Statistical Support of Forensic Auditing

Abstract

During a three year period, $4.38 million of gold was stolen from the vaults of Sammartino's House of Diamonds. A civil suit between the gold's owner and its insurance company centered on when the gold was stolen. The time of the loss was critical to the insurer's contention that it was not liable because the gold was missing prior to the inception date of the policy. The insurer's attorneys retained an accounting firm to conduct a forensic audit (a financial audit to investigate fraud). The key to this investigation was reconstructing a daily gold inventory balance by sampling sales invoices. Since the audit's sampling methodology was open to question, we were retained to use various statistical techniques to verify the process. To test the results of the forensic audit, we constructed two formal hypotheses: The first to test for nonrandomness in the selected sample observations; the second to test if the amount of sales dollars selected in the sample was proportional to the amount of sales dollars in the monthly population. To accomplish this, we utilized a one-sample runs test, time-series charts, regression analysis, and interval estimations. We found that the sampled invoice transactions exhibited no evidence of nonrandomness or bias and appeared to be representative of the population. Hence, these results supported the inventory reconstruction obtained by the accounting firm.

Forensic accounting: a blend of knowledge

“…The descriptive data-mining techniques are used to describe the underlying association in the data such as association rules and clustering (Baesens et al , 2009). Predictive techniques are used to predict value for a certain target variable, such as credit scoring to predict repayment behaviour of loan applicants[4], and logistic regression models, both binary and multinomial logit models, are used for detecting manipulation such as dishonest insurance claims (Olinsky et al , 1996; Major and Riedinger, 2002; Artís et al , 2002; Caudill et al , 2005). Regression is used to predict the value of a continuous target variable such as stock price, credit loss and sales.…”

Fraud Detection by a Multinomial Model: Separating Honesty from Unobserved Fraud

“…Numerous studies develop techniques to identify or classify fraudulent claims. Predictive techniques are used to predict values for a certain target variable, such as credit scoring to predict repayment behaviour of loan applicants, and logistic regression models, both binary and multinomial logit models, are used for detecting manipulation such as dishonest insurance claims (Major and Riedinger, 2002;Olinsky et al, 1996). Artís et al (2002) find a significant portion of the claims that were previously classified as legitimate contain omission errors, and thus are likely to be fraudulent.…”