncratic return volatility surrounding the internet bubble and then subsequently decreasing until the end of the sample period. Clearly, young firms are playing a more dramatic role in the market place over time and had an unprecedented presence during the internet bubble. This is particularly important in considering trends in any accounting quality measures. Specifically, young firms are expected to have the most volatile earnings environment and greatest fluctuations in accrual activity. Given, these firms are also playing an increasing role in the market over much of the sample period, we predict that earnings quality measures will also show an increasing trend over time. However, an important qualification is that this relates to the uncertain operating environments of these firms rather than necessarily poor accounting quality. Pastor and Veronesi (2006) argue that the increase in volatility during the internet bubble was driven by firm-specific uncertainty concerning their future profitability, which is a feature picked up by earnings quality metrics. We also calculate firm-specific volatility using residuals from the Fama-French 3 factor model estimated for each firm using daily data (FF_Volatility). Specifically, we estimate firmspecific Fama-French 3 factor model regressions using daily data for each fiscal year. We then take 3 Inferences are unaltered using firms less than 10 or 5 years as the definition of young.

10 the standard deviation of daily residuals for each month of the fiscal year scaling the monthly standard deviation by the number of trading days in the month and then finally averaging across the monthly standard deviations to obtain a fiscal year measure of idiosyncratic volatility (this follows the procedure in RV). We follow RV and calculate two earnings quality metrics. The first one is the following modified version of Dechow and Dichev (2002) model (modifications were suggested by Francis et al. (2005) and McNichols (2002)): TCA it = ϕ 0 + ϕ1cfoit 1 + ϕ2cfoit + ϕ3cfoit ϕ4 REVit + ϕ5ppeit + ν it (1) where: TCA = ΔCA - ΔCL - ΔCash + ΔSTDEBT 4 CFO = IBEX - TCA + DEPN 5 ΔREV PPE = Change in Sales (Compustat SALES/TURNOVER) = Gross property, plant, and equipment (Compustat PPEGT) All variables in equation (1) are scaled by average total assets (Compustat AT) including the intercept. All variables are winsorized at the upper and lower 1% level for the entire sample period (i.e., winsorization is not performed by year or by industry). The model is estimated annually by the Fama-French 49 industries (we supplement the 48 industries with a miscellaneous industry capturing all other firms) and then the firm-specific residuals over the period t to t-4 are used to calculate the standard deviation measure used as the DD metric of earnings quality (firms must have residuals in all 5 periods to calculate the standard deviation). It is important to note that all firms satisfying only the requirements for equation 1 from Compustat are used to estimate the 4 CA is current assets (Compustat ACT), CL is current liabilities (Compustat LCT), Cash is cash and equivalents (Compustat CHE), STDEBT is short term debt (Compustat DLC), and Δ represents the annual change in the associated measure. 5 IBEX is net income before extraordinary items (Compustat IB) and DEPN is depreciation and amortization expense (Compustat DP).

11 industry/year regressions even if some firms do not meet the 5 year criteria for calculating the standard deviation of the residuals from the model. Also, there is a notable difference in the calculations and the number of observations resulting from using Compustat only data versus using the Compustat data from the merged CRSP/Compustat database. In order to maximize the sample size, we use the Compustat only data to calculate the DD metric. We also estimate Jones (1991) model discretionary accruals using the following regression: TA it = 0 + δ1 REVit ARit ) + δ 2PPEit + δ 3ROAi1 δ ( + η (2) it where: TA ΔAR ROA = TCA DEPN = Change in Accounts Receivable (Compustat RECT) = IBEX Average Total Assets All other variables are as previously defined. ROA is included in the model as a mechanism for controlling for firm performance, which Kothari et al. (2005) shows is an important determinant of abnormal accruals. Consistent with RV, we use the squared estimate of η it as another measure of earnings quality (ABACC 2 ). Finally, we winsorize both DD and ABACC 2 (not FF_Volatility) at the upper and lower 1% levels in order to reduce the influence of extreme observations on the data 6. Table 1 displays means and medians over five year sub-periods covering the entire 45 years of the sample with a total of 104,448 observations. Both measures of volatility (FF_Volatility and Volatility) show a clear upward trend through 2001 followed by a significant drop in the sub-period for both mean and median. The earnings quality metrics DD and ABACC 2 display similar patterns to the idiosyncratic return volatility measures through 2001, but DD continues to rise after that, whereas ABACC 2 falls. Other financial variables have similar increasing or decreasing patterns. For instance, cash flow volatility (CFOVAR), which is the standard deviation 6 Inferences are not sensitive to winsorizing FF_Volatility at the upper and lower 1% levels.

