Relevance of Goodwill Impairments to Cash Flow Prediction and Forecasting

Abstract

This study examines the contribution of goodwill impairment information to the prediction and forecasting of future operating cash flows. Extending the framework of Barth, Cram, and Nelson, we find that explicitly including goodwill impairments incrementally improves 1-year-ahead cash flow prediction and forecasting. Improved cash flow forecasting is present over the entire 2001-2009 study period as well as for each year within the study window. In addition, goodwill impairments retain their significance and predictive power when other non-recurring charges (e.g., restructuring, asset write-downs, and merger and acquisition costs) are added to the model, both individually and aggregately, and when market-related information (i.e., change in market capitalization) is included in the model. While these findings are validated by in-sample prediction techniques, this study is also one of only a few studies to investigate the incremental, out-of-sample predictive power of non-current accruals on reported (as opposed to computed) operating cash flows. Analysts, investors, creditors, and others interested in future cash flows should separately consider goodwill impairment information, when available, to improve the accuracy of cash flow prediction and forecasting.

Keywords

goodwill impairments cash flow forecasting cash flow prediction non-recurring charges goodwill

Introduction

The accounting profession has long recognized that cash flow prediction is one of the fundamental uses of financial information, and this construct provided much of the rationale for the passage of Statement of Financial Accounting Standard No. 142, Goodwill and Other Intangible Assets (FASB; 2001; SFAS No. 142).¹ However, goodwill information is often ignored in cash flow prediction models. Writing contemporaneously with the passage of SFAS No. 142; FASB; 2001, Barth, Cram, and Nelson (2001) find that disaggregated earnings information allows for more accurate cash flow prediction than the use of aggregated earnings data. Thus, we expect that separate inclusion of goodwill impairment information will incrementally improve future cash flows prediction.

We test our hypothesis by first using the cash flow prediction model developed by Barth et al. (2001) and replicating their sample within their original study window (1987-1996). Then, we extend their model by adding a separate term to capture the information content of goodwill impairments and compare the original and expanded models in the post-SFAS 142 time period (2001-2009). Comparing the original and expanded models, we examine the incremental improvement in cash flows prediction and forecasting by separately including goodwill impairment information by testing not only for the statistical association between goodwill impairments and cash flows (to establish comparability with prior studies) but also for the ability of goodwill impairments to improve cash flows forecasting. For comparability with other studies, we also compare both the original Barth et al. (2001) model and our expanded model against a cash-only model. Improvement in forecasting accuracy is assessed via two independent, out-of-sample prediction measures: predicted residual sum of squares (PRESS) and mean squared prediction error (MSPE). As such, this study is also one of only a few studies to investigate the incremental, out-of-sample predictive power of non-current accruals on reported (as opposed to computed) operating cash flows (see Lev, Li, & Sougiannis, 2010; Nam, Brochet, & Ronen, 2012).

We also determine the stability of the relationship between goodwill impairments and future cash flows by including other non-recurring charges in the expanded model. Restructuring charges, asset write-downs, and merger and acquisition costs are separately entered into the expanded model along with goodwill impairments, and the results are compared with the expanded model including only goodwill impairments. In addition, we consider the total effect of including all non-recurring charges in two ways. First, we examine a single variable incorporating the effects of all non-recurring items (including goodwill impairments). Second, we parse the non-recurring items into two variables: one for goodwill impairments and another for all non-recurring items other than goodwill impairments. Furthermore, because it is possible that the information contributed by goodwill impairments may be a reflection of broader economic issues already communicated via market-related information, we test the robustness of our model after including a variable to capture the change in market capitalization for each firm year.

Although goodwill impairments are non-cash events that occur infrequently, our findings indicate that they have individual, incremental relevance for the prediction and forecasting of future cash flows. Our findings are robust even when other types of non-recurring charges are included in the model, both individually and aggregately, and when market-related information is included in the model. Based on these results, we conclude that goodwill impairment data provide unique information relevant to cash flow prediction and forecasting. Therefore, analysts, investors, creditors, and others interested in predicting future cash flows should separately consider goodwill impairment data in their cash flow prediction and forecasting models.

The remainder of this article proceeds as follows: The “Background and Prior Research” section reviews prior research, the “Data and Method” section presents the data and methodology used in this study, the “Results and Analysis” section summarizes and analyzes our findings, and the final section provides concluding remarks.

Background and Prior Research

The Financial Accounting Standards Board (FASB; 1978) has long recognized that one of the primary uses of financial reporting is to help users assess the “amounts, timing, and uncertainties of expected cash flows” of the entity (para. 25). This focus on cash flows is highlighted throughout Statement of Financial Accounting Concept Number 1 (FASB; 1978; SFAC No. 1),² and it is retained in SFAC No. 8 (FASB; 2010) which replaced SFAC No. 1 in 2010.³ Specifically, SFAC No. 8 (FASB; 2010) states that “expectations about returns depend on [users’] assessment of the amount, timing, and uncertainty of (the prospects for) future net cash inflows to the entity” (para. OB3, parenthetical expression in original). SFAC No. 8 (FASB; 2010) also states that “[i]nformation about a reporting entity’s cash flows during a period also helps users to assess the entity’s ability to generate future net cash inflows” (para. OB20). This conceptual focus on cash flows is also reflected in the accounting standards. The Accounting Standards Codification (ASC) explicitly supports the conceptual focus on cash flows, affirming that the primary objective of the cash flows statement is to provide information upon which future cash flows can be predicted (FASB; 2015; ASC 230-10-10-1). In addition, the cash flows statement itself has been required since the adoption of Statement of Accounting Standard No. 95 (FASB; 1987; SFAS No. 95) in 1987.⁴

After the adoption of SFAS No. 95 (FASB;1987), cash flows enjoyed only limited interest among researchers until the past few years. In their “comprehensive review of academic research related to direct method cash flow presentation,” Hales and Orpurt (2013, p. 539) cite only 36 studies, and 9 (25%) of these studies were published within the last 4 years of their 26-year study period (1988-2013). Lorek and Fimia-Moe (2012) find even less scholarly interest in the study of statistically based annual cash flow prediction models. Their review spans from 1986 to 2009 (24 years), and they cite only 14 studies during their research window. Many of these studies were reported by Lorek and his colleagues (Lorek, Schaefer, & Willinger, 1993; Lorek & Willinger, 1996, 2008, 2009, 2010a, 2010b, 2011), and their studies focus primarily on the comparison of various statistical estimation techniques and/or the relative accuracy of using annual or quarterly data. These recent reviews indicate that there has been relatively little scholarly attention to cash flows since the adoption of SFAS No. 95 (FASB; 1987), and that the minimal research interest that has been shown to cash flows has focused more on the presentation of the information rather than on the use of the information to predict cash flows. This apparent deficiency exists despite the fact that FASB initially identified (SFAC No. 1) and continues to emphasize (SFAC No. 8) that cash flow prediction is one of the primary purposes of financial information.

Before the adoption of SFAS No. 95 (FASB; 1987), there was no definitive measure of cash flows from operations (CFO), so researchers developed various proxies for CFO: some to predict CFO itself, but most to serve as variables in the prediction of future earnings (e.g., Ball & Brown, 1968; Beaver & Dukes, 1972; Bowen, Burgstahler, & Daley, 1986, 1987; Greenberg, Johnson, & Ramesh, 1986; Rayburn, 1986; Wilson, 1986, 1987). Operating cash flows were favored for earnings prediction models because they were considered to represent the core or recurring cash flows of the business and were, therefore, more regular and amenable to prediction. After the adoption of SFAS No. 95 (FASB; 1987), research interest in cash flows slowed, and further research regarding the best proxy for CFO became moot. Although some studies continue to use cash flow proxies (e.g., Dechow, Kothari, & Watts, 1998; Finger, 1994; Kim & Kross, 2005),⁵ most recent studies use CFO as reported in the statement of cash flows (e.g., Cheng & Hollie, 2008; Krishnan & Largay, 2000; Lev et al., 2010; Yoder, 2007).

A landmark study on both the progression from the use of cash flows proxies to CFO and the evolution of comprehensive cash flows prediction models is Barth et al.’s (2001) comprehensive model of cash flows prediction. The Barth, Cram, and Nelson model (hereafter, the BCN model) has been cited, examined, and extended by numerous subsequent studies (e.g., Al-Debi’e, 2011; Bandyopadhyay, Chen, Huang, & Jha, 2010; Cheng & Hollie, 2008; Farshadfar, Chew, & Brimble, 2008; Francis, 2011; Habib, 2010; Hales & Orpurt, 2013; Kim & Kross, 2005; Lev et al., 2010; Lorek & Willinger, 2008, 2009, 2010a, 2010b, 2011; Luo, 2008; Orpurt & Zang, 2009; Telmoudi, Noubbigh, & Ziadi, 2010; Telmoudi, Ziadi, & Hedi, 2010). At the time when Barth et al. completed their study, goodwill was still amortized over a period not to exceed 40 years. The adoption of SFAS No. 142 in 2001 prohibited the systemic amortization of goodwill and required companies, instead, to test goodwill annually for impairment. This shift from systematic amortization to annual impairment was intended to more closely correlate goodwill write-offs with future cash flows. Thus, goodwill impairments should be expected to provide useful information in the prediction of future cash flows. Although goodwill impairments could have been recognized prior to 2001, they were rare; thus, goodwill impairments would not have been prevalent at the time of the original BCN study, and they were not separately included in the BCN model.

