Historical cost measurement and the use of DuPont analysis by market participants

The DuPont decomposition is a well-known ratio analysis, which decomposes return on assets into profit margin and asset turnover. It has been used since the early 1900 s and began gaining popularity following Pierre DuPont’s successful turnaround of General Motors in the 1920 s using the DuPont decomposition in the firm’s managerial accounting system. The DuPont decomposition remains a central element of ratio analysis and forecasting in current financial statement analysis textbooks.

Specifically, Konchitchki (2011, 2013) finds evidence that investors underreact to unrealized gains and losses due to inflation, consistent with investors being unable to accurately incorporate unrecognized changes in value.

We assume that differences between modified historical cost and current values are more severe for firms with older assets. Also implicit in our use of this measure is the assumption that the salvage value for all assets is zero and that all firms employ straight-line deprecation. Providing some support for this assumption, Accounting Trends and Techniques (AICPA 2007) finds that the most common depreciation method is straight-line.

Furthermore, we also find that the future profitability of firms in the highest and lowest quintiles of asset age is not significantly different. Thus the linkage between the asset turnover ratio and competitive advantages, at least those that culminate in realized earnings, is unlikely to be correlated with asset age.

Specifically, for the AR(1) model, we assign all firms annually to five portfolios based on asset age in year t − 1 and estimate the persistence of asset turnover using data from t − 1 and t − 2 (regressing ATO _t−1 on ATO _t−2 to form an estimated persistence coefficient within firm age quintiles). We use the persistence coefficients from these regressions to estimate a forecast of asset turnover in year t. For the analyst forecast model, we infer the analysts’ ATO forecast by dividing analysts’ forecasted sales for year t by average NOA over t − 1 (see Fig. 1 for a graphical depiction of the timing of variable measurement).

We use characteristic-adjusted returns following Daniel et al. (1997). The characteristic-adjusted returns remove the average return on a portfolio of firms matched on size, book-to-market, and momentum.

We cannot provide inference on whether the returns we document constitute a profitable trading opportunity for two reasons. First, we do not have access to trading cost data or trade impact data. Second, we examine a stylized model of abnormal returns that does not optimize the returns to co-movement risk; see Richardson et al. (2010) for more details. Anecdotally, biases stemming from asset age appear important to sophisticated market participants. For example, Credit Suisse offers a product called HOLT that it claims removes “differences across firms in asset age, asset life, inflation, and other accounting distortions.” See https://​www.​credit-suisse.​com/​sites/​holt/​en.​html.

Dickinson and Sommers (2012), however, find that many competitive advantages do not lead to a sustainable increase in return on net operating assets.

Technically, any ratio of two nonstationary (i.e., trending variables) like sales and assets are unlikely to yield a mean-reverting ratio unless they are cointegrated. Economic theory suggests that sales and assets are cointegrated, as it is difficult to increase sales over an extended period without increasing assets. If a ratio excludes a trending term such as unrecognized appreciation, increasing the unreported value of an asset, then these ratios will become more like a random walk, which increases persistence; see for example Chapter 19 of Hamilton (1994).

In addition, in the early part of our sample, assets can get “older” due to merger and acquisition activities under the pooling method.

Although we anticipate that more precise measures, for example by the method of Konchitchki (2011), would increase the power of our tests, we do not expect any systematic bias resulting from using our simple measure.

Gormley and Matsa (2014) find that estimates of regressions where only the dependent variable is industry-adjusted will lead to a correlated omitted bias problem and recommend using fixed effects. Our approach is consistent with the inclusion of industry-year fixed effects.

We begin in 1964 as we require lagged data and Compustat data before 1963 is sparsely populated. We end in 2012 as our future returns tests require data two and a half years from each calendar year-end (i.e., we use future returns through June 2014). We use the annual database as we require property, plant, and equipment (gross and net), which are sparsely populated on the quarterly database. Specifically, for fiscal year-ends 1964–2012, the annual database has PPENT and PPEGT populated for 75 % of the sample, whereas the quarterly database has only 38 % of the sample populated for the quarterly equivalents PPENTQ and PPEGTQ (not tabulated).

Specifically, we require nonmissing observations of sales, operating income, total assets, common equity, depreciation expense, and gross and net property, plant, and equipment for the current and lagged fiscal years. All other variables required for the calculation of our main variables are set to zero when data is missing.

We estimate the statistical significance of the difference in the mean values between the extreme quintiles after accounting for both cross-sectional and time-series correlation.