12 of cash flows over the same period as the DD measure, steadily increases throughout the sample, whereas ROA drops over time except for the last sub-period. Overall, most of the measures display clear time-series patterns that are generally consistent with the pattern in firm age. Figure 2 graphs the resulting time-series of average firm-specific volatilities (FF_Volatility) and the two earnings quality measures, DD and ABACC 2, over the sample period. Using the same period in RV we obtain 87,660 observations after merging with CRSP data and eliminating observations with missing data necessary to calculate a reduced form of the panel data model estimated in RV (Table 3, which reports 95,270 in their study). Although the sample sizes are different, the resulting graph in Figure 2 is very similar to the reproduced figure from RV (Figure 1 from their study) displayed at the bottom of Figure 2. The FF_Volatility is reasonably similar to the Volatility measure in Figure 1, with a general increase over the sample period coupled with a shorter spike during the internet bubble followed by a downward trend after Interestingly, the trends in the DD measure and both volatility measures diverge substantially beginning in fiscal year 2001, whereas ABACC 2 falls in a similar fashion to idiosyncratic return volatility, the DD measure reaches an all-time high in fiscal year 2005 before falling just slightly in The measure more than doubles between 2000 and 2005, whereas it climbs much more gradually over the earlier time period. A potential explanation for this dramatic increase relates to the explosion of young firms documented in Figure 1. These firms are included in the annual industry regressions influencing the calculation of the standard deviation of residuals for all firms, but do not end up in the average DD calculation until the early 2000s since it requires 5 years of data prior to entering the sample. We formally investigate the influence of young firms on the accrual measures in the next section.

13 3. Firm Age and Earnings Quality In order to assess the influence of firm age on the earnings quality measures, we sort firms each fiscal year into the oldest and youngest 500 firms and plot the means for each group over time beginning with the DD measure in Figure 3. 7 We did not re-estimate the DD measure using only firms with age related data, but rather we simply merge the age database with the DD measure database and then perform the sorts. Note, if age is an important determinant and the earnings quality metrics are worse for younger firms then this likely biases against finding results since it is likely we are missing age related data on very young firms in any given year. The results in Figure 3 reveal a small positive increase in the DD measure for the group of the 500 oldest firms each year up to 2001 and then a rather steep increase beginning in 2002 extending through The 500 youngest firms show a strong positive trend throughout the sample even through 2001 followed by a similarly large positive shift in fiscal year 2002 extending once again through These data clearly reveal the increasing influence of young firms on the variance of residuals estimated from annual industry cross-sectional regressions of the DD model. Young firms are expected to have greater variation in their accruals and with greater influence over time they have caused a shift in the variance calculations for even the very oldest firms that are included in the sample. The dramatic rise in fiscal year 2002 of the DD measure requires further investigation since it represents such a large shift in the data. Based on the dramatic increase and influence of young firms during the internet bubble, we hypothesize that these same firms have a significant influence on the annual estimations of the DD model even if they do not have five years of data necessary to calculate the variance of the residuals. In an effort to better understand the influence of young firms on the calculation of DD we re-estimate the annual DD regressions using only the 500 oldest firms each year. We estimate the model in equation 1 annually on a cross-sectional basis and include 7 Inferences are unaltered if we use 100 and 250 firms instead of 500. Because of the use of 500 firms, the plots in Figure 3 do not have data during the earliest years in the sample because of the lack of 1,000 firms in the entire sample.

