基于机器学习算法的巨灾债券风险息差定价

Spread Pricing Model of Catastrophe Bonds Based on the Machin Learning

巨灾债券风险息差的实证研究目前已经相对成熟,但具体模型形式和变量选择依然存在一定的争议.将采用地震巨灾债券发行数据,建立巨灾债券的风险息差定价模型,分析风险息差的主要影响因素.首先,构建广义线性模型,发现本文提出的Logit风险附加值效果优异.然后,将广义线性模型的估计结果嵌入深度神经网络,提高广义线性模型的预测能力,同时提高神经网络的迭代效率.最后,比较了深度神经网络、嵌入广义线性模型的深度神经网络、随机森林、XGBoost以及支持向量回归等机器学习算法的定价效果,结果表明支持向量回归对巨灾债券风险息差的预测效果最佳.实证结果表明基于机器学习算法的巨灾债券定价模型明显优于传统回归模型,建议采用支持向量回归算法对巨灾债券风险息差进行定价.

The empirical research on the spread of catastrophe bonds is relatively mature,but there are still some controversies about the model form and variable selection.This thesis summarizes the pricing factors in historical literature and constructs new pricing factors.This thesis establishes a spread pricing model for catastrophe bonds based on the issued data of catastrophe bonds.First,this thesis constructs a Generalized Linear Model and finds that the risk load of Logit performs well.Then,results of the Generalized Linear Model are embedded in to Deep Neural Network.This method improves the prediction ability of the Generalized Linear Model,and improves the iteration efficiency of the Neural Network.Finally,this thesis compares the pricing effects of machine learning algorithms such as Deep Neural Networks,Deep Neural Networks embedded with Generalized Linear Models,Random Forests,XGBoost,and Support Vector Regression.Empirical results show that catastrophe bond pricing models based on machine learning algorithms are significantly better than traditional regression models.This thesis proposes to use Support Vector Regression to price spread of catastrophe bond.