2013年5月15日学术报告通知.pdf

Estimating the Amount of Estimation in Accruals Jason V. Chen University of Michigan Ross School of Business jvchen@umich.edu Feng Li University of Michigan Ross School of Business feng@umich.edu May 8, 2013 Abstract This paper examines the link between the amount of estimation needed during the accrual generating process and the persistence of the accruals portion of earnings. We measure the amount of estimation needed during the accrual generating process using the number of estimation-related linguistic cues in the notes to the financial statements and the critical accounting policies section of the management discussion and analysis. Consistent with the conjecture in Sloan [1996], we find that accruals that need more estimation are less predictive of future earnings. We also find that such accruals map less into the past, current, or future cash flows in the spirit of Dechow and Dichev [2002]. When we decompose the number of estimation linguistic cues into a component that is due to the existence of specific accruals accounts and a component that is due to the within-accounts variations, we find that our results are driven by both components. Lastly, we find mixed evidence as to whether the amount of estimation in accruals is systematically associated with the accrual anomaly. Overall, our results suggest that the estimation needed during the accrual generating process plays an important role in understanding the persistence of accruals. We thank Matt DeAngelis, Ilia Dichev, Scott Richardson, Lakshmanan Shivakumar, Irem Tuna, and the workshop participants at University of Buffalo, Columbia University, Florida Atlantic University, Florida State University, London Business School, London School of Economics, University of Michigan, Michigan State University, Purdue University, Rice University, and Temple University for their comments. We gratefully acknowledge the support of the Ross School of Business and the financial support from the Paton Accounting Fellowship and the Harry Jones Endowment. 1. Introduction We propose a new approach to measure the quality of accruals based on the amount of estimation embedded in the accruals, calculated as the number of estimation-related linguistic cues in the notes to the financial statements and the critical accounting policy (CAP) disclosures in the management discussion and analysis. We empirically investigate the association between this new measure of accruals quality and how predictive the accruals are of future earnings. Our empirical tests are motivated by Sloan [1996] and Richardson et al. [2005], who argue that the greater estimation needed in the accrual generating process explains why accruals are less persistent than cash flows. Unlike cash flows, accruals incorporate estimates of future cash flows, cash flow deferrals, depreciation and amortization, and fair value estimates. While this argument is a well-accepted conjecture, few studies have explicitly examined how the estimation involved in the accrual generating process relates to the persistence of accruals (Dechow et al. [2010]). Rather, prior research has primarily examined how the different components of accruals explain its persistence. For example, Dechow and Ge [2006] finds that accruals exhibit lower persistence when they contain special items. Richardson et al. [2005] examines the difference in persistence of working capital, non-current operating, and financial asset accruals. We provide complementary evidence by linking accruals persistence directly to the characteristics of the accruals generating process. We focus on the footnotes to the financial statements and the critical accounting policies disclosures in the management discussion and analysis to gauge the amount of estimation in accruals since they contain detailed descriptions of the nature of the accruals and how they are generated. Auditors and investors, however, have questioned their usefulness due to the large amount of boilerplate and immaterial information in these disclosures (Radin [2007]). Our empirical tests therefore can also be viewed as joint tests of the usefulness of the footnotes and CAP disclosures for assessing accruals quality. 1 We start out by reading numerous footnotes and CAP disclosures in the financial statements to find estimation-related linguistic cues. We find that in general, these linguistic cues can be broken down into three linguistic relations: (1) An estimation action targets some object (e.g. “we estimated receivables”), (2) An estimation object is the target of a use action (e.g. “we used estimates”), (3) An estimation word is an adjective to an object (e.g. “estimated costs”). We construct a dictionary of estimation-related words (see Appendix 2 for details) and count the number of estimation-related linguistic cues for the three linguistic relations. We then use statistical parsing techniques to automate the search for such linguistic cues in the footnotes and CAP section of firms’ 10-K filings (Appendix 1). Our first hypothesis is that accruals that require more estimation are less predictive of future earnings. Consistent with this hypothesis, we find that accruals are significantly less persistent when there is greater estimation conveyed in the company’s footnotes and CAP disclosures. In contrast, we find that the amount of estimation conveyed in the company’s footnotes and CAP is not informative of the persistence of cash flows. These findings support Sloan’s conjecture that the amount of estimation needed in accruals partially explains their lower persistence. We also find that accruals that involve more estimation also have lower quality measured in the sense of Dechow and Dichev [2002], i.e., these accruals map less into past, current, or future cash flows. This is consistent with the hypothesis that more highly estimated accruals are less precise, and therefore have greater errors, and provides further evidence that the estimation needed during the accrual generating process drives the lower persistence. A natural question arises whether the number of estimation cues simply captures the types of accruals accounts. 1 For instance, firms that have defined benefits pension plans are likely to have more estimation cues in the footnote because pension calculations need to estimate 1 We thank Irem Tuna for this observation. 2 more parameters. We therefore decompose the number of estimation cues in the footnotes and CAP disclosures into two components: a component that is due to the existence of specific accruals accounts and a component that is due to the within-accounts variations. Specifically, we identify 49 common accounting items (see footnote 6) and regress the number of estimation cues on the item fixed effects. The predicted value from this model captures the number of estimation cues due to the existence of different accruals accounts and the residual captures the withinaccounts variations. We find that the persistence of accruals and the quality of accruals are driven by both components of the estimation cues. Lastly, we examine the accrual anomaly as a function of the estimation cues in the footnotes and CAP disclosures. We find mixed evidence as to whether the market reacts as if it does not incorporate the amount of estimation in accruals into its valuation of the firm in a timely manner. Specifically, we find some evidence that the accrual anomaly documented in prior studies (Sloan [1996], Xie [2001], Mashruwala et al. [2006]) is more significant when more estimation is needed during the accrual generating process. 2 This is consistent with the hypothesis that the market more greatly overvalues (undervalues) highly estimated positive (negative) accruals in the short term. However, the robustness tests of our findings using the Carhart four-factor Alpha model provide mixed results. We conduct a battery of robustness checks for our empirical results. First, we build pseudo word counts by randomly selecting a dictionary of words that have similar aggregate frequencies as our estimation cue words in the 10-K samples. The bootstrapping simulation results show that the probability of finding the results documented in the paper using the pseudo dictionary is low. Second, we extract the non-critical accounting policies part of the management discussion and analysis in 10-Ks and count the number of estimation-related linguistic cues in these disclosures. We find that the number of estimation-related linguistic cues in the non-CAP 2 Our study uses one-year abnormal returns which begin 5 days after the 10-K filing. 3 part of MD&As is not associated with accruals or earnings quality. This suggests that our measure of the amount of estimation in accruals does not simply capture generic business uncertainty. Lastly, we control for numerous other textual disclosure characteristics of the financial statements documented in prior studies, including the Fog index (Li [2008]) and the competition measures (Li, Lundholm, and Minnis [2012]), and our results are robust to these controls. We make several contributions to the literature. First, this study incorporates the amount of estimation needed during the accrual generating process as reflected in a firm’s notes to the financial statements and critical accounting policies disclosures into our understanding of the persistence of accruals. Many prior studies have ignored this important source of information about accruals when examining the persistence of accruals.3 Second, our findings strengthen Sloan’s argument that the estimation involved in accruals explains the difference in the persistence of the cash portion of earnings and the accruals portion of earnings. Some prior studies argue that the difference in the persistence of accruals and cash flows documented by Sloan is driven by omitted fundamental differences such as growth. Our findings suggest that estimation does explain the lower persistence of accruals. Finally, this study contributes to the textual analysis accounting literature by using grammatical relationships to extract meaning from qualitative financial information. These relationships provide structure to the qualitative information and allow us to better infer meaning from the text. Additionally, this study adds to a growing field of textual analysis studies which suggest that using qualitative accounting information, in conjunction with quantitative accounting information, helps to provide a richer understanding of firms and their accounting process (Li [2011]). 3 Richardson et al. [2005] is an exception in that they test the association between accruals estimation and persistence by examining the properties of the different components of accruals. 4 The remainder of the paper proceeds as follows. Section 2 provides a discussion of prior literature and motivation for our hypotheses. In Section 3 we discuss how the sample of 10-K footnotes and CAP disclosures and financial information is prepared. In Section 4 we present the research design and main results. Section 5 concludes the paper. 