14 industry indicators rather than estimating by industry since this would severely reduce the sample size. Otherwise the same exact procedures used to produce the measure in Figure 2 are used. The results are presented in Figure 3 in the plot labeled DD_Old_noyoung. The deviation between this plot and DD_Old starts out small and becomes progressively larger over time and ends with a stark difference in trends beginning in Whereas DD_Old begins to spike upwards in 2002, DD_Old_noyoung stays completely flat. Over the entire sample period, there is a very slight upward trend, but nothing relative to young firms or even to the measure for old firms when young firms are included in the annual industry regressions to determine the residuals. Figure 3 demonstrates the dramatic influence young firms have on the estimation of Dechow and Dichev (2002) model residuals over time. Given the varying influence of young firms in the sample, this represents a potentially serious problem for researchers using the measure in panel data analyses or drawing inferences over time. For example, Figure 3 indicates that earnings quality for even the oldest firms has become more than twice as bad between 2001 and 2004, which is hard to comprehend. That means very old firms in 2004 now look like average young firms in the mid- 1980s in terms of earnings quality. This could cause any number of inference problems that is not easily accounted for. One potential solution to the age issue is to use discretionary DD measures following Francis et al. (2005). They estimate innate accrual quality be regressing the resulting standard deviations of DD residuals on the variation of cash flows, variation of sales, firm size, operating cycle, and the incidence of negative earnings. The resulting residual from this regression is deemed to be the true discretionary piece. Since this measure is estimated annually using OLS regressions, there is no time-series trend in the variable since the mean is always 0. This also indicates that the time-series trends of the DD measure plotted in Figures 2 and 3 are plots of the mean innate values of DD. Of course, this procedure may eliminate firms with truly poor accruals quality since the firms with the

15 greatest variance in items like cash flows and sales might also be the ones most likely manipulating earnings. Regardless, it is important to understand the changing nature of this measure over-time given the extensive use of the cross-sectional Dechow and Dichev (2002) model. Turning to Figure 4, we plot the mean ABACC s for the same 500 young (ABACC_Young) and old (ABACC_Old) firms over the time. The graph clearly illustrates the influence of young firms on the calculation of the sample mean, which is also included in the graph. Young firms dominate in terms of the calculation of the sample means, as was true with the DD measure. However, unlike the DD measure, young firms seem to have less influence on the squared residuals of old firms, whose residuals are actually relatively flat over time. To give some perspective, simple pearson correlations between the old and young means during the sample period are 0.93 and 0.43 for DD and ABACC s respectively. Although the correlations are lower, there is still clearly a strong influence of young firms on the residuals of old firms that needs to be considered in research settings using Jones (1991) model accruals over time. We illustrate one example next by reexamining the relation between idiosyncratic risk and earnings quality over time. 4. Age, Idiosyncratic Return Volatility, and Earnings Quality Figures 3 and 4 show a clear relationship between firm age and earnings quality, but as seen in Table 1 many variables change over time in a similar pattern. Although we view firm age as a clearly exogenous variable, nevertheless it is important to distinguish its influence on earnings quality versus other potential metrics. One popular variable that is commonly known to display variation with the earnings quality measures used in this study is firm size, larger firms have better earnings quality. It may be that age is simply a coarse proxy for firm size. In order to investigate this we sort firms first into quintiles by MVE and then within those quintiles on firm age. If age is