Goodwill is “an asset representing the future economic benefits arising from other assets acquired in a business combination . . . that are not individually identified and separately recognized” (FASB; 2015; ASC 350 Glossary, emphasis added). Thus, a market-validated value for goodwill can only be determined at the time of a business combination, and this value is measured as the excess of the purchase price of the assets acquired (both tangible and intangible) over the net fair value (FV) of the assets and liabilities acquired as of the date of the acquisition (FASB; 2015; ASC 805-30-30-1). Upon recognition, goodwill is assigned to a reporting unit (RU), which typically consists of a majority of those assets purchased in the transaction that gave rise to the initial recording of the goodwill amount.

The method by which goodwill is tested for impairment begins with a qualitative assessment of the likelihood that the value of reported goodwill is impaired (see flowchart and related notes at FASB; 2015; ASC 350-20-55-25). If it is determined that impairment is likely, the impairment calculations follow a two-step process.⁶ First, the FV of the RU is compared with the carrying value of the RU including the recorded amount of goodwill. Although the ASC requires no specific method for determining the FV of the RU, the example provided within the ASC (FASB; 2015; ASC 350-20-55-24) uses the expected present value of future cash flows to compute the FV of the RU; and actual financial statement footnotes often indicate that firms use the present value of anticipated future cash flows (either by itself or as a significant component within a more complex valuation model) to compute the FV of RUs for the purpose of testing for goodwill impairments. If the FV of the RU is less than its carrying value, a goodwill impairment may exist, and the process continues to the second step of the impairment testing model. In this step, the value of goodwill must be computed by implication as it can only be measured in actuality at the time of a business combination. The implied value of goodwill is computed as the difference between the FV of the RU and the FV of the net identifiable assets of the RU without considering the recorded value of goodwill. The implied value of goodwill is then compared with the recorded value of goodwill. If the implied goodwill is more than the recorded goodwill, no goodwill impairment exists; but if the implied goodwill is less than the recorded goodwill, a goodwill impairment exists to the extent that the recorded goodwill exceeds the value of implied goodwill.

Goodwill impairments are thus connected to future cash flows both economically and methodologically. Economically, goodwill is generated because the purchaser in a business combination expects future cash flows to be larger than would be otherwise indicated by the underlying net assets themselves. If the anticipated level of cash flows declines (or never materializes), then either the assets themselves or the expected synergies (monetized as “goodwill”) have failed to perform as expected. As the FV of assets is set by the market, the goodwill portion of the transaction is the most likely reason for the decline. Goodwill impairments are also closely tied to anticipated cash flows because anticipated cash flows are a significant (if not the primary) factor in computing the FV of the RU. Thus, if anticipated cash flows decline, the FV of the RU declines. If the FV of the RU declines more quickly than the carrying value of the RU (which also declines via systematic reductions, such as depreciation), then the relative difference between the FV of the RU and the carrying value of the RU decreases. If this relative difference declines enough, a goodwill impairment will result. Thus, both a logical economic connection and a methodological connection exist between goodwill impairments and future cash flows.

The purpose of this study is to examine the cash relevance of goodwill impairments by extending the original BCN model to identify the unique information content of goodwill impairments when predicting future cash flows. We begin by validating the BCN model in both the pre- and post-SFAS No. 142 periods to provide a benchmark for our analyses. We then extend the original BCN framework by adding a term for goodwill impairments, and we anticipate that goodwill impairments will have a significant, inverse relationship to future cash flows. To enhance comparability with other studies, we also compare both the original and extended models with a cash-only model. The improvement in cash flows forecasting accuracy among the models is examined using both the predicted residual sum of squares (PRESS; see Montgomery, Peck, & Vining, 2012) and the MSPE (Clark & McCracken, 2001; Clark & West, 2007; West, 2006) techniques. We also analyze the robustness of our findings by considering the effects of including other non-recurring charges and market-related information in our expanded model.

Data and Method

In their seminal piece, Barth et al. (2001) demonstrate that parsing annual accrual earnings information into specific components improves the prediction of cash flows compared with approaches using lagged, single-item cash flows and/or earnings values to predict future cash flows. Given this result along with the logical and methodological connections between goodwill impairments and future cash flows, we hypothesize that additional parsing of the data to include a separate term for goodwill impairments may further improve the ability to predict future cash flows; and we extend the Barth et al. model by adding a separate term to capture the accounting information from goodwill impairments. Nevertheless, to maintain comparability with prior studies, we begin our analysis with a cash-only prediction model, as follows:

C F_{i, t + 1} = α + β C F_{i, t} + ε_{i, t} .

CF is cash flow from operating activities (OANCF) minus the cash flow effects of extraordinary items and discontinued operations (XIDOC).⁷

Our next step is to replicate the BCN model in its original time period and with the same sample restrictions used by Barth et al. (2001). Replicating the original BCN sample is relevant to the present study in two ways. First, this step ensures that our BCN-type model used to analyze post-SFAS 142 data is comparable with the BCN model used to analyze pre-SFAS 142 data. Second, replicating the original BCN model allows for direct observation of the incremental improvement by including goodwill impairments in the cash flow prediction model. The original BCN model⁸ is as follows:

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{O} OTHE R_{i, t} + ε_{i, t}, \end{matrix}

with OTHER defined as follows:

OTHER = EARN - (CF + Δ AR + Δ INV - Δ AP - DEPR - AMORT) .

Specifically, for data construction, CF is the cash flow from operating activities (OANCF) minus the cash flow effects of extraordinary items and discontinued operations (XIDOC). ΔAR is the change in accounts receivable (RECCH), if available; otherwise to be consistent with the sign of RECCH, ΔAR is calculated using prior year accounts receivable less present year accounts receivable (RECT). ΔINV is the change in inventory (INVCH), if available; otherwise to be consistent with the sign of INVCH, ΔINV is calculated as prior year total inventory less current year total inventory (INVT). ΔAP is the change in accounts payable (APALCH), if available; otherwise, ΔAP is calculated by subtracting the sum of the prior year’s balances for accounts payable (AP) and accrued expenses (XACC) from the present year’s balances for these same accounts. DEPR is computed using depreciation expense (XDP), if available; otherwise DEPR is computed using depreciation and amortization (DP) minus amortization (AM). If DP is available and AM is unavailable, DEPR is DP. AMORT is equal to income statement amortization (AM), if available; otherwise, AMORT is equal to depreciation and amortization (DP) minus depreciation expense (XDP) minus depletion expense (XDEPL).⁹EARN is earnings before extraordinary items (IB). As in Barth et al. (2001), all measures are deflated by the average book value of total assets (AT), computed as the average of present year and prior year total assets.

To verify the accuracy of our construction of the original BCN sample, all BCN sampling restrictions are retained. Data are taken from the Compustat annual file and include only domestic firms for the years 1987 to 1996. All observations with Standard Industrial Classification (SIC) codes from 6000 to 6999 are removed, as are firm-year observations with sales less than US$10 million (SALE) in the Compustat fiscal year or a stock price <US$1 (PRCC_F) on the Compustat firm-year observation date. In addition, for a firm-year observation to remain in the sample EARN must be calculable not only in year t, but also in years t− 1 and t− 2. Finally, for consistency with Barth et al. (2001), observations in the highest and lowest 1% of fiscal-year observations of EARN and CF are deleted, as are observations with studentized residuals greater than 3 in absolute value after initial regressions predicting next fiscal year’s CF are estimated.¹⁰

After confirming that our model reliably approximates the results reported by Barth et al. (2001), we will move to the second phase of our analysis and expand the model, retaining limitations previously employed, into the post-SFAS 142 time period (2001-2009) and add a term to the model to capture goodwill impairments. In prior studies relevant to our current research, Jarva (2009) examines the presence of a relationship between goodwill impairments and future cash flows using the BCN model, but the incremental improvement in the predictive ability of the model is not explored. Lee (2011) and Lee and Yoon (2012) examine the relationship between goodwill impairments and future cash flows in the pre- and post-SFAS 142 periods, but their study focuses on the value of goodwill and goodwill write-offs in isolation, not as part of a unified cash flow prediction model. Furthermore, the literature is silent regarding the incremental contribution of goodwill impairments in the out-of-sample forecasting of future cash flows (rather than in-sample associations). Due to our replication of, and comparison with, the original BCN model, our study is able to measure not only the relationship between goodwill impairments and future cash flows, but also the incremental improvement in the forecasting ability of the cash flow model. Unlike Jarva (2009), we do not limit our sample to firms with goodwill write-offs greater than or equal to 1% of beginning-of-the-year total assets or US$10 million (whichever is larger), a limitation which Jarva (2009) indicates was employed to increase the power of his tests. As our study includes all firm years (2001-2009) regardless of whether or not a goodwill impairment is reported (and if reported, regardless of the relative size of the impairment), we would expect our results to be more difficult to find, but, if found, to be more generalizable. By including all firm years and also covering a longer period of time, 2001-2009 versus 2002-2006, we are able to include more firm years in our study. Unlike Jarva (2009), we do not separately examine the effect of firm size as size is incorporated into our model (as in the original BCN model) by scaling all variables by total assets. Thus, our revised model is as follows:

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{GDWLIP} GDWLI P_{i, t} + β_{O} OTHE R_{i, t} + ε_{i, t} . \end{matrix}

GDWLIP, defined per Compustat documentation as pretax goodwill impairments, is introduced into the original BCN model. All other variables are defined as in Barth et al.’s (2001) formula; and all measures, including GDWLIP, continue to be deflated by average total assets. Under this specification, the calculation of OTHER is also altered thusly:

OTHER = EARN - (CF + Δ AR + Δ INV - Δ AP - DEPR - AMORT - GDWLIP) .

