A neural network approach to measure real activities manipulation

• We use neural networks to measure three varieties of real activities manipulation.
• The multilayer perceptron approach outperforms traditional linear regressions.
• The multilayer perceptron approach performs better than self-organizing maps.
• Individual measures should be used instead of aggregated measures when applying linear regression.

A growing body of literature is examining the concept of real activities manipulation in various contexts. In these studies, the researcher typically models abnormal real activities and draws inferences based on the output measures. Thus, the results of the studies hinge critically on the underlying models. We contribute by examining alternative approaches to measure three varieties of real activities manipulation. Neural network models based on a self-organizing map and a multilayer perceptron are used in variation to a frequently used linear approach. The purpose of the study is to examine whether the neural network models outperform linear-based models in the detection of real activities manipulation. According to the results, the multilayer perceptron models are remarkably strong while the traditional linear models are underachievers. These results are specifically evident when common comprehensive measures are used.