16 just a coarse proxy for firm size, we would expect the average earnings quality measures not to vary significantly within a particular size quintile. Table 2 provides the results of these double sorts. In Panel A, we examine the relationship between firm age and FF_Volatility. Firm size quintiles are reported across the top, while age quintiles are reported vertically. Within each size quintile there is a noticeable difference in idiosyncratic return volatility for the youngest firms relative to the oldest firms. For instance, in the smallest quintile of firms, the youngest firms average idiosyncratic return volatility (8.52) is 225% larger than the corresponding mean for the smallest/oldest firms (3.78). Every difference reported in the Difference (1-5) row is significantly different at the level. The consistency across the size quintiles is strong evidence that firm age is not simply a rough proxy for firm size when it comes to idiosyncratic return volatility. Turning to Panel B, we report average DD measures across the same sorts. Similar to the FF_Volatility results, there is considerable variation within each size quintile moving from the youngest to the oldest firms. Each difference between the youngest and oldest firms is again significant at the level indicating that younger firms earnings quality as measured by the DD measure is much worse than for the oldest firms within each size quintile. In fact, the influence of age is larger than the influence of size on the DD measure as can be seen by the much smaller differences reported in the last column of Table 2 (measures the differences in the DD measure across the smallest and largest firms). Not only is age not acting as a proxy for size, but it is actually more powerful in filtering firms than MVE. Finally, Panel C reports the results for ABACC s, which are consistent with both earlier results, namely that within every size quintile firm age is still an important determinate of variation of the discretionary accrual measure. Although not as dramatic as the DD measure, the differences across the AGE quintiles for ABACC 2 are of similar magnitudes to the differences across the size

17 quintiles for this variable suggesting they have similar power in explaining variation in the measure. Note, although firm age was an important factor in explaining variation in FF_Volatility, it was not as significant as the sorts of firm size. Overall, the results from Table 2 indicate that firm age is a very strong determinant of earnings quality measures. 4.1 Time-Series Regressions Given the strong relationship documented between firm age and earnings quality, along with the time-series variation in both variables, we formally investigate the relationship using aggregate time-series regressions. We use the Campbell et al. (2001) measure of firm specific return volatility estimated annually as in Fink et al. (2010) as the dependent variable in the time-series tests. Fink et al. (2010) estimate a regime switching model of idiosyncratic stock return volatility over time and find three distinct breaks in the data surrounding the oil shock in , the stock market crash in October 1987, and the internet bubble in the late 1990s. As a result they define three indicator variables to capture these shocks to idiosyncratic return volatility that we adopt in our analysis: OIL (CRASH) equals one from (1987), and zero otherwise, while BOOM equals one from , zero otherwise to capture the shock caused by the internet bubble. Following Fink et al. (2010) we use 1 minus the market capitalization of young firms (graphed in Figure 1) as our summary measure of firm age. 8 Finally, we use the fiscal year simple average of the DD and ABACC s measures graphed in Figure 2 to capture earnings quality over time similar to Table 4 in Rajgopal and Venkatachalam (2011). 9 The first set of results in Table 3 simply replicate the findings in Fink et al. (2010), but for the time period instead of Even with this reduced sample period, the results are quite consistent illustrating a positive and significant time trend in idiosyncratic risk. However, just like in Fink et al. (2010) once firm age is accounted for the coefficient on the time trend becomes statistically insignificant 8 AGE is defined this way to be consistent with the expectation that as firms get older idiosyncratic return volatility gets smaller. 9 Results are insensitive to using a value weighted average measures as the summary metrics of earnings quality.

18 and in fact flips signs. The coefficient on AGE is negative and statistically significant at conventional levels even with the reduced time period. The coefficient of implies that a onestandard deviation decrease in the percentage of old firms market capitalization results in a 1.5 standard deviation increase in idiosyncratic return volatility. Over the longer time horizon , Fink et al. (2010) obtain a coefficient estimate of on AGE, which implies a 1.5 standard deviation change in return volatility for a 2 standard deviation change in AGE. It is not surprising that the result is stronger during the sample period since the internet boom plays an even larger role than during the period used in their paper. Turning to the earnings quality metrics, first we illustrate that both have a statistically positive relationship with idiosyncratic return volatility before controlling for firm age. The positive coefficients indicate the worse the earnings quality the higher the idiosyncratic return volatility, which is entirely consistent with the findings in RV using panel data. However, once we add in AGE, neither measure is statistically significant whereas AGE maintains its significance and economic importance. It is also important to note that the adjusted R 2 in models (3) and (6) which include the earnings quality measures are both lower than the adjusted R 2 in model (2) again confirming that earnings quality has no explanatory power beyond its association with firm age. This is important since these short time-series regressions with trending variables could suffer from multicollinearity, which would result in reduced power to find statistically significant coefficients, but would not influence the predictive ability of the model (adjusted-r 2 ). In untabulated results, we do find variance inflation factors in excess of 10 for the coefficients on TIME, AGE, and the earnings quality metrics, which is consistent with multicollinearity (Belsey, Kuh, and Welsch 1980). However, given the adjusted-r 2 results coupled with the stability of the coefficient on AGE, it is clear that the earnings quality metrics association with idiosyncratic return volatility is because of the influence of firm age in their calculations.

