초록 열기/닫기 버튼
본 연구는 단기이자율 확률과정의 특성을 결정하는 핵심 파라미터인 수준효과를 포함하는 단기이자율 확률변동성모형에 대한 효율적인 추론전략과 MCMC(Markov chain Monte Carlo)를 이용한 베이지언 사후표본추출 알고리듬을 제시한다. 수준효과를 설명하는 CEV형태의 확산함수에 포함된 파라미터들의 약식별성과 단기이자율의 단위근에 가까운 지속성으로 인한 파라미터들 간의 강한 상관관계는 단기이자율의 안정성을 특징짓는 수준효과 파라미터들에 대한 식별과 효율적인 추정을 어렵게 한다. 본 연구는 이에 대한 해결방법으로 모수재설정 단기이자율 확률변동성모형과 베이지언 MCMC알고리듬을 제시하고 모의실험을 통한 성과평가를 함께 제시한다. 3개월 만기 미국재무부채권 수익률 주간자료에 대한 실증분석을 통해 본 연구에서 제시하는 모수재설정 모형이 기존 연구에서 제시된 모형설정들보다 수준효과 파라미터 식별과 추정에 효율적임을 확인할 수 있었다. 또한 모수재설정 모형을 이용하여 얻은 수준효과 파라미터의 크기에 대한 베이지언 가설검정 결과는 수준효과의 크기가 0.5라는 귀무가설을 강하게 지지하는 것으로 나타났다.
We propose an efficient Bayesian inference strategy and a new MCMC algorithm for stochastic volatility models of short term interest rates with level effect. The strong correlations among parameters caused by weak identifiability of level effect parameter and near-unit root persistency of short rates makes difficult the identification and efficient inference of the key parameter, the level effect parameter of CEV diffusion function, characterizing stationarity of short term interest rates. Our strategy for efficient Bayesian inference for stochastic volatility models with level effect is a kind of centered parameterization of moving the level effect component which is difficult to identify, included in the diffusion function of CEV observation equation to the log-volatility state equation. Our centered parameterization is contrast to the non-centered parameterzation suggested in Strickland et al. (2008), Kastner and Frühwirth-Schnatter (2014), etc. as an efficient inference strategy for stochastic volatility models. Considering that basic stochastic volatility model with level effect is a model with the characteristics of partial NCP in itself, and that the information of the observed data may not be sufficient to identify the level effect parameter, the non-centered parameterization of Strickland et al. (2008) which moves the parameters of state equation to the observation equation, can make it even more difficult to identify the level effect parameter. The reparameterized model proposed in this study can solve the problem of weak identification problem of level effect parameter and the problem of inefficient inference due to the strong correlation between level effect parameter and mean parameter of log-volatility state equation. We validate the performance of our reparameterization strategy and Bayesian inference algorithm through both simulation experiments and empirical analysis comparing the efficiency of posterior sampling for the level effect parameter with that of the suggested parameterizations in the existing literatures. Empirical analysis results on weekly data of 3 month U.S. T-bill yields reveal that our reparameterized model for Bayesian inference performs better in the identification and efficiency of posterior sampling of the level effect parameter than existing reparameterized models. The better performance in posterior sampling can facilitate efficient Bayesian hypothesis tests for parameters of interest. A Bayesian hypothesis test result for the size of level effect parameter supports the null hypothesis of 0.5, which implies the drift-induced stationarity of 3 month U.S. T-bill yields.
키워드열기/닫기 버튼
Stochastic Volatility Model, Markov Chain Monte Carlo Algorithm, Level Effect, Reparameterization, Effective Sample Size
