基于机器学习的商誉减值预测模型

Predicting Goodwill Impairment Using Machine Learning

本文以我国2008-2020年A股非金融公司为样本,采用机器学习方法对商誉的减值进行了预测,以期帮助投资者识别商誉减值风险,缓解商誉减值对市场的冲击。实证结果显示,随机森林和XGBoost两种集成学习模型的预测表现要优于其他模型。进一步研究发现,随机森林和XGBoost预测的商誉减值公司在预测期的平均持有收益率要显著低于预测不会发生减值的公司,而且随着预测准确率的提升二者之间的差异逐渐向真实差异靠拢。这说明机器学习模型能够有效识别商誉减值风险,从而促进市场对其的吸收。最后,本文还发现由于影响公司行为的因素及其重要性会随时间变化,因此使用机器学习模型预测或者识别公司行为时,训练集的时间跨度并不是越大越好。

This paper compares a range of models in the machine learning repertoire,which are widely used in finance and accounting,in their ability to predict the occurrence of goodwill impairment,using a sample of non-financial listed firms in China from 2008 to 2020.The aim is to help investors predict the occurrence of goodwill impairment and then mitigate the impact of goodwill impairment on the stability of market.We find that the two ensemble classifiers,random forest and XGBoost,outperform all other classifiers.Furthermore,we also find that the average buy-and-hold return of the firms that are predicted to make goodwill impairment charges by random forest and XGBoost is significantly lower than the predicted non-impaired firms,and the return gap between the predicted impaired firms and the predicted non-impaired firms is approaching the real return gap with improved forecast accuracy.That indicates that the prediction model of goodwill impairment based on machine learning algorithm can effectively identify the goodwill impairment risk,thus promoting the market to absorb the goodwill impairment risk.Finally,we examine the effect of the scale of the training set on the predictive performance of the two ensemble classifiers and find that the increase of the scale of the training set does not necessarily improve the predictive performance because the factors of goodwill impairment and their importance change over time.