2. Prior Literature and Hypotheses Sloan [1996] finds that the accruals portion of earnings is less predictive of future earnings than the cash portion earnings (i.e. accruals are less persistent than cash flows). He argues that the difference between the persistence of accruals and cash flows is due to the greater estimation needed when deriving accruals, as accruals incorporate estimates of future cash flows, depreciation and allocations, deferrals, and valuations. Richardson et al. [2005] expands the hypothesis in Sloan [1996] and formally models accruals estimation as an error-in-variables problem. Their model assumes that if accruals can be measured without error, there is no need to correct for accrual errors in future periods. However, actual recorded accruals are measured with error since managerial estimation is needed during the accrual generating process. This error reduces the association between accruals and future earnings. To test the hypothesis, prior studies have examined how the specific components that comprise accruals affect how well accruals predict future earnings. For instance, Dechow and Ge [2006] examines the persistence of low accruals when the firm has special items. Consistent with their hypothesis, they find that accruals are less persistent when the firm has special items. Richardson et al. [2005] disaggregates accruals into financing accruals, working capital accruals, and non-current operating accruals. They posit that the accruals in each of the three categories have different degrees of estimation. Financial accruals require less estimation than working capital or non-current operating accruals because their terms are typically contractually defined. Therefore estimates of future cash flows are well defined and require a lower degree of 5 estimation. On the other hand, estimates of future cash flows, valuations, and other estimates are needed when recording working capital and non-current operating accruals. The greater estimation in these accruals implies that these accruals are less likely to be realized in the cash flows and therefore will be less informative of future earnings. Consistent with their hypothesis, they find that financing accruals are more persistent than working capital and non-current operating accruals. In this paper, we argue that while on average accruals need more estimation than cash flows, not all accruals are created equal. Due to differences in business fundamentals, accounting policies, and earnings management incentives, accruals reported by different companies have different levels of estimations and therefore different quality. Even if two companies have the same total dollar amount of accruals, the amount of estimation in the accruals of these two companies may be vastly different. For example, one company’s total accruals may contain a large amount of estimated fair value accruals while another may contain a large amount of financial accruals involving less estimation. We fill the gap in the literature by explicitly measuring and examining the implications of the cross-sectional differences in the amount of estimation in accruals on accruals persistence. Accordingly, our first hypothesis follows the conjectures of Sloan [1996] and Richardson et al. [2005] and is as follows: Prediction 1: Accruals that involve more estimation are less persistent. Dechow and Dichev [2002] finds that firms that exhibit a lower mapping of accruals into past, current, and future cash flows also exhibit lower earnings persistence. They posit that if accruals map less into these cash flows then accrual errors must be greater (i.e. the accruals are recorded with low precision). Thus, accruals will be less predictive of future earnings. 6 We argue that if there is greater estimation in accruals then these accruals are likely to be recorded with lower precision (i.e. accruals map less into realized cash flows). If managers make a large number of estimations when recording accruals then the range of possible errors in the recorded accruals is greater. When accruals have greater error, they are less realized as cash in prior, current, or subsequent periods. Following this reasoning, greater estimation during the accrual generating process will be associated with a lower mapping of cash flows into the accruals portion of earnings. Prediction 2: Accruals that involve more estimation map less into the firm’s past, current, or future cash flows. Prior studies have found that the lower persistence of accruals is not quickly incorporated by investors in their valuations (Sloan [1996], Hanlon [2005], Richardson et al. [2005]). One explanation for this finding is that investors fixate on total earnings thereby disregarding the affect of the lower persistence of accruals on how predictive current earnings are of future earnings (Sloan [1996], Kraft et al. [2006]). Accordingly, Sloan [1996] finds that the future abnormal returns are negatively associated with the magnitude of firms’ accruals. If investors fixate on total earnings and ignore the difference between accruals and cash flows, then they may not fully incorporate the information in the amount of estimation in accruals in the footnotes or CAP disclosures in a timely manner. This reasoning suggests that the amount of estimation in accruals could exacerbate the accrual anomaly, i.e., accruals that need more estimation are more likely to be mispriced and negatively associated with future stock returns. On the other hand, investors may quickly incorporate the amount of estimation in accruals into their valuation of the firm since this information is readily available in the firm’s 7 10-K filings. More specifically, information provided in a firm’s footnote disclosures has been shown to be incorporated by both investors and analysts (De Franco et al. [2011]). If the amount of estimation involved in the accruals portion of earnings can be found in the notes to the financial statements or CAP disclosures then investors may become informed of the lower persistence of these earnings upon the filing of the 10-K. If so, then the estimation information found in the footnotes or CAP disclosures will not be associated with the future long term abnormal returns of the firm. This leads to our third prediction, stated in the null hypothesis format: Prediction 3: The market reacts as if it does not incorporate the amount of estimation in accruals in their valuation of the firm in a timely manner. 3. Data Preparation 3.1 Extracting the Footnotes to the Financial Statements We download all 10-K documents filed with the SEC for fiscal years between 1995 and 2010 from the SEC EDGAR Website.4 We then extract the notes to the financial statements and the critical accounting policies disclosures in the management discussion and analysis section from each of the 10-K filings using Perl. The extracted footnotes and CAP disclosures are stripped of all HTML tags and tables. To mitigate any data issues related to extracting the notes to the financial statements from each 10-K filing we truncate our sample of notes to the financial statements and critical accounting policies by the total number of words at the 1% and 99% level. We also eliminate any filings which are not explicitly identified as either a “10-K” or “10K405” in the filing’s header. 3.2 Measuring Accruals Estimation 4 Companies began filing using EDGAR beginning in 1994-1995. 8 Our textual analysis approach is used to capture the amount of estimation needed when the accruals are recognized. We focus on the notes to the financial statements section and the critical accounting policies section of the 10-K because they provide information specific to the accounting process. The notes to the financial statements provide a wealth of information not found in other sections of the 10-K filing (Merkeley [2011], Riedl and Srinivasan [2010]). More importantly, this section provides comprehensive information pertaining to the estimations made and the assumptions needed by management during the accrual generating process. In May 2002, the Securities and Exchange Commission (SEC) begin to require firms to include critical accounting policy (CAP) disclosures in their 10-K filings. In particular, the SEC requires managers to provide a discussion of those accounting policies that involve highly uncertain assumptions for which differing estimates would have a material influence on the firm’s financial statements (Billings [2011]). Levine and Smith [2011] show that CAP disclosures influence investors’ valuation decisions. For our purpose, the disadvantage of the CAP disclosures is that it only covers the latter part of our sample period (i.e., post-2002 years). We read numerous notes to the financial statements and critical accounting policies disclosures to identify linguistic cues commonly used to denote that some estimation was needed. The first linguistic cue is when an estimation action targets some object. For example, the phrase “we estimated receivables” contains the estimation action “estimated” which targets the object “receivables”. This cue denotes that receivables were estimated. Another linguistic cue is when a “use action” targets an estimate object. An example of this is the phrase “we used estimates” where the action “used” targets the object “estimates”. Lastly, the use of an estimate adjective to modify some object also conveys that something was an estimate. An example of this is “estimated costs”; here the object “costs” is being modified by the adjective “estimated” thereby conveying that the costs are estimates. 9 We automate the search for these linguistic cues by first parsing each of the sentences in our sample using the Stanford open source statistical parser (Marneffe et al. [2006]). The parse of each sentence identifies its noun modifiers, direct object modifiers, adjective modifier, etc. (i.e. its grammatical relationships). Deconstructing sentences in this manner not only provides us with a map of the qualitative information but, more importantly, allows us to utilize the grammatical relationships between the words in the each sentence. Using the grammatical relationships allows us to more correctly identify linguistic cues that convey that some estimation is needed by management (Klein and Manning [2003], see Appendix 1). Next, we construct fours dictionaries to help us extract meaning from the sentence parses. The first dictionary contains Estimation Actions. Words in the Estimation Actions dictionary convey that an estimation action was performed – this dictionary includes words such as “Estimate,” “Anticipate,” and “Approximate.” The second dictionary is of Estimation Objects (Nouns). This dictionary contains estimation related objects and contains words such as “Belief,” “Estimates,” and “Approximations.” The dictionary of Estimation Objects is used in conjunction with a word from our Use Words dictionary, our third dictionary. This dictionary includes words that denote that management use or need some object and includes words such as “Make,” “Use,” and “Include.” Our fourth and final dictionary is Estimation Adjectives and contains estimation words that are used to modify some object - these words include “Likely,” “Estimated,” and “Anticipated”. Finally, using our sentence parses and dictionaries we examine each sentence in our sample of footnotes and CAP disclosures for the linguistic cues that we identified as conveying that estimation is needed or used by management (see Appendix 2). 5 The number of these 5 We look for direct objects, nominal subjects, noun compound modifiers, adjectival modifiers, and quantifier phrase modifiers which convey that an estimation was made or used. 10 linguistic cues is used as our measure of the amount of estimation needed by management during the accrual generating process. 3.2.1 Measuring Between-Accounts Estimations and Within-Accounts Estimations We also decompose the number of estimation cues into a component that is due to the existence of specific types of accruals accounts (“Between-Accounts Estimations” or BAE) and a component that is due to within-account variations (“Within-Account Estimations” or WAE). For instance, a firm could have more estimation cues simply because it has defined benefits pension plans, which tend to involve more estimations than other accruals. On the other hand, it is also possible that there are different levels of estimations for the same types of accounts or transactions across companies. For example, the calculation of uncollectible receivables of one company may need to estimate more parameters compared with those for another company. We measure BAE and WAE by calculating the expected amount of estimation needed given the specific items in the company’s notes to the financial statements and CAP disclosures. Specifically, we extract all footnote headers from our sample of 10-K filings and sort them based on their frequency. Starting with the most frequent footnote headers, we categorize approximately one thousand unique footnote headers by hand and categorize them into 49 unique footnote items.6 7 Using this list of unique footnote headers we then find which items appear in each company’s notes to the financial statements. We then regress the number of estimation cues on the account item fixed effects by industry and year: 6 The list of footnote headers are: Taxes, Accounting Policies, Commitments, Contingencies, Affiliates, Stock, Long-term Debt, Subsequent Events, PP&E, Inventory, Pension and Retirement, Mergers and Acquisitions, Financial Instruments, Earnings Per Share, Segment Information, Leases, Financial Data, Discontinued Operations, Investments, Stock Options, Payables, Cash, Intangibles, Stock Compensation, Business, Cash Flows, Other Assets, Receivables, Credit Arrangements, Regulatory, Derivatives, Going Concern, Credit Risk, Fair Value, Comprehensive Income, Significant Customers, Accounting Changes, Restructuring, Allowance, Parent Company, Restatement, Shareholder Rights, Loan, Dividends, Real Estate, Other Expenses, Joint Ventures, Supplemental Information, and Reinsurance. 