|
主讲人 |
奚晋 |
简介 |
<p>Machine learning offers a promising set of tools for forecasting. However, some of the well-known properties do not apply to nonstationary data. I propose a simple procedure to extend machine learning methods to nonstationary data that does not require the researcher to have prior knowledge of which variables are non-stationary or the nature of the nonstationarity. I illustrate theoretically that using this procedure with LASSO or adaptive LASSO generates consistent variable selection on a mix of stationary and nonstationary explanatory variables. In a related paper, this approach was also shown to result in consistent estimation of principal components with the presence of nonstationarity. In an empirical exercise, I examine the success of this approach at forecasting U.S. inflation rates and the industrial production index using a number of different machine learning methods. I find that the proposed method either significantly improves prediction accuracy over traditional practices or delivers comparable performance, making it a reliable choice for obtaining stationary components of high-dimensional data. Another application to the FRED-MD macroeconomic dataset demonstrates that the approach offers similar benefits to those of traditional principal component analysis with some added advantages.</p> |
