The predictive ability of entity versus sub-entity earnings : a test using simulated mergers of existent autonomous firms

Simulated mergers of actual firms were used to provide totally comparable CN-SG data sets. In addition, by using quarterly data, it became feasible to use Box-Jenkins analysis and seasonal ARIMA models in a segment reporting context.

Merger candidates were chosen from a subset of COMPUSTAT firms comprised of domestic, autonomous, nonregulated, single-product companies reporting few extraordinary items during the 1967-1975 sample period. These "surrogate segments" (N = 60) were merged to form various n-segment conglomerates (N = 46). There were twenty 3-segment, twelve 5-segment, eight 7-segment, and six 10-segment conglomerates, each accounted for as a pooling of interests. By design, the hypothetical conglomerates were void of problems due to segment definition, intersegment transfers, common cost allocations, and tax allocations.

Because Box-Jenkins analysis entails judgment, all data sets were completely disguised and presented in random order. . These precautions controlled for subconscious experimenter bias and learning effects which might have confounded the results.

A multiple-hypothesis framework, consisting of 128 specific CN-SG comparisons, was formulated by varying six factors: (1) number of segments, (2) size of segments, (3) forecast horizon, (4) error metric, (5) number of estimation observations, and (6) forecast period. Each comparison was tested for mean and median differences. T-tests and Wilcoxon tests were used in parallel.

Test results and descriptive statistics revealed few, if any, differences in predictiveness between CN forecasts and SG forecasts. Only 5.5 percent of the null hypotheses were rejected at the .05 level. It appears, therefore, that there are no differences in predictive ability between CN forecasts and SG forecasts when predicting net income with seasonal ARIMA models.