19 Extending the sample period through 2006, which is the last year for which we have age data, the results are very consistent with those of Fink et al. (2010). The coefficient on AGE is negative and significant, shrinking towards their estimate as a result of adding more years outside of the internet bubble. The only significant change is the lack of association between idiosyncratic risk and DD. The coefficient is positive, but insignificant at conventional levels. This is clearly because of the diverging patterns of the two measures after Overall, the time-series regression tests provide strong evidence that the association of earnings quality with idiosyncratic return volatility is really a function of the changing nature of firm ages over the sample period. 4.2 Panel Data Results Our final set of tests use panel data to investigate the relationship between idiosyncratic return volatility (FF_Volatility) and earnings quality as in RV. In Table 4, we replicate their Table 3 analysis with the exception that we do not include variables related to analysts or institutional investors. These variables are not consistently available over the entire time period and do nothing to alter the results in RV thus we do not include them in our analysis. Table 4, Panel A presents results for both the period used in RV (N = 86,798) and the full sample period (N = 103,369) for both earnings quality metrics, DD and ABACC 2. The results are very similar to the findings in RV. All the signs of the variables are the same except for RET 2 (the square of the annual buy and hold return over the fiscal year) and are of similar magnitude. Virtually all of the variables are significant at conventional levels even after calculating significance using Newey-West (1987) standard errors. Note, it is not possible to cluster by firm and time given the existence of a time trend variable. The primary variable of interest is the interaction between the time trend (TIME, defined as the fiscal year less 1961) and the earnings quality metrics. The interaction terms on both DD and

20 ABACC s are positive and significant, which RV interpret as indicating that there is a stronger association between the volatility of accruals and stock returns later in the sample period. It should be noted that the main effect is negative, which is strange in that it means early in the sample period higher volatility of accruals results in lower stock return volatility. Given all the interaction terms in the analysis, multicollinearity is a possible concern and in fact is very prevalent in the estimations as can be seen from the VIF column next to the parameter estimates. Belsey, Kuh, and Welsch (1980) indicates VIF factors of 10 are indicative of multicollinearity. Although multicollinearity does not bias coefficients, it can make coefficients behave in an unstable fashion, which is evident from the estimations across the two sample periods using the DD measure. In the full sample period, the coefficient on TIME*EQ is negative and significant and the EQ main effect is now positive and significant. Taken literally, this would suggest that the influence of DD had completely switched directions after only adding five more years of data. Because of the multicollinearity, the coefficients from these regressions are quite difficult to interpret. Given collinear models are still fine for predictive purposes, we estimate the models without TIME*EQ and EQ in the model to determine the influence of earnings quality on idiosyncratic return volatility by examining the difference in adjusted-r 2 along with the associated effect on the TIME coefficient. At the bottom of Table 4, Panel A we report R 2 w/o EQ and TIME w/o EQ to measure the influence of the earnings quality measure. The results reveal that EQ has at best a marginal influence on idiosyncratic return volatility. The greatest change in adjusted R 2 is 0.4% as a result of leaving DD out of the model during the period When we extend the model five years, including or excluding EQ has almost no effect on the explanatory power of the model and in fact decreases slightly using the DD measure. Furthermore, the influence on the actual TIME coefficient is minimal, meaning even after controlling for EQ there is still a major time trend in idiosyncratic return volatility according to the results. If EQ were truly associated with the time