7 The 1,000 hand categorized footnote headers directly account for approximately 70% of all footnote headers from our sample. This hand categorized sample was use to seed our Perl script which searched for footnote headers. 11 , where , _ Σ , , , (1) is the number of estimation-related linguistic cues in firm f’s notes to the financial statement and CAP disclosures in year t. _ , , is an indicator which equals 1 if the company’s notes to the financial statements contain the specific footnote item j. The predicted (residual) value from this model captures BAE (WAE), or the number of estimation cues explained by the existence of different accruals accounts (the number of estimation cues due to the within-accounts variations). Appendix (3) presents the top 10 and bottom 10 account items in terms of the amount of estimation as indicated by the coefficients in equation (1). Among the commonly seen transactions, the recording of “fair value,” “intangibles,” “derivatives,” “restructuring,” “discontinued operations,” and “contingencies” has the most amount of estimation-related linguistic cues; “PP&E,” “long-term debt,” “inventory,” and “taxes” have the least amount of estimation-related linguistic cues. We examine whether BAE and WAE have different implications for accruals persistence and quality. Ex ante, we posit that both BAE and WAE explain the lower persistence of accruals in comparison to cash flows. 3.3 Sample Preparation We merge the estimation count data with annual financial information from the Wharton Research Data Services (WRDS) Compustat database and equity market information from the Center for Research in Security Prices (CRSP). For a handful of the firms in our sample were unable to find corresponding financial data or market information. The main reason for many of these stemmed from not being able to find an appropriate GVKEY for the CIK specified in the header of the 10-K filing. We then eliminate financial institutions from our sample due to the 12 potential idiosyncratic nature of their accruals and disclosures.8 This leaves us with a sample size of 64,411 firm year observations. Future long window abnormal returns of the firm are calculated as the compounded returns of the firm minus the compounded returns of the market over the same window. Specifically, we calculated one-year compounded returns beginning five days after the filing of the 10-K. We also calculated one-year value weighted compounded market returns beginning five days after the 10-K filing for each of the firms in our sample. Compounded abnormal returns for each firm are the calculated by subtracting the one-year value weighed compounded market returns from the one-year compounded returns of the firm. 3.3.1 Summary Statistics Table 1 presents the average estimation-related linguistic cues and the average total number of words found in the notes to the financial statements and CAP disclosures. Consistent with prior studies, we find that the average length of the footnotes and CAP disclosures has steadily increased over time (Li [2008]). On average, the length of the notes to the financial statements and CAP disclosures has doubled in size over our sample period going from an average of 5,194 words for the fiscal period 1995 to 14,695 words in 2010. This finding is also consistent with prior studies and anecdotal evidence that suggest that firms’ financial disclosures have been increasing in complexity (Radin [2007]). We also find that the amount of estimationrelated linguistic cues has increased monotonically during our sample period. Specifically, the number of linguistic cues that convey estimation increased from an average 33 in the fiscal period 1995 to an average of 130 in 2010. Figure 1 plots the trend of estimation and of the length of the footnotes and CAP disclosures (as measured by the total number of words in the footnotes and CAP disclosures) and reiterates our description of the trend in both footnotes length and estimation. Overall, the length 8 Financial firms are identified as those firms having SIC codes between 6000 and 6999. 13 of the footnotes and CAP disclosures has been growing over the years in our sample. Moreover, the estimation count has been growing as well. There appears to be a slight leveling off in the growth of the footnotes in the later periods of our sample. To control for the time-trend in the length in footnotes and CAP disclosures, we include year fixed effects and the total number of words in the empirical tests. Table 2 shows the average estimation and the average number of words in the notes to the financial statements and CAP disclosures by industry. Industry appears to play an important role in the amount of estimation. Agricultural production crops, automotive repair, building materials, social services, and construction contractors are the 5 industries with the least amount of estimation having an average of approximately 36 estimation linguistic cues in their footnotes. On the other hand. the industries of coal mining, communications, paper and allied products, and oil and gas constitute the 5 industries with the most amount of accruals estimation and have an average of 102 estimation related linguistic cues in the footnotes. 4. Research Design and Results 4.1 Determinants of Estimation In this section, we perform cross-sectional tests to examine the associations between certain firm characteristics and the amount of estimation needed in its accruals. This exercise serves two purposes. First, it provides some intuitive validation to our measure. Second and perhaps more importantly, these firm characteristics could be potential control variables when we examine the implications of our estimation measure for accruals quality. We examine the following set of variables identified in prior studies (Dechow and Dichev [2002], Francis et al. [2005]) as potential determinants of the amount of estimation in accruals: 14 Size Larger firms typically have more operational complexity than smaller firms. This suggests that greater estimation is needed to convey the activities of the firm through accruals. However, everything else equal, the transactions of larger firms can have diversification effects, which may make the estimation of accruals more precise. For instance, firms with a diverse set of receivables may be able to estimate the bad debt ratio more precisely to the extent that the different sources of receivables offer some diversifications. Therefore, ex ante we do not have a clear prediction on the association between firm size and the amount of estimation in accruals. Negative Earnings Accounting conservatism suggests that accountants tend to require a higher degree of verification to recognize good news as gains than to recognize bad news as losses (Basu [1997]). This suggests that positive earnings need a greater degree of certainty to be recognized while negative earnings may be recognized with less precision (i.e., more estimation). We therefore hypothesize that firms with negative earnings are more likely to have more estimation needed in their accruals. Operating Cycle Longer operating cycles imply a longer horizon for accruals to be mapped into realized cash flows. Therefore more estimates (and more imprecise estimates) may be needed by management when calculating and recognizing accruals. We predict a positive association between the operating cycle of a firm and the amount of estimation in its accruals. Standard Deviation of Cash flows and Standard Deviation of Sales Firms with more volatile business environment are more likely to have accruals that need more estimation. We use two variables, the standard deviation of cash flows and the standard deviation of sales (both scaled by the book value of assets), to test for this positive prediction. 15 We use both variables because sales could be affected by the amount of accruals booked (e.g., receivables) and cash flows are not subject to this problem. We perform our cross-sectional test using the following tobit model left censored at 0 to examine the relationship between the determinants of the amount of estimation in accruals: , Size , , where _ 360 sold. is defined earlier. S , , , (2) , , , , is the log of the market value of the firm’s equity. is the length of the operating cycle of the firm calculated as log , ∗ ∗ 360 , where invt is the average inventory balance and cogs is the cost of goods , is the standard deviation of the firms sales (scaled by the book value of assets) over the past five years. , is the standard deviation of the firms cash flows (scaled by the book value of assets) over the past five years. , is the number of years that the firm had negative earnings over the past 5 years. The empirical results based on a simple Ordinary Least Squared regression are similar to those based on the tobit model. 4.1.1 Determinants of Estimation Findings Table 5 presents the results for the determinants of estimation in accruals. Consistent with our expectations, the standard deviation of sales and the length of the operating cycle are positively associated with Estimation, and the number of the past five years in which the firm has 16 negative earnings is negatively correlated with Estimation. The positive and significant coefficient on size suggests that larger firms tend to report accruals that need more estimation. One of the proxies for the volatility of business environment, the standard deviation of cash flows, is statistically significant but the coefficient loads is in the opposite direction of our prediction. One explanation for this surprising result is that in an uncertain environment, managers simply may not book the imprecisely estimated accruals in the financial statements since they are very uncertain of future cash flows due to the high volatility. 4.2 Estimation and the Persistence of Accruals To test our first prediction (P1), we examine how the amount of estimation in accruals is related to the persistence of earnings and the persistence of accruals relative to cash flows. We follow prior literature (Sloan [1996], Li [2008]) and measure the persistence of earnings (accruals and cash flows) by regressing the following year’s earnings on the current year’s earnings (accruals and cash flows). All variables are scaled by the average book value of assets for the period. Fundamentally, this regression measures how predictive current earnings (accruals and cash flows) are of future earnings. If the estimated coefficient on current earnings (accruals and cash flows) is high then we would conclude that current earnings (accruals and cash flows) are highly persistent since they are highly associated with future earnings and vice versa when the estimated coefficient on earnings in the regression is low. First, we examine the marginal effect of the amount of estimation in accruals on the persistence of total earnings. We include the interaction between the current year’s earnings and the amount of estimation to measure the impact of estimation on the persistence of earnings. For a given level of earnings, how much more (or less) persistent are they for a given level of estimation. If our hypothesis is correct then we should find a negative coefficient on this interaction term. 17 , ∗ , , , Σ IndustryFE where , is defined earlier and (3) , , ∗ , , is income before extraordinary items. The controls include size, operating cycle, standard deviation of sales, standard deviation of operating cash flows, the number of years over the past 5 years in which the firm had negative earnings, and the total length of the footnotes. The first five control variables are the determinants of the amount of estimation in accruals as examined in Table 5. They are included to make sure that Estimation does not simply capture these common firm characteristics. We include the total length of the footnotes to rule of the possibility that Estimation does not simply proxy for the length of the document, since longer footnotes and CAP disclosures are more likely to contain more estimation related linguistic cues. We also include interactions between all control variables and earnings and auditor, year, and industry fixed effects.9 All continuous variables are scaled by average total assets. Next, we disaggregate earnings into cash flows and accruals and interact each component with our measure of estimation. We follow the recommendations made in Hribar and Collins [2002] and calculate accruals using the statement of cash flows. If greater estimation lowers the association between the current year’s accruals the following year’s earnings then the interaction between estimation and the accruals portion of earnings will be negative. Ideally, if the number of linguistic cues that convey estimation in the footnotes and CAP disclosures does not capture 9 We include the interaction between the control variables and earnings since we want to control for the marginal impact of the control variable on the persistence of earnings in addition to the control variables impact on future performance. 