Note that because goodwill impairments were not widely recorded before the adoption of SFAS 142 in 2001, the GDWLIP term would be 0 for the vast majority of observations during the 1986-1997 time period originally studied by Barth et al. (2001). Thus, this expanded model would report nearly identical results to those provided by the BCN model during their original observation window. The reader will further note that Equations 2 and 4 represent a type of “nested” model by which a new potential predictor (GDWLIP) may be evaluated after adding it to the framework of the original model.

Our period of study is for fiscal years after the adoption of SFS 142: 2001-2009. Data from 2010 are also necessary as 1-year-ahead cash flows are the dependent variables for the regressors of the final sample prediction year 2009. For our study, observations must have non-missing data for all variables, excluding disclosure of goodwill impairments on the income statement. Observations with “missing” Compustat data for GDWLIP are assigned values of 0.¹¹

The situation we encounter for analysis compares nested models. The cash-only model is a special subcase of the original BCN model without GDWLIP (Equation 2) which is, itself, a special subcase of our proposed model which includes GDWLIP (Equation 4). Our statistical testing to gauge the contribution of GDWLIP to the modeling of future cash flows proceeds as follows. We begin with an ordinary least squares (OLS) framework which estimates regression coefficients for the predictors of Equations 1, 2, and 4. In doing so, we may gain an initial insight as to whether GDWLIP aids in predicting estimates of future cash flows. To test whether Equation 4 is indeed significantly statistically superior to Equations 1 and/or 2 in terms of explaining associations with future cash flows, we conduct a F test which compares the residual sum of squares (RSS) from each framework. We do this for the full sample and on an annual basis. These in-sample procedures allow for conclusions to be drawn regarding the impact of GDWLIP’s inclusion in the model framework;¹² however, to determine whether or not GDWLIP significantly aids in the forecasting ability of future cash flows, we must consider out-of-sample procedures and ascertain whether predictive ability still improves when GDWLIP enters our modeling. In doing so, we provide new findings to the literature.

As one criterion, we compute the predicted residual sum of squares (PRESS) statistic (see Montgomery et al., 2012) and judge whether the error reflected in the statistic declines upon GDWLIP’s inclusion compared with both the cash-only model (Equation 1) and the original BCN model (Equation 2). PRESS is a cross-validation tool which removes each observation from the regression scheme, one at a time, and compares each actual future cash flow, in turn, from the now “out-of-sample” observation with the predicted value garnered from the procedure. Any improvement in forecasting ability by including GDWLIP is evaluated based on comparing the PRESS statistic between the expanded model with GDWLIP (Equation 4) to both the cash-only model (Equation 1) and the original BCN model (Equation 2). Lower PRESS statistics are indicative of improved forecasting ability. We consider such evaluations for the overall sample and on an annual basis.

While a decreased PRESS statistic is indicative of improved forecasting ability, there is no indication of the degree of improvement by comparing PRESS values. Therefore, a formal statistical test of the improvement in forecasting accuracy is desirable. For this more refined evaluation of the improvement to forecasting accuracy, we consider the testing procedures developed by Clark and McCracken (2001), West (2006), and Clark and West (2007). These benchmark articles provide the framework of refined test statistics and criteria for judging the contribution of proposed new regressors for forecasting in the nested framework.

The procedure begins by calculating the MSPE from each of the nested models. As suggested by Clark and West (2007), we adopt a “rolling” framework to estimate Equations 1, 2, and 4, each using 50% of the data points to construct coefficient estimates (for the full sample and for each year, as shown in Table 3), while the remaining 50% of the data points from each sample are used to evaluate the accuracy of the model.¹³ The rolling procedure then sequentially swaps one “prediction” observation at a time with one “evaluation” observation.¹⁴ After doing so, the compilation of the squared residuals from each proposed model yields a MSPE statistic. While it would be intuitively pleasing to observe the simple difference in the MSPE statistics from the nested models and compare the MSPE values directly, Clark and West (2007) note that this difference is conservatively biased, so that the parsimonious model (i.e., Equations 1 and 2 in this study) will actually exhibit a lower MSPE under the null hypothesis of no forecasting ability added with the inclusion of the additional measure (i.e., GDWLIP in this study). Clark and West (2007) thus develop a correction to the difference of MSPEs (known as MSPE-adjusted) based on the residuals of the parsimonious model.¹⁵ They also establish the following test statistic for comparing models based on forecast errors from the predictions of the parsimonious model and the expanded model.¹⁶

{\hat{f}}_{t + 1} = {\hat{e}}_{1 t + 1}^{2} - [{\hat{e}}_{2 t + 1}^{2} - {({\hat{y}}_{1 t + 1} - {\hat{y}}_{2 t + 1})}^{2}],

where e_1t+1 and e_2t+1 are the one step ahead forecast errors (based, in our article, on the rolling framework) and y_1t+1 and y_2t+1 are the forecasted values from the parsimonious model (in our article, the cash-only model and the original BCN model) and the expanded model (in our article, the BCN model with the added GDWLIP term), respectively. Clark and West (2007) define the sample average of ${\hat{f}}_{t + 1}$ as $\bar{f}$ , which is computed as follows:

\bar{f} = P^{- 1} \sum_{t = R}^{T} {\hat{f}}_{t + 1} .

Clark and West (2007) identify P as the number of observations used to check prediction accuracy, while R is the number of observations used to form the estimates. Clark and West, following the work of Clark and McCracken (2001), detail how the critical values discussed below are valid for ratios of P / R < 20. Furthermore, the number of additional regressors beyond the parsimonious model should be capped at 20. Given the 50%/50% breakdown of P and R used in the present study, P/R = 1, which is well within the tolerances indicated above. The test statistic is thus constructed as

\frac{\sqrt{P} \bar{f}}{{[sample variance of {\hat{f}}_{t + 1} - \bar{f}]}^{1 / 2}} .

Clark and West (2007) also develop practical critical values, via simulation, for the application of statistical inference based on the MSPE difference of models, including their aforementioned adjustment. A critical value of 1.282 (1.645) is established as the 10% (5%) critical value for evaluating the test statistic.¹⁷ Clark and West (2007) note that they do not provide a formal proof of the levels of these critical values. The test statistic based on the adjusted-MSPE difference is not strictly asymptotically normal; however, the simulation evidence strongly suggests that traditional one-sided critical values for t tests are appropriate for inference based on the MSPE-adjusted test statistic as long as P/R < 20, as noted above. Thus, we also expand beyond Clark and West’s (2007) formal discussion and consider the traditional, one-sided 1% significance critical value to be 2.33.

Beyond measuring the improvement in cash flow prediction and forecasting, we also examine the relevance of goodwill impairments with respect to future cash flows when considered along with information provided by other non-recurring charges. Although Jarva (2009) does consider the effects of goodwill impairments with respect to restructuring charges, we choose not to follow Jarva’s use of dummy coding to capture restructuring charges (“1” if present, “0” if absent). We feel that the introduction of dummy-coded data into our models would add unnecessary noise to and would reduce the precision of the results from the otherwise continuous variables.

Instead, we take a three-pronged approach to assessing the reliability of goodwill impairments when other non-recurring charges are considered. First, we include each of the following non-recurring costs, one at a time, along with goodwill impairments in the expanded model (Equation 4): restructuring costs (RCP), asset write-downs (WDP), and merger and acquisition costs (AQP). The resulting equations are as follows:

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{GDWLIP} GDWLI P_{i, t} + β_{RES} RESTRUC T_{i, t} + β_{O} OTHE R_{i, t} + ε_{i, t} . \end{matrix}

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{GDWLIP} GDWLI P_{i, t} + β_{WD} WRITEDOW N_{i, t} + β_{O} OTHE R_{i, t} + ε_{i, t} . \end{matrix}

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{GDWLIP} GDWLI P_{i, t} + β_{AC} {ACQUISITION}_{i, t} + β_{O} OTHE R_{i, t} + ε_{i, t} . \end{matrix}

The OTHER term is adjusted for each model in the same manner as previously described, so that all terms specified in the model are removed from the OTHER term. The regression results for each model are then compared with the expanded model including GDWLIP (Equation 4) by examining the change in adjusted R², the PRESS statistic, and the adjusted-MSPE statistic.

Second, we replace GDWLIP with Compustat’s variable for all special items (SPI)¹⁸ which, itself, includes GDWLIP, resulting in the following equation:

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{SPI} SP I_{i, t} + β_{O} OTHE R_{i, t} + ε_{i, t} . \end{matrix}

Once again, the OTHER term is adjusted, so that no information is included more than once in the equation. The predictive relevance of the SPI model (Equation 12) is evaluated against both the original BCN model (Equation 2) and the expanded model including GDWLIP (Equation 4).

Finally, we include the SPI variable, excluding GDWLIP, along with GDWLIP, and the following equation results:

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{GDWLIP} GDWLI P_{i, t} + β_{SPI - GDWLIP} {(SPI - GDWLIP)}_{i, t} \\ + β_{O} OTHE R_{i, t} + ε_{i, t} . \end{matrix}

The OTHER term is adjusted, so that information in neither GDWLIP nor (SPI−GDWLIP) is included more than once in the equation. The predictive relevance of the GDWLIP with SPI model (Equation 13) is compared with the original BCN model (Equation 2), the expanded model with GDWLIP (Equation 4), and the SPI model (Equation 12).

