How to Forecast Income Statement Items for Auditing Purposes

Generally accepted auditing standards require that auditors use analytical procedures (APs) at certain phases of an audit. The analytical procedure decision process involves comparing financial statement amounts to corresponding expectations of these amounts. A variety of forecasting methods exist for developing these expectations (predictions), ranging from data scanning to complex statistical models, such as regression analysis and time-series models. The current study examines the comparative predictive accuracy of three time-series models (the ARIMA model, the Census X-11 model, and the Holt-Winters model) and a proxy for data scanning (a random walk model) in relation to the AP decision process. The results indicate that data scanning has performed relatively well since the random walk model was not significantly outperformed in many cases. In addition, the Holt-Winters model might be considered a viable forecasting alternative to the ARIMA model in an AP decision context since it was as accurate as the ARlMA model in most cases. Public accounting firms continually search for ways to improve audit quality and reduce costs. Because analytical procedures (APs) contribute to these objectives, auditors use APs in the planning and final review stages of every audit performed in accordance with generally accepted auditing standards. Specifically, Statement of Auditing Standards No. 56 (SAS 56) requires the application of APs (1) to assist in planning the nature, timing, and extent of other auditing procedures and (2) as an overall review of the financial information in the final review stage of the audit. When planning the audit, APs may be used to increase the auditor's understanding of the business and to help identify areas of potential audit risk. During the final review stage, APs may be used to assess the reasonableness of audit conclusions and to evaluate financial statements. SAS 56 also permits the use of APs to supplement test of details in the substantive phase of an audit. In effect, the analytical procedure decision process involves comparing financial statement amounts to corresponding amounts predicted by a model. Thus, APs provide expectations of financial statement amounts for use in determining appropriate auditing strategies. A variety of forecasting methods exist for capturing relationships and converting data into expected values. These forecasting methods range from simple comparisons and ratios to complex mathematical and statistical models. Recent surveys, however, indicate that many auditors continue to rely on data scanning and other "naive" models as their primary APs. Prior research on APs has focused on a comparison of the accuracy of ARIMA (i.e., autoregressive integrated moving average), regression, and "naive" models in diverse settings. Although ARIMA is theoretically the most effective procedure since it subsumes both limited trend and regression analysis, it is not often used by auditors. This may be because of its extensive data requirements, the complex notational appearance, difficulty to interpret results, and larger costs of operationalization. Wheeler and Pany (1990) extended AP research by comparing four naive models, a regression model, and the X-11 model in a "best case scenario" setting. They concluded that the X-11 model was more accurate than the regression model and that both of these models dominated the four naive models. Their study, however, did not compare ARIMA or other time-series models with the X-11 model in their analysis. The purpose of our study is to conduct such an analysis of three time-series models and to compare their predictive accuracy with a proxy for data scanning (examining the prior period's data, which is the most frequently employed AP in current auditing practice) in relation to the AP decision process. The comparative predictive accuracy of: (1) the autoregressive integrated moving average (ARIMA) model, (2) the Census X-11 model, (3) the Holt-Winters exponential smoothing model, and (4) a random walk model (a surrogate for data scanning) are assessed across a variety of income statement items for both quarterly and annual data. …