The granting of credit is a necessary risk of doing business. If companies only accepted cash, sales would be negatively impacted. In a perfect world, all consumers would pay their bills when they become due. However, the fact is that some consumers do default on debt. Companies are willing to accept default risk because the value of defaults does not exceed the value of the additional sales generated. This creates an issue in regards to the valuation of uncollectible accounts. In order for a company to disclose the true value of its accounts receivable, it must establish an allowance for bad debt. Traditionally, companies estimate their bad debt expense and the related allowance for doubtful account by one of two methods: 1) As a percentage of total credit sales or 2) An aging of accounts receivable (that assesses a higher likely rate of default, the older the account becomes past due). By their very nature, these methods take into account only endogenous variables based on past experiences. For many years, the aforementioned methods of estimating bad debt were the only viable ways of determining the allowance for bad debts. However, with the explosion of technology and the easy availability of information, a more comprehensive method of determining bad debts seems appropriate. Neural network computer systems, which mimic some of the characteristics of the human brain, have been developed and may offer an alternative method for estimating the allowance for bad debt. These systems can predict what events may happen, analyze what did happen, and adjust the factor weights accordingly for the next set of event predictions. Thus, it is noteworthy to explore the use of neural networks to predict what a reasonable allowance for bad debt should be for an entity based on an array of interacting variables. Since, a neural network can incorporate both endogenous and exogenous variables one would expect to use such a system to develop a tool which gives a better estimation of the allowance for bad debt than the traditional approaches. In the current study, the findings indicate that neural networks over the balance of the time are better predictors of a company’s ending allowance for bad debt than regression. On a case by case basis, even when neural networks provide a less accurate estimate than regression, statistical analyses demonstrated the neural networks are a less volatile method and their predictions are less likely to result in a significant difference from actual allowance. Neither approach provides results that are exactly the same as the actual ending balance of the allowance for bad debt amount. Even though regression provides a more accurate estimate 45 percent of the time, this result is mitigated by two items: 1) On average, the absolute difference between actual and predicted is much lower when neural networks are used and 2) The standard deviation derived when using neural networks is only a third of the standard deviation derived from regression when applied to the absolute differences between the actual and predicted allowance.
|01 January 2011
|VCU Scholars Compass
|Virginia Commonwealth University
|Theses and Dissertations
|© The Author
Page generated in 0.0018 seconds