Fence methods in model and moment condition selection.

摘要

In the first chapter of this paper, I implement the Fence methods, introduced by Jiang, Rao, Gu, and Nguyen (2008), as new procedures in model and moment condition selection for the Generalized Method of Moments (GMM) estimation. The Fence methods select models and moment conditions by carefully constructing a statistical "fence" to differentiate a set of good models from the wrong models. Then the Fence methods use a second criterion to select the optimal model out of the set of good models. In the setting of GMM estimation, the "J-statistic", which is a measure of lack of fitness, is used as the "Q" in the Fence methods to construct the statistical "fence". I apply these new model and moment condition selection methods to a Monte Carlo experiment in Andrews and Lu (2001). Simulation results show that the traditional Fence method and the Adaptive Fence method have superior performance than the MMSC-AIC, MMSC-BIC, and MMSC-HQIC procedures proposed by Andrews and Lu (2001). This better performance is more significant when we have a large panel dataset. Not only do Fence and the Adaptive Fence have higher chances of selecting the optimal model, their estimated parameters are also more precise and stable. Adaptive Fence method performs better than the traditional Fence method, since it optimally choose a tuning constant cn to maximize the probability of selecting the right model. The Simplified Adaptive Fence not only is easy in implementation but also has superior performance in large sample.;In the second chapter I apply the Fence methods to empirical test of asset pricing models of the US stock returns. Fence methods are used to select the best model among the Capital Asset Pricing Model, the Fama French Three Factor Model, and the Momentum Model. I also use the Fence methods to select moment conditions that better describe the historical stock return patterns. Compared with the MMSC-AIC, MMSC-BIC, and MMSC-HQIC criteria, the Fence methods select a model that has better performance not only in sample but also out of sample.