Because goodwill impairments may be a reflection of broader economic issues already communicated via market-related information, we further test the robustness of our expanded model (Equation 4) by adding a variable to capture the change in the market capitalization (ΔMKTCAP) for each firm year. Market capitalization is computed by multiplying the fiscal-year-end closing price per share by the number of common shares outstanding (Compustat PRCC_F×CSHO), and ΔMKTCAP is computed as the difference between the beginning and ending market capitalization values for each firm year. Thus, the following equation results:

\begin{matrix} C F_{i, t + 1} = α + β_{CF} C F_{i, t} + β_{AR} Δ A R_{i, t} + β_{I} Δ IN V_{i, t} + β_{AP} Δ A P_{i, t} + β_{D} DEP R_{i, t} \\ + β_{AM} AMOR T_{i, t} + β_{GDWLIP} GDWLI P_{i, t} + β_{Δ MKTCAP} Δ MKTCA P_{i, t} \\ + β_{O} OTHE R_{i, t} + ε_{i, t} . \end{matrix}

Because information related to market capitalization is not included in the income-related variables in the equation, the OTHER term is computed identically for Equations 4 and 14. The predictive relevance of the market-related model (Equation 14) is compared with the expanded model with GDWLIP (Equation 4).

We conclude with two exploratory analyses. First, we return to the results of the original BCN model (Equation 2), and we separate firms into two groups based upon whether they exhibit above-median or below-median prediction error. Then, we separately evaluate each group for changes that result from including the GDWLIP variable (Equation 4). Changes are evaluated, as previously noted, based on adjusted R² values, PRESS statistics, and adjusted-MSPE statistics. Second, we subject the variables in the expanded model with GDWLIP (Equation 4) to automated regression with backward selection to derive a statistically generated model.

Results and Analysis

Our replication of the Barth et al. (2001) sample results in 11,488 firm-year observations for the 1987-1996 period as compared with 10,164 firm-year observations reported by Barth et al. The summary statistics of the Barth et al. sample (from their Table 1, pp. 38-39), as well as the summary statistics of our replication of their work, is provided in Table 1. We note the similarity of summary statistics between the Barth et al. sample and our replication effort. The additional firm-year observations could be accounted for by the possible backfilling of Compustat data between the Barth et al. article and our work.

Table 1.

Descriptive Statistics of Scaled Items.

The table below provides descriptive statistics for scaled measures used to predict future cash flows (CF_t+1). For comparison, we list the relevant descriptive statistics exactly as presented by Barth, Cram, and Nelson (2001). We then provide the descriptive statistics of our replication sample for the same period as Barth et al. After adding the Compustat GDWLIP variable to the replication model, we then provide descriptive statistics for the years 2001 to 2009. We begin our sample period in 2001 because goodwill impairments were not prevalent prior to the adoption of SFAS 142 in 2001. The sample period extends through 2010 to calculate future cash flows for fiscal year 2009.

The following restrictions used in the original Barth et al. are retained:

All firms with SIC codes from 6000 to 6999 are removed.

All firm-year observations with sales <US$10 million (SALE) or a stock price <US$1 (PRCC_F) are removed.

EARN must be calculable in years t− 1 and t− 2 for a firm-year observation to remain in the sample.

All variables must be present (not null values) for a firm-fiscal-year observation to enter the sample with the exception of GDWLIP for which missing values are assigned a value of 0.All variables are scaled by the average of the observation year (t) and the prior year (t+ 1) total assets (AT).

Observations are deleted if they are in the highest or lowest 1% of fiscal-year observations for EARN or CF or if they produce studentized residuals greater than 3 (in absolute value) after initial regressions to predict next fiscal year’s CF are estimated.

Barth et al. (2001) levels(1986-1997, n = 10,164)				Barth replication(1986-1997, n = 11,488)				Expanded model(2001-2009, n = 21,509)
Variable	M	Median	SD	Variable	M	Median	SD	Variable	M	Median	SD
EARN	0.04	0.04	0.08	EARN	0.03	0.04	0.09	EARN	0.03	0.04	0.11
CF	0.08	0.09	0.08	CF	0.10	0.09	0.09	CF	0.09	0.09	0.10
ΔAR	0.01	0.01	0.05	ΔAR	−0.01	−0.01	0.05	ΔAR	0.01	0.00	0.05
ΔINV	0.01	0.01	0.05	ΔINV	−0.01	0.00	0.04	ΔINV	0.00	0.00	0.04
ΔAP	0.01	0.01	0.04	ΔAP	0.01	0.01	0.04	ΔAP	0.01	0.00	0.04
DEPR	0.05	0.04	0.03	DEPR	0.06	0.05	0.03	DEPR	0.04	0.03	0.03
AMORT	0.01	0.00	0.02	AMORT	0.00	0.00	0.01	AMORT	0.01	0.00	0.01
OTHER	−0.01	−0.01	0.05	OTHER	0.03	0.02	0.15	GDWLIP	−0.01	0.00	0.04
								OTHER	−0.01	0.00	0.16

Note. The variables are defined as follows (COMPUSTAT data items in parentheses):

EARN = income before extraordinary items (IB); CF = cash flow from operating activities (OANCF) minus extraordinary items and discontinued operations (XIDOC); ΔAR = change in accounts receivable (RECCH), if available; otherwise, this variable is calculated using Accounts Receivable (RECT) as prior year RECT− present year RECT, to be consistent with the sign of RECCH; ΔINV = change in inventory (INVCH), if available; otherwise, this variable is calculated using Total Inventory (INVT) as prior year INVT− present year INVT, to be consistent with the sign of INVCH; ΔAP = change in accounts payable (APALCH), if available; otherwise, this variable is calculated using Accounts Payable (AP) and accrued expenses (XACC) as (present year AP+ present year XACC) − (prior year AP+ prior year XACC); DEPR = depreciation expense (XDP), if available; otherwise, DEPR is computed as depreciation and amortization (DP) minus amortization of intangibles (AM), when AM is present; or DP, when AM is not present; AMORT = amortization of intangibles (AM), if available; otherwise AMORT is equal to depreciation and amortization (DP) minus depreciation expense (XDP) minus depletion expense (XDEPL). Coverage of XDEPL in Compustat dissipates after 1997, thus XDEPL is utilized in our replication effort, but not for the new sample. Most observations of XDEPL were 0 from 1986 to 1997; GDWLIP = pretax goodwill impairments (GDWLIP); OTHER = EARN− (CF+ΔAR+ΔINV−ΔAP−DEPR−AMORT) for the original Barth et al. (2001) model and replication model; EARN− (CF+ΔAR+ΔINV−ΔAP−DEPR−AMORT−GDWLIP) for the expanded model including the GDWLIP variable.

The similarity between the results from Barth et al. (2001) and our replication provides strong evidence that our methodology is adequately replicating the efforts of Barth et al. Maintaining the previously described limitations, our sample includes 21,509 firm years from 2001 to 2009.¹⁹ Goodwill impairments are recognized in 1,918 (approximately 9%) of the total sampled firm years.²⁰ We assign GDWLIP a value of 0 when Compustat data are missing for this variable. We thus proceed in our efforts by comparing the coefficient estimates and significance levels across the cash-only (Equation 1), original BCN (Equation 2), and expanded BCN with GDWLIP (Equation 4) models for 1-year-ahead cash flows prediction. In addition to the explicit inclusion of GDWLIP in Equation 4, we note the difference in the construction of OTHER: For the estimation of Equation 2, OTHER is taken from Equation 3; and for the estimation of Equation 4, OTHER is taken from Equation 5. The results of this analysis are provided in Table 2.

Table 2.

Association and Forecasting of Future Cash Flows With Cash Only, Original BCN, and Expanded Models.

The table below presents coefficient estimates and t statistics obtained from regressing next-year cash flow (CF) on present year predictors, including GDWLIP (Equation 4), for the combined firm-years 2001-2009. For comparison purposes, results are presented from the same regression for the cash-only model (Equation 1) and the original BCN model without GDWLIP (Equation 2). Because prediction of future cash flows is our goal, the sample period extends through 2010 to calculate future cash flows for fiscal year 2009. Variables and sampling restrictions are as described in Table 1 and in the text.

Residual sum of square values for association (RSS) and for prediction via the predicted residual sum of squares (PRESS) statistic are presented, as well as F-test statistics for differences in RSS with and without the inclusion of GDWLIP in the model. Standard errors for t tests are corrected for clustering. We also present the test statistic for difference in MSPE based on the adjusted procedure and critical values of Clark and West (2007). We use the “rolling” scheme with 50% of observations used for estimation and the remaining 50% of observations used for forecast evaluation (i.e., R = P), as discussed in West (2006). Critical values for this test statistic are approximately those of the standard normal distribution for cases where R = P.