18 the precision of the cash portion of earnings the interaction between cash flows and estimation should be statistically insignificant.10 , , ∗ , , Σ , , , ∗ , ∗ , Σ IndustryFE where , , , (4) , , , ∗ , , , and the control variables are as defined in equation (3); is the portion of total earnings due to operating cash flows and A , is the portion of total earnings due to accruals. Like other continuous variables in the regression, both Cash and Accruals are scaled by the average book value of assets. In addition to controlling for the common firm characteristics when examining the accruals and cash flows persistence test in equation (4), we also include three accruals quality measures documented in prior studies. Specifically, we include the absolute value of the magnitude of accruals (Sloan [1996]), the standard deviation of the Dechow and Dichev [2002] residual, and special items (Dechow and Ge [2006]) in our tests of accruals persistence. By examining the incremental power of Estimation in explaining accruals persistence, we can test whether the footnotes and CAP disclosures can provide additional information about accruals quality that is not reflected in other commonly used accruals quality measures. 4.2.1 Estimation and the Persistence of Accruals Findings Table 6 presents the results for the regression of next year’s earnings on current year’s earnings, estimation, the interaction between estimation and the current year’s earnings, and our 10 The measure of estimation may also capture business uncertainty about the firm. If so then the coefficient on the interaction between cash flows and estimation will also be negative and statistically significant. 19 controls (4). The coefficient on the interaction between estimation and earnings in the current year is negative and statistically significant at 1%; this result suggests that earnings that need more estimation are less persistent than those that required less estimation. The economic significance of the effect of estimation on the persistence of earnings when going from the 25th percentile of Estimation to the 75th is approximately -0.063 (-0.0009 * (106 – 36)); this translates into a percentage difference of approximately 10% to 15% when compared to the baseline persistence of accruals (Dechow et al. [2006]). Overall, these findings suggest that the amount of estimation needed during the accrual generating process is associated with an economically significant decrease in the persistence of total earnings. The 3rd Column of Table 6 shows the interactions between BAE and WAE and earnings. Both interaction terms are statistically significant at the 1% level. The magnitude of the difference in the persistence of earnings when going from the 25th percentile of BAE to the 75th percentile of BAE is -0.072 (-0.0010 * (111 – 39)). The economic magnitude when going from the 25th to the 75th percentile of WAE is -0.0168 (-0.0008 * (9 + 12)). As discussed before, accruals are one component of total earnings and the measure of the amount of estimation should only pertain to the accruals portion of earnings and not the cash flows portion. Table 7 shows the results when we interact each component of earnings separately with our measure of estimation.11 As predicted the interaction between accruals and estimation is negative and statistically significant; this result is consistent with our prediction that accruals that need more estimation exhibit lower persistence. The difference in the persistence of accruals between the 25th percentile of Estimation and the 75th percentile is approximately -0.07 (-0.0010 * (106 - 36)). Next, the results show that the coefficient on the interaction between cash flows 11 In untabluated results, consistent with prior research we find that the persistence of accruals is less than that of cash earning and that the magnitudes of the coefficients are similar to those found in prior research. 20 and the amount of estimation is statistically insignificant. This finding suggests that our measure of estimation captures some characteristic of accruals but not cash flows. The 5th Column of Table 7 shows the interaction between the amount of BAE and WAE and both accruals and cash flows. Our findings suggest that the lower persistence of accruals in comparison to cash flows is driven by both components of estimation. The interaction between the amount of BAE and earnings and the interaction between WAE and earnings are both statistically significant. The difference in the persistence of earnings when going from the 25th percentile to the 75th percentile of BAE is approximately -0.0792 (-0.0011 * (111 – 39)) while the difference in the persistence of accruals is -0.0168 (-0.0008 * (9 + 12)). An F-test of the coefficients on the interaction between BAE and accruals and WAE and accruals yields a p-value of 0.6457 thereby suggesting that the coefficients are not statistically different. Table 8 presents the accruals and cash flows persistence test in equation (4) by further including three accruals quality measures documented in prior studies: the absolute value of the magnitude of accruals, the standard deviation of the Dechow and Dichev [2002] residual, and special items. We estimate the Dechow and Dichev model as modified by McNichols [2002] by industry and year and use the standard deviation of the residual from the model over the past 5 years for each firm as a measure of accruals quality (see equation (5) in the next section for more details). Table 8 Column 4 of Panel A shows that consistent with the findings in prior studies, all three measures are statistically significant at the 1% level and negatively associated with the persistence of accruals. Moreover, the standard deviation of the Dechow and Dichev residual and Special Items are associated with a lower persistence of cash flows; this result suggests that these measures pickup uncertainty about the cash flows of the firm as well. Importantly, even after including these alternate measures of estimation we see that our measure of estimation is still associated with a lower persistence of accruals at the 1% level of 21 significance. However, the economic significant of our measure has decreased to -0.049, a reduction of approximately 30% from the economic significance found in Table 7. Even so, these findings suggest that our measure is informative about some aspect of accruals persistence not found in these other measures. Panel B shows the results when we disaggregate estimation into BAE and WAE. BAE remains statistically significant at the 10% level and is still negatively associated with the persistence of accruals. On the other hand, the within-accounts portion of estimation is no longer statistically significant but the sign of the coefficient is still negative. 4.3 Estimation and the Mapping of Accruals into Cash Flows Prediction 2 (P2) suggests that when greater estimation is needed during the accrual generating process accruals are less likely to be realized as cash. We use the measure of how well accruals map into cash flows developed by Dechow and Dichev [2002] (hereafter DD) to capture this effect. This model captures accruals quality by estimating how well working capital accruals map to into realized operating cash flows. The model is based on the premise that accruals are a way to shift the recognition of cash flows.12 If the realized cash flows of the firm map well into the accruals of the firm then the firm’s accruals are deemed to be of high quality. On the other hand, low-quality accruals consist of estimation errors will be reversed without any cash flow implications. DD operationalize their theory by regressing current period working capital accruals on prior period, current period, and next periods operating cash flows. The standard deviation of the residual from this model is the measure of how well the firm’s accruals map into cash flows. 12 This model does not distinguish between managed earnings or those which arise due to unintentional errors or management uncertainty. 22 The specification of the DD model is shown in equation (5). We include the change in revenues and Property, Plant and Equipment (PPE) in the model as proposed in McNichols [2002]. , β , Δ where , , , are the operating cash flows of the firm. accruals of firm. Δ , (5) , , , is the total working capital is the change in sales from the prior year. , is the total property plant and equipment for the current fiscal period. All continuous variables are scaled by average total assets. We estimate the model by industry and year and use STD(DD Residual), the standard deviation of the residual from the model over the past 5 years for each firm, as our measure of how well the accruals of the firm map into cash flows of the firm. We expect a positive association between STD(DD Residual) and Estimation since accruals with more estimation needed are likely to have lower quality and map less into cash flows. 4.3.1 Estimation and the Mapping of Accruals into Cash Flow Findings Table 9 presents the results for how estimation affects how well accruals map into cash flows (P2). Column 2 of Table 9 shows the regression of the determinants of accruals quality as identified in Francis et al. [2005]. All of the determinants of accruals quality are statistically significant and load in the same direction as found in prior studies (Francis et al. [2005], Dechow and Dichev [2002]). The 3rd Column of Table 9 includes our measure of estimation into the model and we find that estimation is statistically significant at the 5% level and is positively associated with the 23 standard deviation of the Dechow and Dichev residual. This finding is consistent with our hypothesis that the amount of estimation in accruals is associated with greater accrual errors and therefore associated with a lower mapping of cash flows into accruals. The 4th Column of Table 9 presents the results when we decompose estimation into BAE and WAE. Once again we see that both components of estimation are statistically significant at the 5% level and positively associated with a lower mapping of accruals into cash flows. An Ftest of the equality of the coefficients on BAE and WAE yields a p-value of 0.5693 therefore we can’t reject the null that BAE and WAE have the same implications for accruals quality. This result suggests that both components of estimation affect the mapping of accruals into cash flows similarly. Overall our findings suggest that when there is a greater amount of estimation during the accrual generating process there is a lower mapping of accruals into cash flows. 