Abstract ABSTRACT: Analytical review can be a relatively inexpensive means for reducing detailed substantive testing in auditing. This article presents a formulation and comparison of four statistical and two naive procedures for conducting analytical review which rely on substantially different information sets. Specifically, ordinary least squares regression predictions and three sets of integreted-autoregressive-moving-averege-(ARIMA) based predictions are compared with martingale and submartingale predictions for a set of monthly accounting series. The data for the empirical work are 15 years of operating revenues of a sample of six railroads in the southwestern United States. All methods exhibit a slight prediction bias but the ARIMA-based univariate transfer function, which requires the largest information set and the greatest computation effort, yields the smallest mean absolute error as well as the smallest prediction bias. Regression predictions are second in predictive power. ARIMA predictions of revenues and ARIMA predictions of the independent variable as input for regression predictions of revenues (which can be used when an independent variable is not economically available or available only at a lag) are found to reduce substantially mean absolute yearly bias and mean absolute errors vis-a-vis the naive procedures. Considering the results, ARIMA bases seem to be potentially useful but not as generally applicable alternative(s) to the more traditional time series regression. Other characteristics of ARIMA predictions are briefly explored.
William R. Kinney (Sun,) studied this question.