基于监督学习模型的央行沟通测度

Measuring Central Bank Communication Based on Supervised Learning Model

央行沟通是受到市场广泛关注的重要叙事文本,如何从高维文本中有效提取关键信息是有待深入研究的科学问题.本文将Ke et al.(2019)提出的基于文本筛选和主题建模情感提取模型运用到央行沟通测度中,具有简单透明、可复制性强的优势.结合中文文本特征和中国货币政策多工具的框架,选取多个货币政策实际干预的变动值作为监督变量进而构建央行沟通指数,并基于广义货币政策规则对未来货币政策实际干预进行预测.研究结果表明,央行沟通文本信息有助于提供额外预测能力,并且与现有文献基于关键措辞、监督词典和LDA主题模型等文本分析方法构建的指数相比,本文构建的指数对未来货币政策实际干预的预测能力更好,尤其是长期预测表现更为优越.本文从预测角度验证了央行沟通引导政策预期的有效性,提供了根据不同预测指标提取文本大数据信息的可行方案.

Central bank communication is an important narrative text that receives a lot of attention from the market, and how to effectively extract key information from the high-dimensional text is a scientific problem to be studied in depth. In this paper, we apply the Sentiment Extraction via Screening and Topic Modeling method proposed by Ke et al. (2019) to measure central bank communication, which has the advantages of simplicity, transparency and replicability. Considering the characteristics of Chinese texts and the multi-instrument framework of China's monetary policy, we select the change values of several actual monetary policy interventions as supervised variables and then construct a central bank communication index, and forecast future actual monetary policy interventions based on generalized monetary policy rules. The results show that textual information on central bank communications helps to provide additional forecasting power. Compared with the indexes constructed by the existing literature based on text analysis methods such as keywords, supervised dictionaries and LDA methods, the index constructed in our paper has better forecasting power, especially with superior performance in long-term forecasting. We verify the effectiveness of central bank communication in guiding expectations from a predictive perspective, and provides feasible solutions for extracting textual information based on different target indicators.