機器學習與深度學習對企業財務危機預測能力之探討

CCUR > College of Business > Department of Accounting & Graduate Institute of Accounting > Thesis > Item 987654321/52988

Please use this identifier to cite or link to this item: https://irlib.pccu.edu.tw/handle/987654321/52988

Title:	機器學習與深度學習對企業財務危機預測能力之探討 Discussion on Machine Learning and Deep Learning on the Ability of Enterprise Financial Distress Prediction
Authors:	許家銘
Contributors:	會計學系
Keywords:	財務危機預測機器學習深度學習卷積神經網路
Date:	2023
Issue Date:	2023-10-04 14:08:58 (UTC+8)
Abstract:	近幾年因市場需求改變，許多企業皆發生破產與倒閉問題，而財務危機預測模型的建立能提早避免不必要的危機發生。本研究以台灣經濟新報(Taiwan Economic Journal, TEJ)收集2012年至2022年發生財務危機之台灣上下市櫃公司為主要研究對象並以1:1進行配對，其中包含19個財務變數及5個非財務變數。以兩階段做建構模型，第一階段使用決策樹C5.0與隨機森林來進行重要變數之篩選。第二階段再以卷積神經網路(CNN)、循環神經網路(RNN)與支援向量機(SVM)分別建立有效之財務危機預測模型。C5.0、RF搭配CNN與C5.0、RF搭配RNN所分析出之平均準確率均皆為85%以上，而C5.0、RF搭配SVM僅有80%左右，這表明了深度學習較機器學習更為優秀，對財務指標變數有更佳的預測能力。 In recent years, due to changes in market demand, many enterprises have experienced bankruptcy and bankruptcy problems, and the establishment of financial crisis prediction models can avoid unnecessary crises in advance. This study focuses on Taiwanese OTC companies that experienced financial crises from 2012 to 2022, collected by the Taiwan Economic Journal (TEJ), and paired them in a 1:1 ratio, including 19 financial variables and 5 non-financial variables. The model is constructed in two stages, with the first stage using decision tree C5.0 and random forest to screen important variables. In the second stage, effective financial crisis prediction models are established using convolutional neural networks (CNN), recurrent neural networks (RNN), and support vector machines (SVM), respectively. The average accuracy of C5.0, RF combined with CNN and C5.0, RF combined with RNN was all above 85%, while C5.0, RF combined with SVM was only about 80%, indicating that deep learning is better than machine learning and has better predictive ability for financial indicator variables.
Appears in Collections:	[Department of Accounting & Graduate Institute of Accounting] Thesis

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	345	View/Open

View Licence

Loading...