Variable	Cash only (Equation 1)			Original BCN (Equation 2)			Expanded BCN (Equation 4)
	Coefficient	t statistics	p value	Coefficient	t statistics	p value	Coefficient	t statistics	p value
Intercept	0.03	48.02	<.001	0.01	17.55	<.001	0.01	15.66	<.001
CF	0.66	133.83	<.001	0.71	140.57	<.001	0.72	142.55	<.001
ΔAR				0.15	−16.41	<.001	0.13	−7.73	<.001
ΔINV				0.06	1.33	.186	0.04	1.04	.297
ΔAP				−0.47	−36.24	<.001	−0.50	−38.75	<.001
DEPR				0.27	18.93	<.001	0.27	19.35	<.001
AMORT				0.18	5.80	<.001	0.18	5.59	<.001
GDWLIP							−0.43	−22.19	<.001
OTHER				0.11	17.76	<.001	0.19	25.01	<.001
n	21,509			21,509			21,509
Adjusted R²	.453			.515			.538
RSS	105.886			90.534			89.958
Full Model Results:								F statistics
				F test: Equation 4 relative to Equation 1				3,234.16	<.001
				F test: Equation 4 relative to Equation 2				122.21	<.001
Forecasting:
PRESS	105.915			90.654			90.088
		Clark and West (2007) adjusted-MSPE^a test statistic: Equation 4 relative to Equation 1						21.81	<.001
		Clark and West (2007) adjusted-MSPE^a test statistic: Equation 4 relative to Equation 2						4.04	<.001

Unadjusted-MSPE test statistics are as follows (caution should be exercised when interpreting unadjusted-MSPE test statistics as they are not normally distributed):

Equation 4 relative to Equation 1: 15.63 (p < .001).

Equation 4 relative to Equation 2: 2.34 (p < .001).

We note that GDWLIP has the predicted negative sign for its estimated beta coefficient (−.43) and is significant²¹ (t = −22.19, p < .001). These results are consistent with our prediction that goodwill impairments are significantly and inversely related to next-year CFO. The adjusted R² of the original BCN model (.515) results in a better degree of explanatory power than the cash-only model (.453), which is consistent with the conclusions of Barth et al. (2001). However, the explanatory power of the expanded model is further increased from an adjusted R², without GDWLIP, of .515 to an adjusted R² of .538 with GDWLIP. An F test of the residual sum of squared (RSS) differences between the original BCN (RSS = 90.534) and the expanded (RSS = 89.958) models indicates a significant improvement in the adjusted R² value (F = 122.21, p < .001) from the separate inclusion of the GDWLIP term. Not surprisingly, the improvement in adjusted R² between the cash-only (RSS = 105.886) and the expanded (RSS = 89.958) models is also significant (F = 3,234.16, p < .001).

Focusing on the improvement in forecast accuracy when GDWLIP is separately included in the model, the reduction in the PRESS statistic between the original BCN model (PRESS = 90.654) and the expanded model (PRESS = 90.088) indicates an improvement in the forecast accuracy of the expanded model vis-à-vis the original model. Improvement is also noted between the cash-only (PRESS = 105.915) and expanded (PRESS = 90.088) models. However, because there is no definitive test of the statistical significance of this improvement in the PRESS statistic, we also assess the improvement in forecast accuracy using the adjusted-MSPE statistic. Using the procedure identified by Clark and West (2007) and outlined previously, the adjusted-MSPE statistic for the expanded model relative to the cash-only model is 21.81 (p < .001), and the adjusted-MSPE statistic also indicates significant improvement in the forecast accuracy of the expanded model as compared with the original BCN model (adjusted-MSPE = 4.04, p < .001). This information is also summarized in Table 2. Based on these findings, we conclude that goodwill impairment information is not only consistently and inversely related to future cash flows, but that goodwill impairments also provide a significant, incremental improvement in the forecasting of future cash flows over both a cash-only model and the disaggregated BCN model.

To further examine the improvement of the expanded model over the original BCN model, we also compare 1-year-ahead cash estimates between Equation 2 (without GDWLIP) and Equation 4 (with GDWLIP) on a year-by-year basis from 2001 to 2009; and the results of these analyses are provided in Table 3. The estimated coefficients for GDWLIP (not tabulated) are consistently negative in each of the years studied and the average (median) of the t statistics for GDWLIP across all 9 years is −2.68 (−2.33). The change in the RSS between the original and expanded models is significant (p < .001) in all 9 years studied. Measuring the improvement in forecast accuracy, the PRESS statistic for the expanded model is also better (i.e., lower) than the PRESS statistic for the original model in all 9 years. It is interesting to note that the PRESS statistic becomes continuously smaller (better) for both models from 2005 through 2009 perhaps indicating that as preparers and auditors have become more familiar with the goodwill impairment standard that its application has also become increasingly relevant to future cash flows. Measuring the improvement in forecast accuracy statistically, the adjusted-MSPE shows significant improvement from the inclusion of the GDWLIP variable at the p < .001 level for 5 of the 9 years (2005-2009); interestingly, these years correspond to the time period during which the PRESS statistic becomes increasingly better (smaller), as well. One additional year (2004) shows statistical improvement by including the GDWLIP variable at the p < .05 level, and 2 more years (2001-2002) show statistical improvement at the p < .10 level. Thus, there is only 1 year (2003) for which the adjusted-MSPE statistic shows no significant increase in predictive accuracy from the addition of the GDWLIP variable. These findings further corroborate the conclusion that goodwill impairments are significantly and inversely related to next-year CFO and that separately including goodwill impairments in cash flow models provides a statistically significant increase in cash flow prediction and forecasting accuracy.

Table 3.

Association and Forecasting of Future Cash Flows, by Year, With and Without GDWLIP Included in Framework.

The table below presents the change in adjusted R², residual sum of squares (RSS), and predicted residual sum of squares (PRESS) when GDWLIP is (is not) included in models estimating future cash flows, that is, when Equation 4 (2) is estimated. The F statistics, comparing the fit of models for association, are provided based on the change in RSS. We also present, by year, the test statistic for difference in MSPE based on the adjusted procedure and critical values of Clark and West (2007). We use the “rolling” scheme with 50% of observations used for estimation and the remaining 50% of observations used for forecast evaluation (i.e., R = P), as discussed in West (2006). Critical values for this test statistic are approximately those of the standard normal distribution for cases where R = P.

Future cash flows (CF in year t+ 1) for observations in firm years 2001 to 2009 are modeled via Equation 4 (for GDWLIP and OTHER, including GDWLIP) and via Equation 2 (for OTHER, not including GDWLIP). All variables must be present for a firm-fiscal-year observation to enter the sample with the exception that if GDWLIP is missing, its value is set to 0. All other variables and sampling restrictions are retained as described in Table 1 and in the text.

Year	ΔAdjusted R²	RSS with GDWLIP	RSS without GDWLIP	F statistic ΔRSS	PRESS with GDWLIP	PRESS without GDWLIP	CW unadjusted-MSPE test statistic^a	CW adjusted-MSPE test statistic
2001	.007	10.396	10.493	19.20***	10.544	10.623	0.99	1.59*
2002	.005	9.776	9.875	22.80***	9.912	9.990	1.02	1.77*
2003	.002	10.635	10.676	9.92***	10.772	10.786	0.58	1.02
2004	.008	11.530	11.699	40.72***	11.663	11.826	1.34*	2.20**
2005	.019	12.076	12.616	126.19***	12.206	12.756	1.99**	3.22***
2006	.008	10.747	10.975	57.54***	10.863	11.087	1.59*	2.85***
2007	.012	8.509	8.675	42.84***	8.640	8.787	1.25	2.55***
2008	.019	8.327	8.602	71.25***	8.242	8.722	2.17**	3.37***
2009	.007	5.452	5.684	79.66***	5.526	5.777	1.78**	3.17***

Caution should be exercised when interpreting unadjusted-MSPE test statistics as they are not normally distributed.

Statistical significance levels of less than or equal to 10%, 5%, and 1% are denoted by *, **, and ***, respectively.

Table 4 presents the results of separately considering three specific types of non-recurring charges—restructuring charges, asset write-downs, and merger and acquisition costs—along with GDWLIP in the expanded model. The results reveal that restructuring charges (t = −2.44, p < .001) and merger and acquisition costs (t = −2.32, p = .020) are, themselves, significant predictors of future cash flows. The results of including asset write-downs indicated that this type of charge was not significantly related to future cash flows (t = −1.19, p = .234). However, rather than reducing the importance of the GDWLIP variable, the t statistic for GDWLIP increases by an almost identical amount and remains significant in every situation. The expanded model with GDWLIP (Equation 4) produced a t statistic for the GDWLIP variable of −22.19 (p < .001). When restructuring charges are included, GDWLIP has a t statistic of −23.34 (p < .001); when asset write-downs are included, GDWLIP has a t statistic of −23.46 (p < .001); and when merger and acquisition costs are included, GDWLIP has a t statistic of −22.37 (p < .001).

Table 4.

Inclusion of Other Non-Recurring Charges in the Expanded Model.

The table below presents coefficient estimates and t statistics obtained from regressing next-year cash flow (CF) on present year predictors, including GDWLIP (Equation 4) for the combined firm-years 2001 to 2009. For comparison purposes, additional non-recurring charges are utilized as well. Restructuring costs (RESTRUCT, Compustat: RCP), asset write-downs (WRITEDOWN, Compustat: WDP), and mergers and acquisition chargers (ACQUISITION, Compustat: AQP) are separately considered as potential additional explanatory variables, inside of the Equation 4 framework, after being scaled by the average of the prior 2 years total asset values. The inclusion of each of these variables results in Equations 9 (RESTRUCT), 10 (WRITEDOWN), and 11 (ACQUISITION). All other variables and sampling restrictions are retained as described in Table 1 and in the text.

