Using a novel ensemble learning framework to detect financial reporting misconduct

Abstract

Our research focuses on detecting financial reporting misconduct and derives a comprehensive misconduct sample using AAERs and intentional restatements. We develop a novel ensemble learning method, Multi-LightGBM, for highly imbalanced classification learning. We adopt a human-machine cooperation feature selection method, which can mitigate the limitation of incomplete theories, enhance the model performance, and guide researchers to develop new theories. We propose a cost-based measure, expected benefits of classification, to evaluate the economic performance of a model. The out-of-sample tests show that Multi-LightGBM, coupled with the features we selected, outperforms other predictive models. The finding that introducing intentional material restatements into our predictive model does not reduce the effectiveness of capturing AAERs has important implications for research on AAERs detection. Moreover, we can identify more misconduct firms with fewer resources by the misconduct sample relative to the standalone AAERs sample, which is quite beneficial for most model users.