Our research focuses on detecting financial reporting misconduct and derives acomprehensive misconduct sample using AAERs and intentional restatements.We develop a novel ensemble learning method, Multi-LightGBM, for h...Our research focuses on detecting financial reporting misconduct and derives acomprehensive misconduct sample using AAERs and intentional restatements.We develop a novel ensemble learning method, Multi-LightGBM, for highlyimbalanced classification learning. We adopt a human-machine cooperationfeature selection method, which can mitigate the limitation of incompletetheories, enhance the model performance, and guide researchers to develop newtheories. We propose a cost-based measure, expected benefits of classification,to evaluate the economic performance of a model. The out-of-sample testsshow that Multi-LightGBM, coupled with the features we selected, outperformsother predictive models. The finding that introducing intentional materialrestatements into our predictive model does not reduce the effectiveness ofcapturing AAERs has important implications for research on AAERsdetection. Moreover, we can identify more misconduct firms with fewerresources by the misconduct sample relative to the standalone AAERs sample,which is quite beneficial for most model users.展开更多
基金National Natural Science Foundation of China under[grant numbers 72071038,72121001].
文摘Our research focuses on detecting financial reporting misconduct and derives acomprehensive misconduct sample using AAERs and intentional restatements.We develop a novel ensemble learning method, Multi-LightGBM, for highlyimbalanced classification learning. We adopt a human-machine cooperationfeature selection method, which can mitigate the limitation of incompletetheories, enhance the model performance, and guide researchers to develop newtheories. We propose a cost-based measure, expected benefits of classification,to evaluate the economic performance of a model. The out-of-sample testsshow that Multi-LightGBM, coupled with the features we selected, outperformsother predictive models. The finding that introducing intentional materialrestatements into our predictive model does not reduce the effectiveness ofcapturing AAERs has important implications for research on AAERsdetection. Moreover, we can identify more misconduct firms with fewerresources by the misconduct sample relative to the standalone AAERs sample,which is quite beneficial for most model users.