Standard errors for t tests are corrected for clustering. Residual sum of square values for association (RSS) and prediction via the predicted residual sum of squares (PRESS) statistic are presented. We present the test statistics for difference in MSPE based on the adjusted procedure and critical values of Clark and West (2007). We use the “rolling” scheme with 50% of observations used for estimation and the remaining 50% of observations used for forecast evaluation (i.e., R = P), as discussed in West (2006). Critical values for this test statistic are approximately those of the standard normal distribution for cases where R = P.

Variable	Expanded model with GDWLIP (Equation 4)			Including restructuring costs (Equation 9)			Including asset write-downs (Equation 10)			Including M&A costs (Equation 11)
	Coefficient	t statistics	p value	Coefficient	t statistics	p value	Coefficient	t statistics	p value	Coefficient	t statistics	p value
Intercept	0.01	15.66	<.001	0.01	13.91	<.001	0.01	14.91	<.001	0.01	15.48	<.001
CF	0.72	142.55	<.001	0.73	141.24	<.001	0.73	143.18	<.001	0.72	143.46	<.001
ΔAR	0.13	7.73	<.001	0.15	9.49	<.001	0.10	5.98	<.001	0.13	7.79	<.001
ΔINV	0.04	1.04	.297	0.06	1.93	.054	−0.01	−0.55	.585	0.04	1.55	.121
ΔAP	−0.50	−38.75	<.001	−0.50	−38.56	<.001	−0.51	−39.80	<.001	−0.51	−39.22	<.001
DEPR	0.27	19.35	<.001	0.27	18.79	<.001	0.27	18.86	<.001	0.27	19.43	<.001
AMORT	0.18	5.59	<.001	0.13	4.03	<.001	0.16	5.07	<.001	0.16	5.06	<.001
GDWLIP	−0.43	−22.19	<.001	−0.45	−23.34	<.001	−0.46	−23.46	<.001	−0.43	−22.37	<.001
RESTRUCT				−0.16	−2.44	<.001
WRITEDOWN							−0.03	−1.19	.234
ACQUISITION										−0.26	−2.32	.020
OTHER	0.19	25.01	<.001	0.20	25.60	<.001	0.21	26.69	<.001	0.19	25.18	<.001
n	21,509			21,509			21,509			21,509
Non-zero values	GDWLIP = 1,918			RESTRUCT = 6,503			WRITEDOWN = 4,172			ACQUISITION = 1,310
Adjusted R²	.538			.540			.537			.539
RSS	89.958			89.873			89.942			89.916
Forecasting:
PRESS	90.088			90.062			90.086			90.068
Clark and West (2007) unadjusted-MSPE test statistic^a relative to Equation 4				1.17			0.02			0.77
Clark and West (2007) adjusted-MSPE test statistic relative to Equation 4				1.69*			0.16			1.28

Caution should be exercised when interpreting unadjusted-MSPE test statistics as they are not normally distributed.

A statistical level of significance less than or equal to 10% is denoted by *.

Considering the improvement in the overall model, the adjusted R² for the expanded model (Equation 4) is .538, and it changes little when restructuring charges (adjusted R² = .540), asset write-downs (adjusted R² = .537), or merger and acquisition charges (adjusted R² = .539) are included. Conversely, the RSS and PRESS statistics for each of the three additional models are slightly lower (better) than those of the expanded model with only GDWLIP (Equation 4). However, the adjusted-MSPE statistic reveals that only the restructuring charges model (adjusted-MSPE = 1.69) shows a significant improvement at the p < .10 level when compared with the expanded model with GDWLIP only (Equation 4). Therefore, we conclude that GDWLIP retains its highly significant relationship to future cash flows even after the inclusion of other non-recurring charges. In addition, the expanded model with GDWLIP is not greatly improved by the inclusion of additional non-recurring charges.

To further examine the potential effect of non-recurring items on the GDWLIP variable, we also introduce Compustat’s variable for all special items (SPI), which includes all non-recurring items (e.g., goodwill impairments, restructuring charges), into the model in two ways. First, we replace GDWLIP with SPI (Equation 12) to determine whether or not the total of all non-recurring charges might have a stronger relationship to future cash flows than GDWLIP alone. Second, we include both GDWLIP and (SPI−GDWLIP) in the model (Equation 13) to assess the continued reliability of the GDWLIP variable when the total of all other non-recurring charges is also included in the model. The results of these tests are presented in Table 5.

Table 5.

Inclusion of Special Items in the Expanded Model.

The table below presents coefficient estimates and t statistics obtained from regressing next-year cash flow (CF) on present year predictors, including GDWLIP (Equation 4) for the combined firm-years 2001 to 2009. For comparison purposes, special items (Compustat: SPI) excluding GDWLIP (i.e., SPI−GDWLIP) are considered, after being scaled by the average of the prior 2 years total asset values, as an additional potential explanatory variable both in replacement of GDWLIP and along with GDWLIP. The replacement of GDWLIP with SPI results in Equation 12, and the inclusion of (SPI−GDWLIP) along with GDWLIP results in Equation 13. The results of the original BCN model (Equation 2) are also presented for comparative purposes. All other variables and sampling restrictions are retained as described in Table 1 and in the text. Standard errors for t tests are corrected for clustering.

Variable	Original BCN model (Equation 2)			Expanded model with GDWLIP only (Equation 4)			Expanded model with SPI only (Equation 12)			Expanded model with GDWLIP and SPI−GDWLIP (Equation 13)
	Coefficient	t statistics	p value	Coefficient	t statistics	p value	Coefficient	t statistics	p value	Coefficient	t statistics	p value
Intercept	0.01	17.55	<.001	0.01	15.66	<.001	0.01	15.22	<.001	0.01	12.05	<.001
CF	0.71	140.57	<.001	0.72	142.55	<.001	0.72	140.07	<.001	0.75	144.34	<.001
ΔAR	0.15	16.41	<.001	0.13	7.73	<.001	0.14	8.14	<.001	0.08	2.04	.041
ΔINV	0.06	1.33	.186	0.04	1.04	.297	0.06	1.99	.047	0.09	4.64	<.001
ΔAP	−0.47	−36.24	<.001	−0.50	−38.75	<.001	−0.49	−37.22	<.001	−0.56	−42.70	<.001
DEPR	0.27	18.93	<.001	0.27	19.35	<.001	0.27	19.13	<.001	0.26	18.54	<.001
AMORT	0.18	5.80	<.001	0.18	5.59	<.001	0.15	4.52	<.001	0.12	3.96	<.001
GDWLIP				−0.43	−22.19	<.001				−0.39	−18.74	<.001
SPI							−0.29	−17.73	<.001
SPI−GDWLIP										−0.11	−8.57	<.001
OTHER	0.11	17.76	<.001	0.19	25.01	<.001	0.11	16.44	<.001	0.08	12.24	<.001
n	21,509			21,509			21,509			21,509
Non-zero values				GDWLIP = 1,918			SPI = 13,516
Adjusted R²	.515			.538			.540			.541

As shown in Table 5, the SPI variable is significant in the SPI model (t = −17.73, p < .001), and the adjusted R² of the SPI model (.540) is an improvement over the original BCN model (adjusted R² = .515). However, the adjusted R² of the SPI model is only slightly better than the adjusted R² of the expanded model with GDWLIP (.538; Equation 4). Furthermore, including both GDWLIP and (SPI−GDWLIP) in the model (Equation 13) produced an adjusted R² value (.541) that improves upon the original BCN model (adjusted R² = .515) but provides only a very small degree of improvement over the expanded equation with GDWLIP only (adjusted R² = .538). In addition, although both GDWLIP and (SPI−GDWLIP) are significant at p < .001, GDWLIP continues to produce a t statistic (−18.74) that is both large relative to the other variables in the model and larger than the t statistic for (SPI−GDWLIP; −8.57). Thus, the explanatory and predictive relevance of GDWLIP with respect to future cash flows is robust not only to the inclusion of certain specific non-recurring charges (i.e., restructuring costs, asset impairments, and merger and acquisition costs), but also to the inclusion of all non-recurring charges, other than goodwill impairments, in a single term. In addition, the GDWLIP term provides an almost as much explanatory power as does the SPI term.

To examine whether or not goodwill impairments provide unique information beyond what is communicated through market-related data, we add a variable to our expanded model (Equation 4) to capture the change in market capitalization (ΔMKTCAP) for each firm year resulting in Equation 14. ΔMKTCAP is computed as the difference between the beginning and ending market capitalization values for each firm year. The results of this analysis are presented in Table 6.

Table 6.

Inclusion of Market-Related Data in the Model.

The table below presents coefficient estimates and t statistics obtained from regressing next-year cash flow (CF) on present year predictors, including GDWLIP (Equation 4) for the combined firm-years 2001 to 2009. In addition, the change in market capitalization (ΔMKTCAP), in millions of dollars, is included as an explanatory variable. Market capitalization is computed using Compustat PRCC_F times Compustat CSHO, and ΔMKTCAP is measured as the present year market capitalization minus the prior year market capitalization. ΔMKTCAP is included inside of the Equation 4 framework resulting in Equation 14. All other variables and sampling restrictions are retained as described in Table 1 and in the text.

