Artificial neural networks and the accounting method choice in the oil and gas industry

Abstract

On three occasions, accounting regulators considered eliminating full cost accounting as an acceptable method and at the same time requiring all oil and gas producing companies to adopt successful efforts accounting. In response, full cost companies appealed to the Securities and Exchange Commission to allow the continued use of full cost accounting arguing that companies using each method are different. They outlined three primary variables along which full cost and successful efforts companies can be differentiated: exploration aggressiveness, political costs, and debt-recontracting costs. Prior studies used these variables to explain the accounting method choice by oil and gas producers. Although these variables were significant from the standpoint of model development, the overall classification error rate for the traditional statistical models used by these studies has ranged from 28% to 57%. We propose that the high classification error is driven by strong non-linearities and high interactions among the posited variables and/or by the inability of binary statistical models to properly model the accounting method choice dynamics. On the other hand, the ability of artificial neural networks to model non-linear dynamics and to deal with noisy data make them potentially useful for this type of application. In this paper, we develop three supervised artificial neural networks (general regression, backpropagation, and probabilistic) to predict the accounting method choice by oil and gas producing companies. We compare the prediction accuracy generated by the artificial neural networks with those generated using logit regressions and multiple discriminant analysis. Consistent with the findings of prior studies, the overall prediction error for logit regressions and multiple discriminant analysis has ranged from 32% to 46%. Threelayer backpropagation and three-layer probabilistic networks performed no better than their equivalent traditional statistical models with the overall prediction error ranging from 24% to 43%. On the other hand, our three-layer general regression network performed much better with the overall prediction error ranging from 8% to 11%. More importantly, our general regression network performed extremely well in predicting both full cost and successful efforts companies.