A comparative analysis of earnings forecast models augmented for interindustry dependencies and phased for economic cycles

The purpose of this study was to investigate and compare the behavior and forecastability of quarterly earnings at the industry level, via the Box and Jenkins methodology. On the basis of existing interindustry relationships, 10 manufacturing industries were selected. The quarterly earnings series of the respective industries were subjected to univariate and bivariate modeling procedures in each of four historical periods. The investigation was directed toward providing a descriptive and predictive evaluation of various combinations of the identified augmented and phased models.

Earnings data for the firms constituting the industries of this study were obtained from COMPUSTAT Quarterly Industrial tapes and the Wall Street Journal Index. Statistical analyses were conducted using the BMDP Statistical Package developed by the University of California, Los Angeles. Data contained in the 1967 Input-Output Tables, published by the Department of Commerce, were relied upon for identification of industries expected to be of potential value in the constructed bivariate models.

Of the 40 univariate models identified, the most frequently occurring (20 times) was a multiplicatively differenced model with a moving average seasonal component (010 x 011, in Box-Jenkins notation). The Griffin-Watts-Lorek multiplicative first-order moving average model (011 x 011) was identified in 10 instances. In two instances, the Brown-Rozeff first-order autoregressive seasonal moving average model (100 x 011, with a drift adjustment, was identified.

The bivariate models applied to the various industry earnings series generally performed poorly. Although the bivariate, industry model (010 x 011) proved to be superior when models were constructed over the total period, the superiority was not maintained when the overall period was subdivided into phases. Apparently, the presence of feedback in the investigated earnings numbers affected the descriptive ability of all of the alternative models.

The industry, Griffin-Watts-Lorek, and Brown-Rozeff models performed about equally well in a predictive sense. While the Foster (100 X 010) model was generally outperformed, it did excel in selected quarters. Forecasts produced from bivariate models exhibited little improvement over univariate forecasts. Although the expected improvement in predictive ability from bivariate models was not realized, this should not be a deterrent to future research into the behavior and forecastability of quarterly earnings at the industry level. Such research is vital to policy-makers who require knowledge about the time-series properties of accounting numbers.