Name	Expanded model with GDWLIP (Equation 4)			Expanded model with market capitalization (Equation 14)
	Coefficient	t statistics	p value	Coefficient	t statistics	p value
Intercept	0.01	15.66	<.001	0.01	17.13	<.001
CF	0.72	142.55	<.001	0.71	140.73	<.001
ΔAR	0.13	−7.73	<.001	0.19	8.84	<.001
ΔINV	0.04	1.04	.297	−0.01	0.21	.839
ΔAP	−0.50	−38.75	<.001	−0.49	−37.85	<.001
DEPR	0.27	19.35	<.001	0.27	19.00	<.001
AMORT	0.18	5.59	<.001	0.14	4.50	<.001
GDWLIP	−0.43	−22.19	<.001	−0.33	−17.20	<.001
ΔMKTCAP				0.15	5.94	<.001
OTHER	0.19	25.01	<.001	0.15	19.82	<.001
n	21,509			21.509
Adjusted R²	.538			.550

As one might expect, the results in Table 6 show that ΔMKTCAP is positively and significantly associated with future cash flows (t = 5.94, p < .001). However, even when this market-related information is included in the model, GDWLIP maintains its inverse and highly significant relationship with future cash flows (t = −17.20, p < .001). Thus, the inclusion of market-related information in the model does not appreciably diminish the predictive relevance of goodwill impairments to future cash flows.

To further explore these results, we return to the original BCN model (Equation 2) and split our firms into two groups: those with above-median prediction errors and those with below-median prediction errors. We then apply the expanded model with GDWLIP (Equation 4) to each group to determine whether or not the improved model produces similar results for each group. The results of our analysis are shown in Table 7. As we would expect, the below-median prediction error group has a higher adjusted R² (.904) than the above-median prediction error group (adjusted R² = .422). However, the adjusted R² for both groups improves when GDWLIP is included in the model. In addition, the PRESS statistic improves for both groups, with the below-median prediction error group improving from 3.281 to 3.236 and the above-median prediction error group improving from 87.372 to 86.853. However, the adjusted-MSPE statistic shows a significant improvement in model performance only for the above-median error prediction group (MSPE = 2.28, p < .05). This indicates that the less reliable the other financial information is in assisting users to predict future cash flows, the more goodwill impairments disclosures improve cash flow forecasting ability.

Table 7.

Comparison of Improvement in Future Cash Flows Prediction by Including Goodwill Impairments for Firms With Below-Median and Above-Median Prediction Errors Based on the Original BCN Model.

The table below presents coefficient estimates and t statistics obtained from regressing next-year cash flow (CF) on present year predictors. The results are presented when the full sample of firm-year observations is split into those observations that demonstrate a below-median (above-median) prediction error when the original model (without GDWLIP, that is, Equation 2) is estimated. The relative contribution of GDWLIP is then demonstrated by estimating specifications after including GDWLIP (i.e., Equation 4) for each part of the sample. All variables and sampling restrictions are retained as described in Table 1 and in the text.

Standard errors for t tests are corrected for clustering. Residual sum of square values for prediction via the predicted residual sum of squares (PRESS) statistic are presented. We also present the test statistic for difference in MSPE based on the adjusted procedure and critical values of Clark and West (2007). We use the “rolling” scheme with 50% of observations used for estimation and the remaining 50% of observations used for forecast evaluation (i.e., R = P), as discussed in West (2006). Critical values for this test statistic are approximately those of the standard normal distribution for cases where R = P.

Variable	Firm-years with below-median prediction errors						Firm-years with above-median prediction errors
	Original BCN model (Equation 2)			Expanded model with GDWLIP (Equation 4)			Original BCN model (Equation 2)			Expanded model with GDWLIP (Equation 4)
	Coefficient	t statistics	p value	Coefficient	t statistics	p value	Coefficient	t statistics	p value	Coefficient	t statistics	p value
Intercept	0.01	41.41	<.001	0.01	41.12	<.001	0.01	8.27	<.001	0.01	7.78	<.001
CF	0.72	307.91	<.001	0.72	303.59	<.001	0.70	81.97	<.001	0.72	83.79	<.001
ΔAR	0.25	3.23	<.001	0.24	−29.74	<.001	0.23	9.35	<.001	0.08	−2.98	.003
ΔINV	0.17	20.85	<.001	0.16	−18.22	<.001	0.15	5.26	<.001	0.02	0.57	.567
ΔAP	−0.47	−79.84	<.001	−0.47	−79.43	<.001	−0.48	−21.41	<.001	−0.51	−23.12	<.0001
DEPR	0.25	40.72	<.001	0.25	40.61	<.001	0.30	11.82	<.001	0.29	11.69	<.001
AMORT	0.21	16.4	<.001	0.21	16.22	<.001	0.19	3.24	.001	0.13	2.35	.019
GDWLIP				−0.14	−11.92	<.001				−0.55	−17.07	<.001
OTHER	0.12	37.99	<.001	0.16	33.37	<.001	0.12	11.16	<.001	0.18	17.08	<.001
n	10,755			10,755			10,754			10,754
Adjusted R²	.904			.905			.422			.444
Forecasting:
PRESS	3.281			3.236			87.372			86.853
CW unadjusted-MSPE test statistic^a relative to Equation 2				0.12						1.39*
CW adjusted-MSPE test statistic relative to Equation 2				0.49						2.28***

Caution should be exercised when interpreting unadjusted-MSPE test statistics as they are not normally distributed.

Statistical significance levels of less than 10% and 1% are denoted by * and ***, respectively.

In our final exploratory procedure, we estimate the expanded model with GDWLIP (Equation 4) under an automated backward selection procedure.²² The results in Table 8 show that the GDWLIP variable provides significant incremental adjusted R² and that it ranks ahead of depreciation, amortization, and change in accounts receivable variables in uniquely contributing to the explanatory/predictive power of the model. These results also corroborate the conclusion that although goodwill impairments are “non-cash” events, they provide a statistically significant improvement in the prediction of future cash flows. Thus, analysts, investors, creditors, and others interested in future cash flows should separately consider goodwill impairments in their prediction and forecasting models.

Table 8.

Backward Selection Criteria of Regression Estimating Future Cash Flows.

The table below presents the results for the estimation of Equation 3 utilizing backward selection criteria. Future cash flows (CF in year t+ 1) for observations in firm years 2001 to 2009 are modeled via Equation 3, and variables with coefficient p values < .15 are included in the final model. All variables are scaled by the average of the observation year and the prior year total assets. This scaling is identical to Barth, Cram, and Nelson (2001). To remain consistent with Barth et al., EARN must additionally be calculable in years t− 1 and t− 2 for a firm-year observation to remain in the sample. All SIC codes from 6000 to 6999 are removed, as are firm-year observations with sales <US$10 million in the Compustat fiscal year or a stock price <US$1 on the Compustat firm-year observation date. Finally, for consistency to Barth et al., observations in the highest and lowest 1% of fiscal-year observations of EARN and CF are deleted, as are observations with studentized residuals greater than 3 in absolute value after initial regressions predicting next fiscal year’s CF are estimated. All variables must be present for a firm-fiscal-year observation to enter the sample with the exception that if data for GDWLIP are missing, its value is set to 0. The F statistics, which are the sum of squares for the variable divided by the mean square error for the full model, are presented as are the incremental contributions to R² of each variable that is included in the final selection model.

Variable	Coefficient	Partial R²	F statistics	p value
CF	0.73	88.99	20,319.40	<.001
ΔAP	−0.50	6.28	1,501.62	<.001
OTHER	0.19	2.61	625.33	<.001
GDWLIP	−0.44	2.06	492.34	<.001
DEPR	0.27	1.57	374.59	<.001
ΔAR	0.12	.25	59.79	<.001
AMORT	0.17	.13	31.25	<.001

Given the findings above, future research might proceed along one of the following lines. While our study uses regression models based on annual data, similar to Barth et al. (2001), future research could examine whether or not these findings persist when using other cash flow prediction models (e.g., Autoregressive Integrated Moving Average (ARIMA) models; see Lorek & Willinger, 1996, 2008, 2011) and/or quarterly data. Hayn and Hughes (2006) indicate that the quality of disclosures may also affect the usefulness of information to investors, thus the quality of goodwill impairment disclosures—instead of, or perhaps in addition to, the quantitative amounts of the impairments—may be another fruitful line of inquiry. In light of the recent Private Company Council (PCC)/FASB decision related to the treatment of goodwill for private versus public companies (FASB, 2014), the relative ability of goodwill impairments to predict future cash flows for public and private firms could also be examined.²³ Future studies could also examine whether industry affiliation strengthens or weakens the relationship between goodwill impairments and future cash flows. Finally, future studies could explore the manner and extent to which various types of market-related information attenuate the relationship between goodwill impairments and future cash flows. Such studies could also assess whether such intervening information leads, lags, or coincides with the relationship between goodwill impairments and future cash flows.

Conclusion

This study contributes to the understanding of cash flow prediction by demonstrating that goodwill impairments provide a significant, incremental improvement in the prediction and forecasting of future cash flows. The predictive relevance of goodwill impairments to future cash flows is examined using not only in-sample techniques, but also out-of-sample testing. Thus, this study is also one of only a few studies to investigate the incremental, out-of-sample predictive power of non-current accruals on reported (as opposed to computed) operating cash flows. In addition, we find that, as predicted, goodwill impairments are inversely related to future cash flows both aggregately across all years and for each year individually within the 2001-2009 study window. These findings are robust both when other non-recurring items (i.e., restructuring costs, asset write-downs, and merger and acquisition costs) are separately considered along with goodwill impairments and when the total of all other non-recurring items is considered together with goodwill impairments. In addition, goodwill impairment information has a more significant relationship to future cash flows than the sum of all other non-recurring items considered as a whole. We also find that even when market-related information is included in the model, goodwill impairments still provide unique information related to future cash flows. Finally, we find that while the inclusion of goodwill impairments improves the prediction of future cash flows for all firms, it provides a more significant improvement for firms with otherwise above-median prediction errors than for firms with below-median prediction errors. Thus, it would appear that the less reliable the other financial information is, the more useful the inclusion of goodwill impairments information becomes for predicting future cash flows. As a result of these findings, analysts, investors, creditors, and others interested in predicting cash flows should separately consider goodwill impairment information, when available, to improve cash flow prediction and forecasting.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

References

Al-Debi’e

M. M.

(2011). Are operating cash flows a superior predictor of future operating cash flows than earnings? Evidence from Jordan. European Journal of Economics, Finance and Administrative Sciences, 40, 36-46.