21 variation in idiosyncratic return volatility then it would at least diminish the economic significance of the time trend variable. In an effort to better understand the economic significance of the EQ measure, we standardize all regression variables (besides TIME) to be mean zero and have standard deviations equal to one. This allows for a relative interpretation of the significance of the variables. Furthermore, we eliminate the interactions with TIME since they do not capture much information as evident from the comparison of the adjusted R 2 across the two panels. The results from these regressions are reported in Panel B of Table 4. The first thing to notice is that the multicollinearity issue has been completely removed with no VIF greater than 3. Given the full model results in Panel A, it is not surprising that the EQ measures have essentially no influence on the adjusted R 2 of the models especially in the full sample estimations. Interpretation of the coefficients is also simplified in this model. The coefficient on EQ in the full sample DD regression is 0.025, which means a one standard deviation increase in DD results in a standard deviation increase in FF_Volatility. That means that a two standard deviation increase in DD would result in a increase in FF_Volatility, whose standard deviation is Overall, the evidence is not consistent with EQ having a major impact on idiosyncratic return volatility using the panel data methodology. 5. Conclusion This study documents the decreasing trend in earnings quality metrics measured using annual cross-sectional regressions of the Dechow and Dichev (2002) and Jones (1991) models are highly related to variation in firm age over time. Using data from Fink et al. (2010), we properly measure firm age from the earlier of the date of incorporation or the founding date and show there has been a steady decline over the sample period in the median firm age since birth. This means firms are undergoing IPOs much earlier in their life cycles during the later portion of the

22 sample, and especially during the internet bubble. These same firms are expected to have greater volatility in the accruals simply because of the greater uncertainty in their operating environments. This is consistent with the Pastor and Veronesi (2003) explanation of the rise in idiosyncratic risk surrounding the internet bubble. Not only have firms been going public at any earlier age and more frequently over the sample period, but the market capitalization of these young firms drastically increased over time with a large spike occurring during the internet bubble. The combination of the increase in the number of young firms in the sample with greater influence in the market provides young firms with unprecedented influence in the calculation of Dechow and Dichev (2002) measures of earnings quality estimated cross-sectionally on an industry basis. These age related phenomenon also are associated with temporal changes in the Jones (1991) model discretionary accrual measures. As such, researchers need to carefully consider their research designs when using accrual quality measures in order to avoid the confounding effects of firm age. As an example, we illustrate that once age is accounted for the temporal association between idiosyncratic risk and earnings quality measures is completely subsumed. This indicates there is no evidence that accounting is performing worse over time in terms of its potential influence on idiosyncratic risk. This helps to inform the debate on the usefulness of accounting data currently underway in both academic and practitioner circles. More importantly, this study illustrates the time series trends in earnings quality measures need to be carefully considered given their strong association with firm age over time.

23 References Brandt, M., A Brav, J. Graham, and A. Kumar The Idiosyncratic Volatility Puzzle: Time Trend or Speculative Episodes? Review of Financial Studies 23, Brown, G. and N. Kapadia Firm Specific Risk and Equity Market Development. Journal of Financial Economics 84, Campbell, J., M. Lettau, B. Malkiel, and Y. Xu Have Individual Stocks Become More Volatile? An Empirical Exploration of Idiosyncratic Risk. Journal of Finance 56, Cao, C., T. Simin, and J. Zhao Can Growth Options Explain the Trend in Idiosyncratic Risk? Review of Financial Studies 21, Dechow, P. and I. Dichev The Quality of Accruals and Earnings: The Role of Accrual Estimation Errors. The Accounting Review 77 (supplement), Fama, E. and K. French Industry Costs of Equity. Journal of Financial Economics 43, Fama, E. and K. French Disappearing Dividends: Changing Firm Characteristics or Lower Propensity to Pay? Journal of Financial Economics 60, Fink, J., K. Fink, G. Grullon, and J. Weston What Drove the Increase in Idiosyncratic Volatility During the Internet Boom? Journal of Financial and Quantitative Analysis 45 (5), Francis, J., R. Lafond, P. Olsson, and K. Schipper The Market Pricing of Accruals Quality. Journal of Accounting and Economics 73, Jones, J Earnings Management During Import Relief Investigations. Journal of Accounting Research 29, Jovanic, B. and P. Rousseau Why Wait? A Century of Life before IPO. American Economic Review 91, Kothari, S.P., A.J. Leone, and C. E. Wasley Performance Matched Discretionary Accrual Measures. Journal of Accounting and Economics 39, Landsman, W. R., and E. L. Maydew Has the Information Content of Quarterly Earnings Announcements Declined in the Past Three Decades? Journal of Accounting Research 40, Lev, B. and P. Zarowin The Boundaries of Financial Reporting and how to Extend Them. Journal of Accounting Research 37, Loughran, T. and J. Ritter. Why Has IPO Underpricing Changed Over Time? Financial Management 33, 5-37.