4.4 Estimation and Future Abnormal Returns For our test of P3 we follow the research design of Sloan [1996] and Richardson et al. [2005] to determine whether the market reacts as if it quickly incorporates the estimation information found in the footnotes and CAP disclosures. Sloan [1996] regresses future abnormal returns on total accruals and finds a negative association between the two. Therefore, positive accruals are associated with negative future abnormal returns. On the other hand, negative accruals are associated with positive future abnormal returns. These findings are consistent with his hypothesis that the market over-values the persistence of accruals. We make several small but important modifications to the research design for our study. Since we are interested in the incremental effect of the amount of estimation on the persistence of accruals we include the interaction between the amount of estimation and total accruals into the model. The interaction term models the marginal effect of the amount of estimation on the association between current accruals and future abnormal returns. If the interaction effect is 24 negative then this suggests that the market overvalues more highly estimated positive accruals and vice versa for negative accruals. Next, we make two small changes to the specification of the model to better coincide with our research design. First, Sloan [1996] calculates future abnormal returns beginning four months after the end of the firm’s fiscal period. In contrast, our abnormal returns accumulation begins 5 days after firm files their 10-K form with the SEC. The information about the estimation of accruals used in this study is found in the firm’s 10-K filing. Therefore, we need to ensure that the estimation information found in the footnotes to the financial statements and CAP disclosures is available to the market before we can assess whether the market incorporated the information in a timely manner. Of course, some of the estimation information may have been released prior to the filing of the 10-K but this would only bias results away from our prediction since the market would have had more time to incorporate the information. Second, rather than using a decile ranking of accruals we use the raw amount of accruals. One of the purposes of Sloan [1996] is to show that a trading strategy could be implemented by purchasing stock in firms with extreme low accruals and shorting those with extreme high accruals. The purpose of this study isn’t to implement a trading strategy but rather the provide evidence that the markets appear to not quickly incorporate the estimation information found in the footnotes. Therefore, to preserve more of the information in accruals, we use the raw accruals amount rather than the decile ranking of the amount of accruals. β , , , ∗ , , , , , 25 , , (6) where , is defined earlier. A operating cash flows. S , , is income before extraordinary items minus is the log of the market value of the firms equity. book to market ratio of the firm. , is the firms earnings to price ratio. B beta of the firm for fiscal period t. , , , is the is the market is one-year market-adjusted abnormal returns beginning 5 days after the filings of the 10-K. We also use the Fama-French Carhart four-factor model to further test the association between accruals estimation and future abnormal returns. More specifically, we construct 25 portfolios each month based on the amount of accruals of the firm and the amount of estimation conveyed in the notes to financial statements – 5 rankings of accrual x 5 rankings of estimation. For each portfolio we then estimate the four-factor alpha using the following model. (7) where is the monthly excess return of the value or equal weighted portfolios. is the monthly return of the value-weighted index minus the risk free rate. monthly premium of the book-to-market factor. is the is the monthly premium of the size factor; is the monthly premium on winners minus losers.13 4.4.1 Estimation and Returns Findings The second columns of Table 10 present the baseline results shown in Sloan [1996]. For our sample we find that accruals are negatively associated with the future abnormal returns of the firm. Column 3 of Table 10 shows the results of our test of P3 - the association between future long-term abnormal returns and the amount of estimation information found in the footnotes to 13 RF, HML, SMB, and UMD factors are from Ken French’s website. 26 the financial statements and CAP disclosures. The coefficient on the interaction between estimation and total accruals is negative and statistically significant at the 1% level. Therefore, accruals that require more estimation are more negatively associated with future long-term abnormal returns. This is consistent with the hypothesis that investors are more likely to overvalue firms with less persistent positive accruals and undervalue those with less persistent negative accruals when the amount of estimation is larger. Column 4 presents the results of the model when we decompose estimation into BAE and WAE. We find that the negative association between abnormal returns and the interaction of accruals and estimation is primarily driven by BAE. Table 11 presents the results for our test of P3 using the Carhart four-factor model. Overall we find mixed results as to whether the accrual anomaly is most concentrated in those firms with the greatest amount of accruals estimation. Panel A of Table 11 presents the results for the top and bottom most portfolios in terms of estimation and accruals. We construct hedged returns by going long in those firms with the greatest accruals and short in those with the least amount of accruals. When using equal weighted hedged returns we find evidence that accruals are associated with future abnormal returns in our sample. However, we do not find any association between the future abnormal returns of the firm and accruals when using value weighted hedge returns. We then examine whether the association between future abnormal returns and the current years accruals are greater for those firms with the greatest amount of accruals estimation. We do not find any evidence consistent with the hypothesis that the accrual anomaly is exacerbated in those firms with greater estimation. This result is consistent with the hypothesis that investors utilize the estimation information found in the notes to the financial statements and CAP disclosures when valuing the firm. 27 Lastly, Panel B disaggregates estimation into BAE and WAE. We find no discernible association between future abnormal returns and the amount of estimation and accruals when using value-weighted returns. Using equal-weighted returns we find little difference between the high and low estimation portfolios when using BAE. However, we do find some evidence that those firms with the least WAE are associated with the future abnormal returns of the firm. 5. Robustness Tests 5.1 Placebo Tests Based on Bootstrapping To address the question of whether the results using our linguistic cues approach are “random” we conduct bootstrap tests of our main findings, Prediction 1 (P1) and Prediction 2 (P2). We begin this test by ranking all of the words in all of the notes to the financial statements in our sample by their frequencies. Then, for each of the 40 words in our Estimation Actions, Estimation Objects, and Estimation Adjectives dictionaries that appear at least once in a firm’s notes to the financial statements we select the 5 words which lie above and the 5 words which lie below that word in our frequency list.14 This process yields 10 unique dummy words for each estimation words and a total of 400 unique placebo estimation words which we will refer to as our placebo estimation words dictionary. This placebo dictionary contains words which have a similar frequency of use as our main dictionaries but whose variation of use is likely different from the dictionaries that we used in our study. We begin our simulation by randomly selecting 1 word from the 10 dummy words chosen for each estimation word to arrive at a list of 40 placebo estimation words and count the number of times each of the placebo words is mentioned in a firms footnotes or CAP section; 14 The words anticipating, approximation, approximations, beliefs, believing, estimations, and expecting were never mentioned in a firms footnotes but were included in the original dictionaries for completeness. 28 this count is our placebo estimation count.15 We then estimate the models for the tests of (P1) and (P2) using the placebo estimation count in place of our original count of the number of estimation related linguistic cues. We use this simple word count approach instead of the statistical parser due to computing constraints. This process of selecting random words and retesting our hypotheses is repeated 1,000 times. Table 12 presents our results for the joint tests of significance (insignificance). Specifically, we are interested in what percentage of the bootstrapped tests yields a similar pattern of results as those found using out estimation dictionaries. We see that only 3% of the bootstrap placebo tests yield similar results to our main findings for the persistence of earnings and the quality of accruals, (P1) and (P2) respectively. Specifically, only 3% of the bootstraps yielded a negative and statistically significant result on the coefficient of earnings interacted with the placebo estimation count (P1a), a negative and statistically significant results on the interaction between accruals and the placebo estimation count and insignificant result on the coefficient of the interaction between cash flows and the placebo estimation count (P1b), and a positive and statically significant association between the placebo estimation count at the standard deviation of the Dechow and Dichev residual (P2). This finding suggests that it is unlikely that the results of study are purely random. Next, we see that only 6% of the bootstrapped tests yield similar joint results for (P1a) and (P1b). Lastly, 11% of the bootstraps yield similar joint results to our main findings for (P1b). We also use the number of estimation-related linguistic cues in a firm’s non-critical accounting policy part of the management discussion and analysis in its 10-K filing to replace our Estimation variable in the empirical tests for P1 and P2. Such an estimation count is not directly related to the accruals recognition, and is likely to capture the generic business 15 As an additional robustness check, we randomly select 40 words from the 400 placebo words instead of 1 word from the 10 placebo words chosen for each estimation word. This change in the selection process does not significantly affect the results of our simulation. 29 uncertainty only. Therefore, we do not expect to find similar empirical results using this variable. In untabulated results, we find that indeed this variable is not statistically significantly associated with earnings persistence, accruals persistence, or the Dechow-Dichev accruals quality. 5.2 Controlling for Other Textual Characteristics of 10-Ks Prior studies have found that the persistence of earnings is lower for those firms whose notes to the financial statements are more difficult to read (Li [2008]). It is possible that the details about the accruals that require more estimation are more difficult to convey in writing. Therefore, notes to the financial statements that are more difficult to read may actually convey that greater estimation was needed during the accruals generating process. To address this issue we calculate the Gunning-Fog index score for each of the notes to the financial statements in our sample and include the calculated fog score (and its interaction with earnings, accruals, and cash flows) in each of our main tests. In untabluated results we find that our results are not sensitive to this specification. Li, Lundholm, and Minnis [2012] show that the perceived competition intensity as reflected in firms’ 10-K filings is associated with the mean-reverting speed of abnormal profit. Untabulated results confirm that our results remain essentially the same if the competition measure and its interactions with earnings, accruals, and cash flows are included as additional control variables. 