论文部分内容阅读
信用债估值是金融机构资产管理与风险控制的核心问题之一,主流估值方法诸如CreditGrades等模型无法捕捉违约事件新闻舆情及市场投资者情绪的变化。基于文本情感挖掘方法,把新闻舆情分为情感和语义两个维度,在CreditGrades模型基础上增加了一项信用点差,建立了量化修正的一个信用债估值改进模型。相比传统方法,新模型具有三点优势:(1)很多债券在日常市场活动中交易不够频繁,很多情况下某一段时间内市场没有交易,传统时间序列模型在对这种情况的预测会存在误差,而且新模型通过最近一段时间的文本挖掘,可以获取有关债券最新信息更加有效的预测债券价格走势。(2)传统方法过于依赖数字类的系统内推预测模型而无法规避系统风险和行业风险,新模型挖掘本文信息有效地弥补这个不足,具有较广普适性。(3)新模型具有自我进化功能,通过针对某一个金融领域不断更新模型中的情感词典、停用词词典、用户词典,模型的预测精度将会随着模型读取的文本数据的数量不断提高。实验表明,改进模型与原模型相比其估值的均方误差从0.134降为0.056,取得了明显高于传统方法的效果。
Credit debt valuation is one of the core issues of asset management and risk control in financial institutions. The mainstream valuation methods such as CreditGrades can not capture the public opinion of the default events and the changes of market sentiment. Based on the text emotion mining method, the news public opinion is divided into emotional and semantic dimensions. Based on the CreditGrades model, a credit spread is added and an improved credit debt valuation model is established. Compared with the traditional method, the new model has three advantages: (1) Many bonds do not trade frequently in daily market activities, and in many cases the market does not trade for a certain period of time. The traditional time series model will predict this situation Error, and the new model by the recent period of text mining, you can get the latest information on the bond more effective forecast bond price movements. (2) The traditional method relies too much on the system prediction model of the digital class to avoid the system risk and industry risk. The new model mining information in this paper can make up for the problem effectively and has a wide range of applicability. (3) The new model has the function of self-evolution. By continuously updating the sentiment dictionary, the stop word dictionary and the user dictionary for a certain financial field, the prediction accuracy of the model will increase as the number of text data read by the model increases . Experiments show that the mean square error of the estimated value of the improved model is reduced from 0.134 to 0.056 compared with the original model, and the effect is obviously higher than that of the traditional method.