American Institute of Certified Public Accountants. (1963). Accounting Principles Board Opinion No. 3: The statement of source and application of funds. New York, NY: Author.

American Institute of Certified Public Accountants. (1971). Accounting Principles Board Opinion No. 19: Reporting changes in financial position. New York, NY: Author.

Ball

Brown

(1968). An empirical evaluation of accounting income numbers. Journal of Accounting Research, 6, 159-178.

Bandyopadhyay

S. P.

Chen

Huang

A. G.

Jha

(2010). Accounting conservatism and the temporal trends in current earnings’ ability to predict future cash flows versus future earnings: Evidence on the trade-off between relevance and reliability. Contemporary Accounting Research, 27, 413-460.

Barth

M. E.

Cram

D. P.

Nelson

K. K.

(2001). Accruals and the prediction of future cash flows. Accounting Review, 76, 27-58.

Beaver

W. H.

Dukes

R. E.

(1972). Interperiod tax allocation, earnings expectations, and the behavior of security prices. Accounting Review, 47, 320-332.

Bowen

R. M.

Burgstahler

Daley

L. A.

(1986). Evidence on the relationships between earnings and various measures of cash flow. Accounting Review, 61, 713-725.

Bowen

R. M.

Burgstahler

Daley

L. A.

(1987). The incremental information content of accrual versus cash flows. Accounting Review, 62, 723-747.

10.

Cheng

C. C.

Hollie

(2008). Do core and non-core cash flows from operations persist differentially in predicting future cash flows?Review of Quantitative Finance and Accounting, 31, 29-53.

11.

Clark

McCracken

(2001). Tests of equal forecast accuracy and encompassing for nested models. Journal of Econometrics, 105, 85-110.

12.

Clark

West

(2007). Approximately normal tests for equal predictive accuracy in nested models. Journal of Econometrics, 138, 291-311.

13.

Dechow

P. M.

Kothari

S. P.

Watts

R. L.

(1998). The relation between earnings and cash flows. Journal of Accounting & Economics, 25, 133-168.

14.

Farshadfar

Chew

Brimble

(2008). The relative ability of earnings and cash flow data in forecasting future cash flows: Some Australian evidence. Pacific Accounting Review, 20, 254-268.

15.

Financial Accounting Standards Board. (1978). Statement of financial accounting concept no. 1: Objectives of financial reporting by business enterprises. Norwalk, CT: Author.

16.

Financial Accounting Standards Board. (1987). Statement of financial accounting standard no. 95: Statement of cash flows. Norwalk, CT: Author.

17.

Financial Accounting Standards Board. (2001). Statement of financial accounting standard no. 142: Goodwill and other intangible assets. Norwalk, CT: Author.

18.

Financial Accounting Standards Board. (2010). Statement of financial accounting concept no. 8: Conceptual framework for financial reporting: The objective of general purpose financial reporting and qualitative characteristics of useful financial information. Norwalk, CT: Author.

19.

Financial Accounting Standards Board. (2014). Accounting Standards Update No. 2014-02: Intangibles—Goodwill and other (Topic 350). Retrieved from https://asc.fasb.org/imageRoot/79/49129379.pdf

20.

Financial Accounting Standards Board. (2015). Accounting standards codification. Available from https://asc.fasb.org/

21.

Finger

C. A.

(1994). The ability of earnings to predict future earnings and cash flow. Journal of Accounting Research, 32, 210-223.

22.

Francis

R. A.

(2011). Out-of-sample cash flow prediction and cash distributions to shareholders. Advances in Accounting, 27, 1-9.

23.

Greenberg

R. R.

Johnson

G. L.

Ramesh

K. K.

(1986). Earnings versus cash flow as a predictor of future cash flow measures. Journal of Accounting, Auditing & Finance, 1, 266-277.

24.

Habib

(2010). Prediction of operating cash flows: Further evidence from Australia. Australian Accounting Review, 20, 134-143.

25.

Hales

Orpurt

S. F.

(2013). A review of academic research on the reporting of cash flows from operations. Accounting Horizons, 27, 539-578.

26.

Hayn

Hughes

P. J.

(2006). Leading indicators of goodwill impairment. Journal of Accounting, Auditing & Finance, 21, 223-265.

27.

Investor Responsibility Research Center Institute. (2012). Controlled companies in the Standard & Poor’s 1500: A ten year performance and risk review. Retrieved from irrcinstitute.org/pdf/FINAL-Controlled-Company-ISS-Report.pdf

28.

Jarva

(2009). Do firms manage fair value estimates? An examination of SFAS 142 goodwill impairments. Journal of Finance and Accounting, 36, 1059-1086.

29.

Kim

Kross

(2005). The ability of earnings to predict future operating cash flows has been increasing-not decreasing. Journal of Accounting Research, 43, 753-780.

30.

Krishnan

G. V.

Largay

J. A., III

. (2000). The predictive ability of direct method cash flow information. Journal of Business Finance & Accounting, 27, 215-245.

31.

Lee

(2011). The effect of SFAS 142 on the ability of goodwill to predict future cash flows. Journal of Accounting and Public Policy, 30, 236-255.

32.

Lee

Yoon

S. W.

(2012). The effects of goodwill accounting on informativeness of earnings: Evidence from earnings persistence and earnings’ ability to predict future cash flows. Journal of Accounting and Finance, 12, 124-147.

33.

Lev

Sougiannis

(2010). The usefulness of accounting estimates for predicting cash flows and earnings. Review of Accounting Studies, 15, 779-807.

34.

Lorek

K. S.

Fimia-Moe

(2012). A critical assessment of the state of the art of statistically-based annual cash-flow prediction models. Journal of Theoretical Accounting Research, 7, 68-81.

35.

Lorek

K. S.

Schaefer

T. F.

Willinger

G. L.

(1993). Time-series properties and predictive ability of funds flow variables. Accounting Review, 68, 151-163.

36.

Lorek

K. S.

Willinger

G. L.

(1996). A multivariate time-series prediction model for cash-flow data. Accounting Review, 71, 81-101.

37.

Lorek

K. S.

Willinger

G. L.

(2008). Time-series properties and predictive ability of quarterly cash flows. Advances in Accounting, 24, 65-71.

38.

Lorek

K. S.

Willinger

G. L.

(2009). New evidence pertaining to the prediction of operating cash flows. Review of Quantitative Finance and Accounting, 32, 1-15.

39.

Lorek

K. S.

Willinger

G. L.

(2010a). The predictive ability of quarterly cash-flow models. Journal of Theoretical Accounting Research, 5, 59-81.

40.

Lorek

K. S.

Willinger

G. L.

(2010b). Time series versus cross-sectionally derived predictions of future cash flows. Advances in Accounting, 26, 29-36.

41.

Lorek

K. S.

Willinger

G. L.

(2011). Multi-step ahead quarterly cash-flow prediction models. Accounting Horizons, 25, 71-86.

42.

Luo

(2008). Unusual operating cash flows and stock returns. Journal of Accounting and Public Policy, 27, 420-429.

43.

Montgomery

D. C.

Peck

E. A.

Vining

G. G.

(2012). Introduction to linear regression analysis (5th ed.). Hoboken, NJ: John Wiley.

44.

Nam

Brochet

Ronen

(2012). The predictive value of accruals and consequences for market anomalies. Journal of Accounting, Auditing & Finance, 27, 151-176.

45.

Orpurt

S. F.

Zang

(2009). Do direct cash flow disclosures help predict future operating cash flows and earnings?Accounting Review, 84, 893-935.

46.

Rayburn

(1986). The association of operating cash flow and accruals with security returns. Journal of Accounting Research, 24, 112-133.

47.

Telmoudi

Noubbigh

Ziadi

(2010). Forecasting of operating cash flow: Case of the Tunisian commercial companies. International Journal of Business and Management, 5(10), 198-210.

48.

Telmoudi

Ziadi

Hedi

(2010). Factors determining operating cash flow: Case of the Tunisian commercial companies. International Journal of Business and Management, 5(5), 188-200.

49.

West

(2006). Chapter 3 forecast evaluation. Handbook of Economic Forecasting, 1, 99-134.

50.

Wilson

(1986). The relative information content of accruals and cash flows: Combined evidence at the earnings announcement and annual report release date. Journal of Accounting Research, 24, 165-200. doi:10000

51.

Wilson

(1987). The incremental information content of the accrual and funds components of earnings after controlling for earnings. Accounting Review, 62, 293-322.

52.

Yoder

T. R.

(2007). The incremental predictive ability of accrual models with respect to future cash flows. Social Science Research Network. Retrieved from http://ssrn.com/abstract=962700