24 Newey, W. K. and K. D. West A Simple Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix. Econometrica 55, Pastor, L. and P. Veronesi Stock Valuation and Learning about Profitability. Journal of Finance 58, Rajgopal, S. and M. Venkatachalam Financial Reporting Quality and Idiosyncratic Return Volatility. Journal of Accounting and Economics 51, Ryan, S. and P. Zarowin Why has the Contemporaneous Linear Returns-Earnings Relation Declined? The Accounting Review 78,

25 Figure 1 Volatility and Market Capitalization of Young Firms This figure plots volatility calculated using the methodology in Campbell et al. (2001) which does not require calculation of firm level covariances like in Figure 1. Firm Age is calculated from the earlier of the firm incorporation date, listing year, or the founding date. Young firms are defined as those firms less than or equal to 20 years of age. The graph plots the percentage of total CRSP market capitalization represented by young firms as of December of each calendar year.

26 Figure 2 Firm Volatility, Dechow and Dichev Model, and Jones Model Accruals over Time This figure plots volatility calculated by averaging the monthly volatility of residuals from the Fama French 3 factor model using daily data within a month multiplied by the number of trading days within the month. The Dechow and Dichev (2002) measure (DD) and Jones (1991) measure (ABACC 2 ) are calculated within industry each fiscal year where there are at least 20 observations within an industry/fiscal year. The DD model is supplemented with the change in revenues and gross property plant and equipment. All variables in the model are scaled by average total assets including the intercept (the model is estimated without a traditional intercept term). All variables are winsorized at the upper and lower 1% levels prior to estimating the DD regression. The DD measure is the standard deviation of firm specific residuals over the period t-4 to t. Firms are required to have observations in all 5 periods to calculate the standard deviation. DD and ABACC 2 are winsorized at the upper and lower 1% levels prior to being graphed. The graph on the bottom is a reproduction of the graph included in Rajgopal and Venkatachalam (2011) as Figure 1 in that study.

27 Figure 3 Dechow and Dichev (2002) Measure for Young and Old Firms This figure plots the Dechow and Dichev (2002) measure (DD) for the 500 youngest and oldest firms each fiscal year. The DD measure is calculated within industry each fiscal year where there are at least 20 observations within an industry/fiscal year. The model is supplemented with the change in revenues and gross property plant and equipment. All variables in the model are scaled by average total assets including the intercept (the model is estimated without a traditional intercept term). All variables are winsorized at the upper and lower 1% levels prior to estimating the DD regression. The DD measure is the standard deviation of firm specific residuals over the period t-4 to t. Firms are required to have observations in all 5 periods to calculate the standard deviation. The standard deviation is winsorized at the upper and lower 1% levels prior to being graphed. The plot of DD_Old_noyoung represents the average DD metric estimated using only the 500 oldest firms each year. This regression is estimated annually on a cross-sectional basis controlling for industry rather than by industry. Otherwise the same procedures are used to obtain the plot other than the fact that DD_Old_noyoung is not winsorized prior to being plotted (winsorization does not alter the inferences).