6. Conclusion The primary focus of this study was to examine the association between the estimation needed during the accrual generating process and the persistence of accruals. Sloan [1996] and Richardson et al. [2005] suggest that the estimation needed when recording accruals reduces the persistence of accruals. Their hypotheses are based on the idea that if accruals require a greater 30 degree of estimation they are more likely to be recorded with error (i.e. accruals are less precise). If so, then accruals will be less predictive of future earnings. While this conjecture has been generally accepted in the accounting literature few have explicitly examined the association between the estimation involved during the accrual generating process and the persistence of accruals. This study provides evidence consistent with the conjecture that the estimation needed during the accrual generating process plays a key role in the persistence of accruals. Specifically, we find that when accruals have more estimation they are less predictive of future earnings. We also find that accruals map less into the past, current, or future cash flows of the firm when they require more estimation. Next, we find that both the between-accounts and within-accounts portions of estimations drive our results. Lastly, we find mixed evidence as to whether the markets do not quickly incorporate the estimation information found in firms’ footnotes into their valuation of the firm. In conclusion, the findings in this study provide insight into the accrual generating processing of the managers. More importantly, the findings in our study suggest that understanding the process which managers undergo when recording accruals plays an important role in understanding the persistence and quality of accruals. 31 References Basu, S. 1997. The conservatism principle and the asymmetric timeliness of earnings. Journal of Accounting & Economics 24 (December): 3-37. Billings, M.B., 2011, Discussion of ‘Critical accounting policy disclosures’ by C. Levine and M. Smith, Journal of Accounting, Auditing and Finance. Dechow, P., I. Dichev, 2002, The Quality of Accruals and Earnings: The Role of Accrual Estimation Errors, The Accounting Review 77, 35-59. Dechow, P., W. Ge, 2006, The Persistence of Earnings and Cash Flow and the Roles of Special Items: Implications for the Accrual Anomaly, Review of Accounting Studies, 253-296. Dechow, P., W. Ge, C. Schrand, 2010, Understanding earnings quality: A review of the proxies, their determinants and their consequences. Journal of Accounting and Economics, 50(2-3), 344– 401. Dechow, P. R. Sloan, A. Sweeney, 1995, Detecting Earnings Management, The Accounting Review 70, 193-225. De Franco, G., F. Wong, Y. Zhou, 2011, Accounting Adjustments and the Valuation of Financial Statement Note Information in 10-K Filings, The Accounting Review, Forthcoming. Francis, J., R. LaFond, P. Olsson, K. Schipper, 2005, The Market Pricing of Accruals Quality, Journal of Accounting and Economics 39, 295-327. Hanlon, M., 2005, The Persistence and Pricing of Earnings, Accruals, and Cash Flows When Firms Have Large Book-Tax Differences, The Accounting Review 80, 137-166. Hribar, P., D. Collins, 2002, Errors in Estimating Accruals: Implications for Empirical Research, Journal of Accounting Research 40. Jones, J., 1991, Earnings Management During Import Relief Investigations. Journal of Accounting Research 29, 193–228. Klein, D. and C. Manning, 2003, “Accurate Unlexicalized Parsing” Proceedings of the 41st Meeting of the Association for Computational Linguistics, 423-430. Kraft, A., A. Leone, C. Wasley, 2006, An Analysis of the Theories and Explanations Offered for the Mispricing of Accruals and Accrual Components. Journal of Accounting Research, 44(2), 297–339. Levine, C. B., and Michael J. S. 2011. “Critical Accounting Policy Disclosures.” Journal of Auditing, Accounting and Finance 26 (Winter): 39–75. Li, F., 2008, Annual Report Readability, Current Earnings, and Earnings Persistence, Journal of Accounting and Economics 45, 221-247. 32 Li, F., 2011, Textual Analysis of Corporate Disclosures: A Survey of the Literature, Journal of Accounting Literature, 29, 143-165. Li, F., R. Lundholm, M. Minnis, 2012, A Measure of Competition Based on 10-K Filings, Journal of Accounting Research. Marneffe, M., B. MacCartney, C. Manning, 2006, Generating Typed Dependency Parses from Phrase Structure Parses. LREC 2006. Merkley, K., 2010, More Than Numbers: R&D-related Disclosure and Firm Performance, Working Paper, Cornell University. Mashruwala, C., S. Rajgopal, & T. Shevlin, 2006, Why is the accrual anomaly not arbitraged away? The role of idiosyncratic risk and transaction costs. Journal of Accounting and Economics, 42(1-2), 3–33. McNichols, M., 2002, Discussion of The Quality of Accruals and Earnings: The Role of Accrual Estimation Errors, The Accounting Review 77, 61-69. Radin, A., 2007, Have We Created Financial Statement Disclosure Overload?, The CPA Journal. http://www.nysscpa.org/cpajournal/2007/1107/perspectives/p6.htm Richardson, S., R. Sloan, M. Soliman, I. Tuna, 2005, Accrual Reliability, Earnings Persistence and Stock Prices, Journal of Accounting and Economics 39, 437-485. Riedl, E., S. Srinivasan, 2010, Signaling Firm Performance Through Financial Statement Presentation: An Analysis Using Special Items, Contemporary Accounting Research 27, 289332. Sloan, R. 1996, Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future Earnings?, The Accounting Review 71, 289-315. United States Securities and Exchange Commission, SEC Form 10-K General Instructions, http://www.sec.gov/about/forms/form10-k.pdf. United States Securities and Exchange http://www.sec.gov/about/laws/sea34.pdf. Commission, SEC Act of 1934, White, H., 2010, An Accrual Balance Approach to Assessing Accrual Quality. Working Paper. Xie, H., 2001, The Mispricing of Abnormal Accruals, (3), 357–373. 33 Figure 1: Length of the Footnotes and Estimation Trend This figure shows the yearly trend in the length of the notes to the financial statements, as measured by the total word count, and the average number of estimation related linguistic cues found in the footnotes. 16000 140 14000 120 12000 100 10000 80 8000 60 6000 40 4000 20 2000 0 0 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Total Length Estimation 34 Table 1 : Estimation Trend This table presents the average number of estimation related linguistic cues and the total number of words found in the notes to the financial statements and the critical accounting policies sections for our sample. Our sample spans from fiscal year 1995 to 2010. Year 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Average Estimation 33 35 39 41 43 46 65 83 93 100 104 118 123 126 128 130 82 35 Length 5194 5534 6101 6444 6744 7642 9331 11360 12419 12688 13054 13861 14612 14938 14924 14695 10596 N 2667 4686 4849 4717 4819 4671 4373 4144 3914 3850 3557 3566 3604 3839 3673 3482 4026 Table 2: Estimation by Industry This table shows average number of estimation related linguistic cues and the average number of words in the notes to the financial statements and critical accounting polices section by industry for fiscal periods between 1995-2010. Industry Two Digit SIC Estimation Total Length N Agricultural Production Crops 1 31 5651 16 Metal Mining 10 87 11642 380 Coal Mining 12 132 15852 55 Oil And Gas Extraction 13 89 11391 2445 Mining And Quarrying Of Nonmetallic Minerals, Except Fuels 14 52 6862 77 Building Construction General Contractors And Operative Builders 15 69 9872 423 Heavy Construction Other Than Building Construction Contractors 16 80 10006 154 Construction Special Trade Contractors 17 41 5955 98 Food And Kindred Products 20 69 9269 1504 Textile Mill Products 22 52 7368 308 Apparel And Other Finished Products Made From Fabrics 23 67 9384 682 Lumber And Wood Products, Except Furniture 24 65 7974 315 Furniture And Fixtures 25 63 7951 450 Paper And Allied Products 26 90 10311 660 Printing, Publishing, And Allied Industries 27 71 9035 894 Chemicals And Allied Products 28 82 11518 6255 Petroleum Refining And Related Industries 29 85 11132 410 Rubber And Miscellaneous Plastics Products 30 68 8895 838 Leather And Leather Products 31 48 7794 203 Stone, Clay, Glass, And Concrete Products 32 68 8425 401 Primary Metal Industries 33 74 9578 1024 Metal Products, Except Machinery And Transportation Equipment 34 68 8041 1068 Industrial And Commercial Machinery And Computer Equipment 35 77 9344 4304 Electronic And Other Electrical Equipment And Components 36 84 10051 5723 Transportation Equipment 37 84 9695 1458 Measuring, Analyzing, And Controlling Instruments 38 75 9429 4351 Miscellaneous Manufacturing Industries 39 66 8346 730 36 Table 2: Estimation by Industry (continued) This table shows average number of estimation related linguistic cues and the average number of words in the notes to the financial statements and critical accounting polices section by industry for fiscal periods between 1995-2010. Industry Two Digit SIC Estimation Total Length N Railroad Transportation 40 51 7905 49 Motor Freight Transportation And Warehousing 42 61 7519 540 Water Transportation 44 83 11461 265 Transportation By Air 45 84 10289 479 Transportation Services 47 74 10533 276 Communications 48 93 13085 2467 Electric, Gas, And Sanitary Services 49 108 14284 3279 Wholesale Trade-durable Goods 50 60 8521 1848 Wholesale Trade-non-durable Goods 51 69 10029 1037 Building Materials, Hardware, Garden Supply, And Mobile Home Dealers 52 35 5961 36 General Merchandise Stores 53 71 8465 427 Food Stores 54 67 8320 437 Automotive Dealers And Gasoline Service Stations 55 79 11313 344 Apparel And Accessory Stores 56 67 8839 752 Home Furniture, Furnishings, And Equipment Stores 57 47 6585 290 Eating And Drinking Places 58 67 8896 1235 Miscellaneous Retail 59 64 9419 1500 Hotels, Rooming Houses, Camps, And Other Lodging Places 70 61 8077 264 Personal Services 72 57 8632 143 Business Services 73 82 10794 8680 Automotive Repair, Services, And Parking 75 34 5022 21 Motion Pictures 78 78 11216 401 Amusement And Recreation Services 79 73 11030 856 Health Services 80 82 11110 1378 Educational Services 82 76 10012 291 Social Services 83 41 6949 65 Engineering, Accounting, Research, Management, And Related Services 87 78 10986 1526 Nonclassifiable Establishments 99 66 9805 299 37 Table 3: Summary Statistics This table shows the summary statistics for the sample used in this study. Total Earnings is the firms income before extraordinary items scaled by average total assets. Accruals are total accruals scaled by average total assets. Operating Cash Flows are operating cash flows scaled by average total assets. Estimation is the number of estimation related linguistic cues found in the footnotes section and the critical accounting policies section of the firm’s 10-K. BAE and WAE is estimation broken down into the between accounts estimation and within accounts estimation. BAE and WAE are estimated by industry and year. Length is measured as the total number of words in the footnotes and the CAP section of the firm’s 10-K. Operating Cycle is the log of a operating cycle of the firm. Log(Market Value) the market value of the firm’s equity is calculated as the share price of the firm’s stock at the filing date multiplied by the number of shared outstanding. NEGEARN is the number of years over the past 5 years in which the company had negative earnings. BTM is the book to market ratio. This ratio is calculated as the book value of assets divided by the market value of equity plus liabilities. ETP is the earnings to price ratio calculated as the firms income before extraordinary items divided by price. Stdev(DD Residual) is the standard deviation of the Dechow and Dichev residual over the past 5 year. The DD Residual is estimated by industry and year. Beta is the firm annual beta. Variable Total Earnings Accruals Operating Cash Flows Estimation BAE WAE Length Operating Cycle log(Market Value) NEGEARN BTM ETP Stdev(DD Residual) Beta N 64,411 64,411 64,411 64,411 61,300 61,300 64,411 62,480 60,107 53,542 59,946 60,107 45,433 52,670 Mean -0.068 -0.086 0.018 79 80 0 10327 