以類神經網路建構風險值模型

Abstract

本研究利用2008/01/04至2018/09/28台灣加權指數進行分析與評估風險值之預測效果。研究成果與貢獻如下：一、延伸GARCH模型，假設即期波動度與非預期變動、歷史波動度呈非線性關係，透過類神經網路來捕捉更多非線性槓桿、波動叢集等效果。二、針對不同模型，分別利用最大概似法、反向傳播演算法等進行參數估計與訓練網路。實證結果顯示新模型對於波動度具有較好之配適能力。三、與變異數法、歷史模擬法、GARCH 模型比較，在95%信心水準下新模型所計算之風險值具有較低之穿透率。因此新模型所計算之風險值應可有效改善企業投資時所需提撥之準備金，主管機關亦可透過此模型來訂定投資人所需付出之保證金，皆可降低信用風險與穩定金融市場。本研究可提供上述產業評估風險時較為精準、客觀與較有效率之工具。

We analyze the VaR prediction by using the TWII data from 2008/01/04 to 2018/09/28. The contribution and results are as following: First, to extend the GARCH model, we assume spot volatility, unexpected volatility and historical volatility have non-linear relationship. By training neural network, we capture more non-linear lever effects and cluster volatility effects. Second, compared with different VaR models, we use Max Likelihood method to estimate the parameters and Backpropagation to train the neural network. The results show that the new model fits the volatility better than others. Third, compare the new model with other methods, VaR values predicted by new model have lower ABLF values. Therefore, the VaR values evaluated by new model can improve the reserve fund when the enterprise invests. The financial authority also can set the security deposit by using new model. This study can provide the abovementioned industrial a precise and objective tool to evaluate the risk.