آرشیو

آرشیو شماره ها:
۴۴

چکیده

هدف: این پژوهش بر آن است که با اضافه کردن متغیرهای مربوط به شبکه مالی، عملکرد الگوی درخت تصمیم تقویت گرادیان را که شاخص هایش با الگوریتم بهبودیافته گرگ خاکستری بهینه شده است، با الگو های منتخب در حوزه پیش بینی درماندگی مالی ارزیابی کند.روش: الگوی پیشنهادی این پژوهش روی داده های 123 شرکت تولیدی پذیرفته شده در بورس و فرابورس ایران در بازه زمانی 2015 تا 2021 اجرا شد. ابتدا شبکه مالی تشکیل شد و سپس با ترکیب متغیرهای مبتنی بر شبکه با برخی نسبت های مالی و با استفاده از الگوی درخت تصمیم تقویت گرادیان که شاخص های آن با الگوریتم بهبودیافته گرگ خاکستری بهینه شده است، درماندگی مالی شرکت ها پیش بینی شد.نتایج: الگوی درخت تصمیم تقویت گرادیان با اضافه شدن متغیرهای مربوط به شبکه مالی هم ازنظر دقت و هم ازنظر خطای نوع یک عملکرد بهتری در مقایسه با دو الگوی k نزدیک ترین همسایه و رگرسیون لجستیک از خود نشان داد. شرکت های با مرکزیت بینابینی و مرکزیت درجه زیاد، کمتر مستعد قرارگرفتن در شرایط درماندگی مالی هستند و برعکس. در شرایطی که نسبت فراوانی نمونه های هر طبقه از طبقات دیگر بسیار متفاوت باشد، استفاده از روش درخت تصمیم تقویت گرادیان بسیار کارآمد خواهد بود.نوآوری: برای نخستین بار متغیرهای مربوط به شبکه مالی با نسبت های مالی، ترکیب شد و ازطریق روش نوین درخت تصمیم تقویت گرادیان که شاخص هایش با الگوریتم گرگ خاکستری بهبودیافته بهینه شده، درماندگی مالی پیش بینی شد.

A Novel Approach to Predicting Financial Distress by Using Financial Network-Based Information and the Integrated Method of Gradient Boosting Decision Tree

This study aimed to evaluate the performance of the gradient boosting decision tree model, the parameters of which were optimized with the improved Gray Wolf Algorithm (GWO) by adding financial network-related variables via the selected models of predicting financial distress. The proposed model of this study was implemented on the data of 123 manufacturing companies admitted to the Tehran Stock Exchange and Iran Fara Bourse Co. (IFB) from 2014 to 2021. Initially, the financial network was formed and then, the financial distress of companies was predicted by integrating the network-based variables with financial ratios and using a gradient boosting decision tree model. The model of the gradient boosting decision tree had better performance in terms of precision and Type I error by adding Financial Network Indicators (FNI) compared to the two models of K-Nearest Neighbor (KNN) and Logistic Regression (LR). Companies with betweenness centrality and high degree centrality were found to be less prone to financial distress and vice versa. This is the first study to predict financial distress by using financial network-related variables integrated with financial ratio variables through the novel gradient boosting decision tree method, the parameters of which were optimized with the improved GWO.Keywords: Financial Distress, Financial Network, Gradient Boosting Decision Tree, Improved Gray Wolf Algorithm (GWO), Centrality Criteria. IntroductionPrevious studies on predicting financial distress have mainly adopted financial variables in financial statements as explanatory variables, while ignoring some other potentially useful information, such as financial network-related information. Disregarding such information for predicting financial distress is one of the significant gaps in the literature. Therefore, the present study aimed to evaluate the performance of the gradient boosting decision tree model, the parameters of which were optimized with the improved Gray Wolf Algorithm (GWO), along with the financial network-related variables, which showed another gap in the literature, and then compare its findings with the two recently widely used models of K-Nearest Neighbor (KNN) and Logistic Regression (LR). The following questions were posed in the present study.Can the integrated IGWO-GBDT model provide a better prediction of financial distress compared to the widely selected models of LR and KNN?Does the inclusion of financial network variables improve the performance of the integrated IGWO-GBDT model for predicting financial distress?Does the inclusion of financial network variables improve the performance of the widely used models of LR and KNN for predicting financial distress? Materials & MethodsThe sample of this study included all the manufacturing companies listed in the Tehran Stock Exchange and Iran Fara Bourse Co. (IFB) from 2014 to 2021, of which 123 companies were selected. The information of the last 30 trading days in each fiscal year was used to form the financial network variables. Finally, 8 financial variables were selected based on those adopted by Ebrahimi Sarvolia et al. (2018), including current ratio, net ratio of working capital to total asset, ratio of current asset to total asset, profit margin, return on assets, return on equity, book to market ratio, and size of company. Furthermore, the total debt ratio, a widely used variable in previous studies, was added to the financial variables. Regarding the financial network variables, the 4 variables of degree centrality, betweenness centrality, eigenvector centrality, and closeness centrality were selected similar to those selected by Montasheri and Sadeqi (2020) and Liu et al. (2019). More than one criterion was applied to measure financial distress in this study. Any companies that met at least one of the four mentioned criteria were considered distressed in that year. These criteria were subject to Article 141 of the Commercial Law, suffering losses for 3 consecutive years (Damoori & Hozhabrie, 2019), subject to Article 412 of the Commercial Law, referring to a debt ratio higher than 1 (Poorheidari & Koopaei, 2011), and had negative equity. FindingsThe findings indicated that the mean degree of centrality for companies with financial health was 0.203, while it was 0.193 for the distressed companies. Betweenness centrality was 0.007 for the distressed and non-distressed companies. Regarding eigenvector centrality, the means of healthy companies and those with financial distress were 0.083 and 0.104, respectively. Also, the values of closeness centrality were 0.525 and 0.496 for the mentioned companies, respectively. Table 5 shows that all the 3 models have performed quite well in terms of prediction of the class of healthy companies, while the error was below 10%. Still, the distinguished performances of these models were revealed when they could accurately and appropriately predict the class of financially distressed companies, which included only 12% of observations. The type I errors in the LR and KNN models were respectively 20 and 40%, bearing huge costs. In other words, these two models mistakenly predicted distressed companies as healthy ones in 20 and 40% of cases. However, the type I error of the proposed research model was only 5%, indicating high accuracy of the model. Table 5. Comparison of the performances of the 3 studied models with and without financial network variablesFN-GBDTGBDTFN-KNNKNNFN-LRLR 0.9820.960.8950.9060.9070.901Accuracy0.050.110.400.310.200.24Error 10.0130.0330.0650.0110.0780.086Error 20.9330.9890.9070.9160.9590.932AUC0.9680.9270.7480.8260.8580.833G-mean Discussion and Conclusion(1) Financial network variables can be employed to explore useful information in the financial network and improve the prediction performances of classifiers.(2) According to the definition of centrality, companies with high centrality, especially betweenness centrality and degree centrality, are less prone to financial distress and vice versa.(3) The improved GWO is a practical method for selecting the parameters of the gradient boosting model.(4) The gradient boosting decision tree model is highly efficient when the frequency ratio of the samples of each class is significantly different from those of other classes. The experimental findings revealed that the proposed model outperformed the other two models in predicting financial distress. This was confirmed by the adopted evaluation criteria.

تبلیغات