We find that the positive association between asset age and ATO ratios is robust to controlling for economic determinants of performance such as size, book-to-market, beta, sigma, firm age, lagged operating performance, and future performance (not tabulated).

When we examine profit and loss firms separately, we find that ATO is significantly higher for older asset firms for both samples. In contrast, we find that profitable older asset firms have significantly lower RNOA and PM than newer asset firms but older asset firms with a loss have significantly higher RNOA and PM than newer asset firms with a loss (not tabulated).

Following Richardson (2006), we measure total investment (TI) as the sum of capital expenditures, research and development expenditures, and acquisitions, less the sale of property, plant, and equipment, plus amortization and depreciation (CAPX + XRD + AQC − SPPE + DPC). We then scale this measure by total assets (TI/TA).

Our inferences are unchanged when estimating Eqs. 5–7 using the Fama–MacBeth approach where the regression is estimated by year and obtaining the t-statistics based on the distribution of the parameter estimates for each asset age quintile (not tabulated).

The difference in the persistence estimates for the top and bottom quintiles is also statistically significant when quintiles are formed on the basis of the average asset age in t and t–1 (not tabulated).

We also forecast RNOA using prior RNOA, the percentage change in ATO, sales growth, and the interaction of ATO and sales growth following Richardson et al. (2006). Richardson et al. (2006) highlight that accounting distortions will be manifest in change in ATO and diminishing marginal returns in sales growth. We find some evidence that accounting distortions are observed in older asset firms significantly more than in newer firms (not tabulated).

Thus the AR(1) forecast errors are out-of-sample forecast errors and are estimated from 1965–2012, as 1965 is the first year with lagged estimates of Eqs. 5–7 and 2012 is the last year that allows us to calculate future returns from July 2013 through June 2014.

Another common approach to forming portfolios is to assume that accounting information is available 4 months after the fiscal year-end. Our results are similar, but the future return spreads (Q5 − Q1) are approximately 100 basis points higher for the forecast errors of asset turnover (not tabulated). As the inferences are otherwise similar, we report the more conservative tests.

Sales forecasts are only available on I/B/E/S for the largest firms. For our sample, beginning in 1996, the average market capitalization of the analyst forecast sample is $4994.9 million, which is significantly greater (t = 28.7, p < 0.001) than the average market capitalization of $1082.4 million for the firms not in the analyst forecast sample in the period 1996–2012 based on a two sample t test (not tabulated). For these smaller firms without a sales forecast on I/B/E/S, the results are similar to those reported for the full sample period (not tabulated).

Note that we forecast ΔNOA _t rather than NOA. An alternative is to estimate the AR(1) forecast errors for NOA as NOA_t − \(\widehat{NOA}_{t}\) from an AR(1) model where \(\widehat{NOA}_{t} = \hat{\phi }_{0Pt - 1} + \hat{\phi }_{1Pt - 1} NOA_{t - 1}\). When included in the multivariate model in Column 2, Panel B of Table 6, we find similar inferences on the AR(1) ATO forecast errors, but the NOA based forecast errors do not have a statistically significant association with future returns. This is largely due to many of the estimates of \(\hat{\phi }_{1Pt - 1}\) being >1, making the forecasts unreliable (not tabulated).

We report models where the forecast errors are relatively consistent for brevity. We also estimated models examining the robustness of the AR(1) forecast errors for ATO when including industry-adjusted ΔNOA _t and (unadjusted) ΔNOA _t , which is most similar to Richardson et al. (2005). In both cases, the t-statistic on the AR(1) forecast errors for ATO declined by less than 1, and both are above 3 (not tabulated). We also included the random walk forecast of ATO and the forecast error of ATO in the same model to examine the incremental explanatory power of the model-based forecast errors. We find that ATO model-based forecast errors continue to be positively associated with future returns, whereas the random walk forecast errors become negatively associated with future returns (not tabulated). This supports our conjecture that the driver of the association between future returns and the change in ATO is asset age bias.

An alternative test would be to examine whether changes in inflation over time are associated with variation in the asset age bias, which we could test in our US sample. The benefit would be to remove any between country effects associated with the historical cost bias and the asset turnover. The shortcomings of this approach include that the significant decline in inflation in the US setting is associated with systematic changes in productive assets confounding interpretation of results. We choose to investigate the cross-country sample for this reason. Nonetheless, in the US sample, we do find a positive association between asset-turnover persistence and inflation, consistent with inflation increasing the measurement bias in reported asset turnover ratios (not tabulated).

http://​data.​worldbank.​org/​indicator/​NY.​GDP.​DEFL.​KD.​ZG.