FORECASTING TECHNIQUES FOR ANALYTICAL REVIEWS: AN EMPIRICAL INVESTIGATION OF LINEAR REGRESSION AND ARIM

Regression analysis is currently used in the analytical review phase of the audit process. Proper application of regression analysis requires that the data fulfill the necessary statistical assumptions in order to ensure valid interpretation of the results. The objectives of this study are to determine whether accounting data used by auditors in analytical review violate any of the regression assumptions, whether application of standard transformations can remove existing violations, and whether ARIMA predicts at least as well as regression using such data.;The research methodology entails (1) acquiring accounting data of various types of entities for which analytical review procedures were performed in audits, (2) constructing regression models, (3) performing standard tests for the regression assumptions of normality of error terms, homoscedasticity, nonautocorrelation, and noncontemporaneous correlation, (4) applying standard transformations on the data when violations are encountered, and (5) comparing the prediction results of the regression model with those obtained from ARIMA models constructed from the same data.;All assumptions except noncontemporaneous correlation are violated when applying regression analysis to the audited data. Transformations removed some but not all of the violations. Certain violations occurred more or less frequently and were more easily or less easily removed from the data, depending on the industry, type of account, and whether nonfinancial data were included in the model.;The comparison of relative prediction errors between regression and ARIMA models indicated no significant difference when predicting revenue or cost of sales accounts, but when predicting major expense accounts ARIMA had a significantly smaller prediction error.;The conclusions in this study are that assumption violations occur when auditors apply regression analysis to accounting data, these violations are not always removed via transformations, and use of a model containing such violations may affect the auditor's interpretations of results. ARIMA, which avoids the problem of assumption violations, appears to have predictive capabilities at least as good as regression, and may be a more appropriate audit tool.