UC San Diego

This dissertation consists of two essays on econometric methods for forecasting using high-dimensional data, along with an essay that integrates econometric analysis with mechanism design. In Chapter 1, we delve into the utilization of machine learning for forecasting with nonstationary data. We demonstrate that our proposed detrending method is easily implementable in high-dimensional settings and exhibits desirable properties both analytically and empirically. Specifically, this method automatically eliminates nonstationarity without requiring knowledge of the data generating process. The paper theoretically illustrates that LASSO and adaptive LASSO effectively identify useful predictors using detrended data, which shows that the proposed method preserves key information in the raw data. Regardless of the machine learning approach employed, the detrended data consistently produces prediction outcomes for forecasting inflation and the growth of industrial production that are comparable to or significantly better than those obtained using traditional detrending methods. Importantly, our proposed method generates robust prediction results in response to the COVID-19 shock.

Chapter 2 examines the proposed detrending method in principal component analysis. We analytically demonstrate that the detrended data enables principal components analysis to consistently estimate factors under specific conditions. In empirical applications, the proposed method showcases outstanding performance in characterizing large datasets. Factors estimated using the proposed method exhibit either comparable or significantly better performance compared to those estimated using traditional methods in forecasting exercises.

Chapter 3 examines the mechanism design problem for public goods provision from an econometric point of view by investigating asymptotic properties in a large economy. We introduce a new class of dominant-strategy incentive compatible and ex-post individually rational mechanisms. This class of mechanisms is both eventually ex-ante budget balanced and asymptotically efficient under certain specified conditions. The proposed class of mechanisms takes on a simple form and requires minimal knowledge of agents' valuation distribution, making it easy to implement for the public-good provider.