4.591 5.296 1.720 0.694 -0.250 0.088 0.841 Minimum -1.608 -0.953 -1.018 11 13 -75 1788 1.825 0.026 0 0.068 -9.164 0.008 -5.081 P1 -1.608 -0.953 -1.018 11 13 -75 1788 1.825 0.026 0 0.068 -9.164 0.008 -0.386 P25 -0.082 -0.118 -0.015 36 39 -12 4784 4.183 3.792 0 0.423 -0.931 0.032 0.376 38 Median 0.024 -0.058 0.067 61 66 0 8009 4.660 5.296 1 0.669 0.025 0.057 0.779 P75 0.071 -0.014 0.128 106 111 9 13145 5.094 6.776 3 0.914 0.061 0.105 1.228 P99 0.320 0.298 0.376 285 238 104 46224 6.659 10.599 5 1.919 0.360 0.580 2.634 Maximum 0.320 0.298 0.376 285 238 104 46224 6.659 10.599 5 1.919 0.360 0.580 5.809 Std. Dev. 0.294 0.166 0.217 58 51 28 8064 0.817 2.182 1.820 0.363 1.152 0.095 0.650 Table 4: Univariate Correlations This table presents the Spearman (above diagonal) and Pearson (below diagonal) correlation for the main variables used in this study. Accruals are total accruals scaled by average total assets. Operating Cash Flows are operating cash flows scaled by average total assets. Estimation is the number of estimation related linguistic cues found in the footnotes section of the firm’s 10-K. BAE and WAE is estimation broken down into the between accounts estimation and within accounts estimation. BAE and WAE are estimated by industry and year. Length is measured as the total number of words in the footnotes and CAP section of the firm’s 10-K. Operating Cycle is the log of a operating cycle of the firm. Log(Market Value) the market value of the firm’s equity is calculated as the share price of the firm’s stock at the filing date multiplied by the number of shared outstanding. NEGEARN is the number of years over the past 5 years in which the company had negative earnings. BTM is the book to market ratio. This ratio is calculated as the book value of assets divided by the market value of equity plus liabilities. ETP is the earnings to price ratio calculated as the firms income before extraordinary items divided by price. Stdev(DD Residual) is the standard deviation of the Dechow and Dichev residual over the past 5 year. The DD Residual is estimated by industry and year. Beta is the firm annual beta. Variable 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Total Earnings Accruals Operating Cash Flows Estimation BAE WAE Length Operating Cycle log(Market Value) Negative Earnings BTM ETP Stdev(DD Residual) Beta 1 0.575 0.781 -0.036 -0.008 -0.058 -0.086 -0.024 0.309 -0.596 0.009 0.415 -0.432 -0.025 2 0.366 -0.044 -0.077 -0.062 -0.048 -0.074 0.141 0.108 -0.243 -0.002 0.401 -0.242 -0.026 3 0.663 -0.309 0.015 0.038 -0.035 -0.050 -0.135 0.303 -0.555 0.010 0.204 -0.351 -0.010 4 -0.109 -0.106 -0.013 0.870 0.491 0.737 -0.055 0.379 0.072 -0.021 -0.053 0.049 0.327 5 -0.071 -0.090 0.015 0.893 0.006 0.645 -0.064 0.365 0.043 -0.020 -0.035 0.035 0.337 6 -0.077 -0.044 -0.046 0.345 -0.057 0.359 0.001 0.126 0.070 -0.007 -0.046 0.037 0.068 39 7 -0.149 -0.108 -0.060 0.879 0.801 0.276 -0.058 0.277 0.119 -0.015 -0.073 0.083 0.240 8 -0.011 0.233 -0.199 -0.097 -0.098 -0.011 -0.109 -0.042 -0.002 -0.005 0.032 0.018 0.016 9 0.383 0.064 0.351 0.371 0.354 0.104 0.337 -0.078 -0.381 -0.420 0.253 -0.250 0.346 10 -0.641 -0.211 -0.502 0.103 0.070 0.063 0.150 0.012 -0.405 -0.027 -0.265 0.462 0.080 11 -0.300 -0.013 -0.209 -0.020 -0.025 0.015 -0.023 -0.023 -0.392 0.018 -0.152 -0.139 -0.188 12 0.788 0.388 0.471 -0.110 -0.082 -0.059 -0.137 -0.033 0.237 -0.607 0.044 -0.184 -0.021 13 -0.276 -0.141 -0.228 0.079 0.069 0.019 0.108 0.072 -0.320 0.518 -0.137 -0.331 0.068 14 0.005 -0.031 0.012 0.369 0.365 0.060 0.339 0.012 0.398 0.068 -0.205 -0.087 0.072 - Table 5: Determinants of Estimation This table shows the regression of determinants of estimation on the estimation. P-values are reported in parenthesis below their respective coefficients. Please refer to section 4 of the study for detailed descriptions of each of the variables. All continuous variables are winsorized at 1% and 99% of their respective sample distributions. ***, **, and * denote two-tailed statistical significance at 1%, 5%, and 10% significance levels respectively. log(Market Value) Operating Cycle Stdev(Sales) Stdev(Operating Cash Flows) Negative Earnings Constant Observations Pseudo R-squared Estimation Count (p-value) 8.5600*** (0.000) 0.9758* (0.077) 5.0063*** (0.002) -48.3253*** (0.000) 6.5510*** (0.000) -29.2978*** (0.000) 49,733 0.0607 40 Table 6: Earnings Persistence This table presents the results for our tests of the amount of estimation on the persistence of earnings. The controls include size, operating cycle, standard deviation of sales, standard deviation of operating cash flows, the number of years over the past 5 years in which the firm had negative earnings, and the total length of the footnotes. Please refer to section 4 of the study for detailed descriptions of each of the variables. P-values are reported in parenthesis below their respective coefficients. All continuous variables are winsorized at 1% and 99% of their respective sample distributions. ***, **, and * denote two-tailed statistical significance at 1%, 5%, and 10% significance levels respectively Earnings Estimation Estimation x Earnings Earnings t+1 (p-value) Earnings t+1 (p-value) 0.3024** (0.015) -0.0000 (0.963) -0.0009*** (0.000) 0.3179** (0.016) BAE 0.0000 (0.719) -0.0010*** (0.000) -0.0000 (0.318) -0.0008*** (0.009) BAE x Earnings (1) WAE WAE x Earnings (2) Equivalence of Coefficients (1) = (2) Constant Controls Controls Interact w/ Earnings Fixed Effects Cluster Observations Adjusted R-squared p-value: 0.918 Yes Yes Yes Industry, Year, Auditor Industry, Year 45,727 0.540 41 Yes Yes Yes Industry, Year, Auditor Industry, Year 43,167 0.539 Table 7: Accruals Persistence This table presents the results for our tests of the amount of estimation on the persistence of cash and accruals. Please, refer to section 4 of the study for detailed descriptions of each of the variables. P-values are reported in parenthesis below their respective coefficients. All continuous variables are winsorized at 1% and 99% of their respective sample distributions. ***, **, and * denote twotailed statistical significance at 1%, 5%, and 10% significance levels respectively. Accruals Operating Cash Flows Estimation Estimation x Accruals Estimation x Operating Cash Flows Earnings t+1 (p-value) Earnings t+1 (p-value) Earnings t+1 (p-value) Earnings t+1 (p-value) 0.5367*** 0.5389*** 0.2431*** 0.2539*** (0.000) (0.000) (0.008) (0.007) 0.8867*** 0.8882*** 0.6013*** 0.6168*** (0.000) (0.000) (0.000) (0.000) -0.0000 -0.0001** (0.262) (0.025) -0.0010*** -0.0010*** (0.000) (0.000) 0.0002 0.0000 (0.149) (0.838) BAE BAE x Accruals (a) BAE x Operating Cash Flows -0.0001* (0.955) (0.059) -0.0010*** -0.0011*** (0.002) (0.000) 0.0001 -0.0000 (0.251) (0.853) -0.0001*** -0.0001*** (0.003) (0.008) -0.0011*** -0.0010*** (0.001) (0.006) 0.0002 0.0002 (0.352) (0.249) p-value: 0.8241 p-value: 0.6457 WAE WAE x Accruals (b) WAE x Operating Cash Flows Equivalence of Coefficients (a) = (b) 0.0000 Constant Yes Yes Yes Yes Controls No No Yes Yes Controls Interact w/ Cash earnings - - Yes Yes Controls Interact w/ Accrual earnings - - Yes Yes Industry, Year, Auditor Industry, Year, Auditor Industry, Year, Auditor Industry, Year, Auditor Industry, Year Industry, Year Industry, Year Industry, Year Observations 58,703 55,795 45,727 43,167 Adjusted R-squared 0.604 0.603 0.570 0.569 Fixed Effects Cluster 42 Table 8: Estimation on Persistence with other Measures of Accruals Quality Table 8 Panel A presents the results for our tests of estimation on the persistence of cash flows and accruals including other measure of accruals quality found in the accounting literature. Please, refer to section 4 of the study for detailed descriptions of each of the variables. P-values are reported in parenthesis below their respective coefficients. All continuous variables are winsorized at 1% and 99% of their respective sample distributions. ***, **, and * denote two-tailed statistical significance at 1%, 5%, and 10% significance levels respectively. Panel A: Estimation Accruals Operating Cash Flows Estimation Estimation x Accruals Estimation x Operating Cash Flows Stdev(DD Residual) Stdev(DD Residual) x Accruals Stdev(DD Residual) x Operating Cash Flows Earnings t+1 Earnings t+1 Earnings t+1 Earnings t+1 (p-value) (p-value) (p-value) (p-value) 0.2619*** 0.2837*** 0.3650*** 0.3659*** (0.003) (0.002) (0.001) (0.000) 0.6302*** 0.6526*** 0.7223*** 0.7683*** (0.000) (0.000) (0.000) (0.000) -0.0001 -0.0001** -0.0001*** -0.0001** (0.107) (0.043) (0.006) (0.043) -0.0008*** -0.0009*** -0.0008*** -0.0007*** (0.004) (0.000) (0.002) (0.007) 0.0001 0.0001 0.0001 0.0002 (0.716) (0.686) (0.514) (0.430) -0.0931*** -0.0911*** (0.000) (0.001) -0.5856*** -0.3021*** (0.000) (0.004) 0.1627 0.2064 (0.169) (0.104) Abs(Accruals) Abs(Accruals) x Accruals Abs(Accruals) x Operating Cash Flows -0.0561* -0.0755** (0.090) (0.033) -0.2883*** -0.2184*** (0.000) (0.003) -0.1544*** -0.1029* (0.001) (0.093) Special Items Special Items x Accruals Special Items x Operating Cash Flows 0.0074 0.0131** (0.150) (0.016) -0.1836*** -0.1684*** (0.000) (0.000) -0.1164*** -0.1132*** (0.000) (0.000) Constant Yes Yes Yes Yes Controls No No Yes Yes Controls Interact w/ Cash earnings - - Yes Yes Controls Interact w/ Accrual earnings - - Yes Yes Industry, Year, Auditor Industry, Year, Auditor Industry, Year, Auditor Industry, Year, Auditor Industry, Year Industry, Year Industry, Year Industry, Year Observations 39,943 43,155 43,155 39,943 Adjusted R-squared 0.572 0.571 0.574 0.578 Fixed Effects Cluster 43 Table 8: Estimation on Persistence with other Measures of Accruals Quality Table 8 Panel B presents the results for our tests of estimation on the persistence of cash flows and accruals including other measure of accruals quality found in the accounting literature. Please, refer to section 4 of the study for detailed descriptions of each of the variables. P-values are reported in parenthesis below their respective coefficients. All continuous variables are winsorized at 1% and 99% of their respective sample distributions. ***, **, and * denote two-tailed statistical significance at 1%, 5%, and 10% significance levels respectively. Panel B: BAE and WAE Accruals Operating Cash Flows BAE BAE x Accruals BAE x Operating Cash Flows WAE WAE x Accruals WAE x Operating Cash Flows Stdev(DD Residual) Stdev(DD Residual) x Accruals Stdev(DD Residual) x Operating Cash Flows Earnings t+1 (p-value) 0.2637*** (0.002) 0.6186*** (0.000) -0.0001* (0.083) -0.0008** (0.015) 0.0001 (0.889) -0.0001 (0.263) -0.0007* (0.050) -0.0004 (0.123) -0.0954*** (0.000) -0.6190*** (0.000) 0.1888 (0.105) Earnings t+1 (p-value) 0.2964*** (0.002) 0.6448*** (0.000) -0.0002* (0.057) -0.0011*** (0.001) 0.0000 (0.983) -0.0001* (0.079) -0.0006* (0.092) -0.0001 (0.723) Abs(Accruals) 0.0069 (0.172) -0.1872*** (0.000) -0.1118*** (0.000) Earnings t+1 (p-value) 0.3565*** (0.000) 0.7447*** (0.000) -0.0002** (0.012) -0.0006* (0.073) 0.0003 (0.528) -0.0001 (0.170) -0.0005 (0.172) -0.0002 (0.491) -0.0942*** (0.001) -0.3566*** (0.001) 0.2302* (0.065) -0.0722* (0.050) -0.2031*** (0.007) -0.1050* (0.089) 0.0139** (0.013) -0.1670*** (0.000) -0.1134*** (0.000) Yes Yes Yes Yes Industry, Year, Auditor Industry, Year 43,155 0.574 Yes Yes Yes Yes Industry, Year, Auditor Industry, Year 39,943 0.578 -0.0583* (0.077) -0.2914*** (0.000) -0.1519*** (0.001) Abs(Accruals) x Accruals Abs(Accruals) x Operating Cash Flows Special Items Special Items x Accruals Special Items x Operating Cash Flows Constant Controls Controls Interact w/ Cash earnings Controls Interact w/ Accrual earnings Fixed Effects Cluster Observations Adjusted R-squared Earnings t+1 (p-value) 0.3667*** (0.001) 0.7035*** (0.000) -0.0002*** (0.007) -0.0008** (0.031) 0.0002 (0.698) -0.0001** (0.015) -0.0005 (0.153) 0.0000 (0.938) Yes No Industry, Year, Auditor Industry, Year 39,943 0.572 Yes No Industry, Year, Auditor Industry, Year 43,155 0.571 44 Table 9: Estimation and Accruals Quality This table shows the association between accrual quality, as measure by the Dechow and Dichev 2002 model, and estimation. P-values are reported in parenthesis below their respective coefficients. Refer to section 4 of the study for detailed descriptions of each of the