经济开放与货币需求:国际金融风险及持币成本的测度

在中国开放经济体制下的基准货币需求模型中,本文将源于国际金融市场的持币成本设为遗漏潜变量,并构建特定的国际金融综合指数(CIFI)作为该潜变量的测度。借鉴机器学习与测度理论,本文利用对数误差修正模型提出了分步降维的CIFI构造算法,构造了长期CIFI和短期CIFI。结果表明,CIFI构造中的无监督降维步骤有助于减少高维金融数据中的冗余信息。实证分析发现,国际机会成本对中国货币需求具有规律性的前导影响,而在2007至2008年国际金融危机期间,央行的应急措施对长期CIFI所代表的非均衡冲击起到明显的阻截效果,对短期CIFI的影响基本是持续不变的。通过综合指数构造与宏观货币需求模型的算法连接,可以利用CIFI的构成结构从前导时间与影响强度两方面追踪冲击货币需求的国际金融风险的具体来源,这为宏观决策者监测国际金融市场提供了颇有规律的信息。在方法论上,本研究为如何利用模型监测国际金融市场影响宏观经济开辟了一条新路。

Openness and Money Demand:Measuring the Opportunity Cost Effects of International Financial Markets

Standard money demand models neglect the direct effects of economic openness. This omission is problematic when domestic opportunity cost variables fail to fully reflect the dynamics of international financial markets. Examining the effect of this omission is of great practical importance given the ever-increasing openness of China's economy. We propose composite international financial indices (CIFIs) to measure the latent variables that are omitted in standard money demand models. Using techniques from machine learning and measurement theory, we develop a novel model-based approach to construct CIFIs that combines both unsupervised and supervised dimension reduction methods. The choice of the popular error-correction model for the money demand function leads us to construct two types of CIFIs: long-run and short-run CIFIs. We collect a large set of around 100 financial input indicators to construct CIFIs using monthly data for the 1993M9-2015M6 period. These input indicators are obtained from 21 economies, covering almost all of China's major trading partners. The CIFI construction algorithm contains two stages of aggregation. First, it produces composite financial input indicators by aggregating groups of financial indicators. These groups are formed using clustering methods under the unsupervised learning approach. Second, it uses supervised dimension reduction methods to aggregate the composite financial input indicators following the principle of partial least-squares (PLS). The algorithm produces short-run CIFIs by targeting money growth rates, whereas it forms the target of long-run CIFIs using the error-correction term of standard money demand models. The second supervised aggregation stage sets the input indicators as leading indicators by construction, allows for dynamic dis-synchronization among them, and performs dynamic backward selection of different lags to make the dynamic input forms of the leading indicators as simple as possible. Concatenation is imposed on the resulting CIFIs during regular data updates. Experiments with CIFI-enhanced money demand models yield positive outcomes. Our key findings are as follows: (i) We find strong evidence of the effects of foreign opportunity costs on China's money demand based on the statistical significance and constancy of the coefficients of CIFIs and overall comparisons of model explanatory power; (ii) the effect of the short-run CIFIs is particularly robust, as evidenced by the 2007-2008 US-led financial crisis; however, in the enhanced error-correction term of the long-run CIFIs, a temporary coefficient variation toward insignificance is observed, which is interpreted as resulting from the emergency measures taken by the People's Bank of China in response to the crisis; (iii) model performance comparisons of the CIFIs produced with and without the first step of unsupervised dimension reduction show the necessity of this step in that it helps reduce redundant information in large financial datasets; (iv) tracing the compositions of CIFIs back to individual financial input indicators yields various patterns and features that enable the identification of the sources of the aggregate foreign opportunity cost effects.The explicit links between disaggregate input indicators and aggregate CIFIs provide valuable tools for policymakers to monitor external financial shocks from different geographical regions and markets and assess their aggregate risks in real time. Our CIFI algorithm opens a novel route of model-based composite construction. This route also sheds light on why the conventional route of principal component-based factor analysis is insufficient to construct composite indices for macro-modeling.