Asset Write-Offs Prediction by Support Vector Machine and Logistic Regression

Abstract

ABSTRACT The purpose of asset write-offs by a firm is to provide an accurate valuation of the firm and to reveal its true business performance from the perspective of economic conditions. However, the decision to write-off assets might be manipulated by the manager of the firm and thus misguide the public to an incorrect firm value. The aim of this study is to provide quantitative prediction models for asset write-offs based on both firms’ financial and managerial incentive factors. The prediction is achieved in two stages, where the first stage conducts a binary prediction of the occurrence of asset write-offs by a firm, while the second stage predicts the magnitude of such asset write-offs if they took place. The prediction models are constructed by support vector machine (SVM) and logistic regression for the binary decision of asset write-offs, and support vector regression (SVR) and linear regression for the write-off magnitude. The performances of different models are compared in terms of various criteria. Moreover, the bagging approach is used to reduce the variance in samples to improve prediction performance. Computational results from empirical data show the prediction performances of SVM/SVR are moderately superior to their counterpart logit/linear models. Moreover, the prediction accuracy varies with the distinctive types of asset write-offs.