variables. Length is scaled by 100,000. All continuous variables are winsorized at 1% and 99% of their respective sample distributions. ***, **, and * denote two-tailed statistical significance at 1%, 5%, and 10% significance levels respectively. Stdev(DD Residual) (p-value) Estimation Stdev(DD Residual) (p-value) 0.0622** (0.024) BAE (a) 0.0603** (0.039) 0.0801** (0.018) WAE (b) Equivalence of Coefficients (a) = (b) p-value: 0.5693 Length log(Market Value) sum(NEGEARN) Stdev(Sales) Stdev(Operating Cash Flows) Avg(Operating Cycle) Constant Fixed Effects Cluster Observations Adjusted R-squared Stdev(DD Residual) (p-value) 0.0374*** (0.001) -0.0024*** (0.000) 0.0116*** (0.000) 0.0480*** (0.000) 0.4605*** (0.000) 0.0034*** (0.002) 0.0150 (0.741) -0.0016** (0.013) 0.0123*** (0.000) 0.0493*** (0.000) 0.4558*** (0.000) 0.0037*** (0.000) 0.0122 (0.794) 0.0547*** (0.000) -0.0024*** (0.000) 0.0116*** (0.000) 0.0474*** (0.000) 0.4596*** (0.000) 0.0037*** (0.001) 0.0124 (0.788) Industry, Year, Auditor Industry, Year, Auditor Industry, Year, Auditor Industry, Year Industry, Year Industry, Year 41,655 41,655 39,208 0.419 0.421 0.419 45 Table 10: Estimation and Abnormal Returns This table presents the association between one-year abnormal returns beginning 5 days following the filing date and accruals. All regressions were estimated using ordinary least squares regressions. P-values are reported in parenthesis below their respective coefficients. Refer to section 4 of the study for detailed descriptions of each of the variables. Length is scaled by 100,000. All continuous variables are winsorized at 1% and 99% of their respective sample distributions. ***, **, and * denote two-tailed statistical significance at 1%, 5%, and 10% significance levels respectively. Abnormal Returns t+1 Abnormal Returns t+1 Abnormal Returns t+1 (p-value) (p-value) (p-value) Accruals -0.0934*** -0.1385** -0.1133* (0.005) (0.013) (0.075) Estimation -0.0000 (0.905) Estimation x Accruals -0.0024*** (0.001) BAE 0.0001 (0.756) BAE x Accruals -0.0029*** (0.000) WAE 0.0000 (0.987) WAE x Accruals -0.0014 (0.189) Length 0.1078 0.1129 0.0973 (0.201) (0.144) (0.186) Length x Accruals 0.0000*** 0.0000*** 0.0000*** (0.000) (0.000) (0.000) log(Market Value) -0.0147*** -0.0149*** -0.0138*** (0.000) (0.000) (0.000) BTM 0.1830*** 0.1794*** 0.1829*** (0.000) (0.000) (0.000) Beta 0.0333*** 0.0317*** 0.0326*** (0.000) (0.000) (0.000) ETP -0.0669*** -0.0657*** -0.0678*** (0.000) (0.000) (0.000) Constant -0.1643*** -0.1601*** -0.1677*** (0.000) (0.000) (0.000) Fixed Effects Observations Adjusted R-squared Year 51,807 0.066 Year 51,807 0.066 46 Year 49,117 0.067 Table 11: Estimation and Future Abnormal Returns Using Fama-French Carhart Four-Factor Alpha This table presents the results for association between estimation and future abnormal returns using the Carhart Four-Factor Alpha. All regressions were estimated using ordinary least squares. P-values are reported in parenthesis below their respective coefficients. ***, **, and * denote two-tailed statistical significance at 1%, 5%, and 10% significance levels respectively. Panel A: Accruals Estimation All Observations Estimation 1 5 Equal Weighted Returns Value Weighted Returns Accruals Accruals 1 5 Hedge 1 5 Hedge 0.6759** (0.016) 0.0566 (0.594) 0.6193** (0.032) 0.0697 (0.775) -0.2851** (0.049) 0.3548 (0.235) 1 5 Hedge 1 5 Hedge 0.7644*** (0.006) 0.6637* (0.056) 0.0325 (0.888) -0.0405 (0.822) 0.7319*** (0.001) 0.7042** (0.028) 1.0640** (0.001) -0.0942 (0.807) -0.5032* (0.050) 0.0299 (0.876) 1.5672*** (0.000) -0.1240 (0.780) -0.0277 (0.943) Difference -1.6912*** (0.004) Panel B: BAE and WAE All Observations BAE 1 5 Equal Weighted Returns Value Weighted Returns Accruals Accruals 1 5 Hedge 1 5 Hedge 0.6764** (0.017) 0.1067 (0.594) 0.5697** (0.032) 0.0323 (0.894) -0.2426 (0.163) 0.2750 (0.410) 1 5 Hedge 1 5 Hedge 0.7586*** (0.006) 0.6833** (0.039) -0.0805 (0.705) -0.0541 (0.762) 0.8391*** (0.000) 0.7374** (0.020) 0.4587 (0.126) -0.0275 (0.939) -0.3631 (0.118) -0.1932 (0.333) 0.8218** (0.017) 0.1656 (0.682) -0.1017 -(0.270) Difference WAE 1 5 Difference -0.6562 (0.215) 1 5 Hedge 1 5 Hedge 0.6983** (0.031) 0.6458* (0.099) 0.2108 (0.374) 0.1799 (0.384) 0.4875* (0.078) 0.4659 (0.139) 0.1813 (0.652) 0.0716 (0.862) -0.2600 (0.304) -0.3486 (0.247) 0.4413 (0.353) 0.4202 (0.426) -0.0217 (0.959) 47 -0.0210 (0.976) Table 12: Bootstrapped Joint Tests of Prediction 1 and Prediction 2 This table presents the results of our placebo test simulation. For each word in our estimation dictionaries we randomly choose a placebo word from the list of 10 nearest words in terms of frequency. We repeat this procedure 1,000 times and jointly test our main hypothesizes. The simulated tests follow our main hypotheses and are as follows: (P1a): Earnings x Placebo Estimation (-), P1b: Accruals x Placebo Estimation (-) and Operating Cash Flows x Placebo Estimation (I) P2: Placebo Estimation (+). Statistical Significance Level (P1a) and (P1b) and (P2) (P1a) and (P1b) (P1b) 48 1% 5% 10% 3% 6% 11% 6% 9% 13% 8% 11% 15% Appendix 1: Overview of Grammatical Relationships In this study we are interested in determining when estimation was needed by management during the accrual generating process. Since estimation is a function of the actions taken during the accrual generating process and not necessarily of the accruals themselves it is difficult to determine the amount of estimation by simply examining the magnitude of a company’s accruals. For public companies the notes to the financial statements proves a wealth of information about the accrual generating process and, more importantly, information about the estimations needed by management. Unlike the accruals which are denoted quantitatively, the notes to the financial statements, and hence the information pertaining to the estimation involved, is qualitative in nature. More importantly to assess when an estimate was needed by management we need to infer from the words and the placement of the words in the sentence (i.e. the grammatical relationships) when an estimate was needed during the accrual generating process. The words and the grammatical relationships that are used in written language are not random and are in fact highly structured. For example in the sentence “I like football.” The object “football” is the target of “like”. Since the word “like” is conveying the enjoyment of something and “like” is targeting “football” this linguistic cue is conveying that football is enjoyed. More over the association between “I” and “like” denote that the person performing the action is “I”. Even though this example may be simple it illustrates a powerful idea and provides us with a structure to help us infer meaning from the qualitative footnotes. To infer meaning from the footnotes we therefore need the grammatical relationships in each sentence in the notes to the financial statements and several dictionaries of terms associated with estimation (Appendix 2). To get the grammatical relationships in each sentence we use a technique pioneered in the field of Natural Language Processing called Statistical Parsing to map the structure of the sentences in the notes to the financial statements. We use a specific implementation of statistical parsing from the Stanford Natural Language Processing Group – see http://nlp.stanford.edu/software/index.shtml 49 for details (Marneffe et al [2006]). Essentially, this implementation finds the most likely map of the sentence by matching the sentence to a tree bank of manually parsed sentences to find the layout of the sentence which is most likely. To illustrate this technique, the following sentence was parsed using the Stanford parser. “We estimated receivables and purchased inventory.” [nsubj(estimated-2, We-1), nsubj(purchased-5, We-1), dobj(estimated-2, receivables-3), conj_and(estimated-2, purchased-5), dobj(purchased-5, inventory-6)] We see that the object “receivables” is the direct object of the action (verb) “estimated”. This linguistic cue indicates that the sentence is conveying that receivables were estimated. As illustrated, using the grammatical relationships removes any ambiguity about the meaning of the sentence. 50 Appendix 2: Estimation Dictionaries and Grammatical Relationships Estimation Dictionaries We construct four dictionaries to help measure the estimation conveyed in each companies notes to the financial statements: Estimation Actions Estimate, Estimating, Estimated, Anticipate, Anticipates, Anticipating, Anticipated, Approximate, Approximates, Approximated, Approximating, Assess, Assesses, Assessed, Assessing, Believe, Believed, Believes, Believing, Determine, Determined, Determining, Determines, Evaluate, Evaluated, Evaluating, Evaluates, Expect, Expects, Expected, Expecting, Forecast, Forecasts, Forecasted, Forecasting Estimation Objects Estimate, Estimates, Estimation, Estimations, Approximation, Approximations, Assumption, Assumptions, Belief, Beliefs, Forecast, Forecasts Estimation Adjectives Estimated, Anticipated, Approximately, Expected, Forecasted, Likely, Probable Use Words Make, Makes, Made, Making, Use, Uses, Used, Using, Include, Includes, Included, Including Estimation Actions are verbs which convey that an estimation action was performed (e.g. “we estimated accruals”). Estimation Objects are estimation related objects/nouns (e.g. “we used estimates”). Estimation Adjectives modify an object to convey that the object was estimated (e.g. “estimated accruals”). Lastly, the Use Words dictionary contains action words which convey that something was used or done by management. 51 Grammatical Relationships We use the following grammatical relationships in conjunction with the Estimation Dictionaries above to find the linguistic cues that infer that estimation was needed. Direct Object The accusative object of an action (i.e. “estimate receivables” or “used estimates”). For this grammatical relationship we look for when a word from the Estimation Action dictionary targets some object or when a word from the Use Word dictionary targets a word from the Estimation Object dictionary. An example of the first case is “estimate receivables” here the action “estimate” targets the object “receivables” thereby implying that receivables were estimated. An example of the second scenario is “used estimates”. In this example one of the words from the Use Words dictionary, “used”, targets a word from the Estimation Object dictionary, “estimates”, which implies that they used an estimate. (Passive) Nominal Subject This grammatical relationship is similar to the direct object in that it relates information about an object. This grammatical association is similar to the Direct Object and the relationships that we look for are the same as those for the direct object. Adjective Modifier Adjective Modifies modify the meaning of an object (i.e. “likely receivable” or “anticipated value”). For this grammatical relationship we find when a word from the Estimation Adjective dictionary targets some object – this implies that the object was estimated. Quantifier Phrase Modifier This grammatical relationship is a modifier to a number (i.e. “approximately $100”). This grammatical association is similar to the adjective modifier but specific to numbers. We look for when a word from the Estimation Adjectives dictionary targets a number. 52 Noun Compound Subjects A noun used to modify another noun (i.e. “value estimates”). This grammatical relationship is similar to the adjective modifier except that a word from the Estimation Object dictionary is modifying another object. 53 Appendix 3: Top 10 and Bottom 10 Accrual Accounts by Estimation This table presents the accrual accounts with the highest and lowest coefficients on the accrual account dummies from the regression of estimation on accrual account flags. The regression was performed using OLS with an intercept as described in equation (1). Top 10 Accrual Accounts Fair Value Intangibles Derivatives Regulatory Restatements Restructuring Discontinued Operations Segments Accounting Changes Contingencies Bottom 10 Accrual Accounts PP&E Credit Arrangements Long-Term Debt Leases Significant Customers Inventory Accounting Policies Reinsurance Taxes Home Loan 54