Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Statistics and Actuarial Sciences

Supervisor

Reg J. Kulperger

2nd Supervisor

Hao Yu

Joint Supervisor

Abstract

In this thesis, we propose a systematic approach called the doubly adaptive LASSO tailored to time series analysis, which includes four specific methods for four time series models, respectively:

The PAC-weighted adaptive LASSO for univariate autoregressive (AR) models. Although the LASSO methodology has been applied to AR models, the existing methods in the literature ignore the temporal dependence information embedded in AR time series data. Consequently, the methods may not reflect the characteristics of underlying AR processes, especially, the lag order of AR models. The PAC-weighted adaptive LASSO incorporates the partial autocorrelation (PAC) into the adaptive LASSO weights. The PAC-weighted adaptive LASSO estimator has asymptotic oracle properties and a Monte Carlo study shows promising results.

The PAC-weighted adaptive positive LASSO for autoregressive conditional heteroscedastic (ARCH) models. We have not found any results in the literature that apply the LASSO methodology to ARCH models. The PAC-weighted adaptive positive LASSO incorporates the PAC information embedded in squared ARCH process into adaptive LASSO weights. The word positive reflects the fact that the parameters in ARCH models are non-negative. We introduce a new concept named the surrogate of the second-order approximate likelihood, and propose a modified shooting algorithm to implement the PAC-weighted adaptive positive LASSO computationally. The PAC-weighted adaptive positive LASSO estimator has asymptotic oracle properties and a Monte Carlo study shows promising results.

The PLAC-weighted adaptive LASSO for vector autoregressive (VAR) models. Although the LASSO methodology has been applied to building VAR time series models, the existing methods in the literature ignore the temporal dependence information embedded in VAR time series data. Consequently, the methods may not reflect the characteristics of VAR time series data, especially, the lag order of VAR models. The PLAC-weighted adaptive LASSO incorporates the partial lag autocorrelation (PLAC) into the adaptive LASSO weights. The PLAC-weighted adaptive LASSO estimator has oracle properties and Monte Carlo studies show promising results.

The PLAC-weighted adaptive LASSO for BEKK vector ARCH (VARCH) models. We have not found any results in the literature that apply the LASSO methodology to VARCH processes. We focus on the BEKK VARCH models. The PLAC-weighted adaptive LASSO incorporates the PLAC information embedded in the squared BEKK VARCH process into the adaptive LASSO weights. We extend the concept of the surrogate of the second-order approximate likelihood, and propose a modified shooting algorithm to implement the PLAC-weighted adaptive LASSO computationally. We conduct a Monte Carlo study and have preliminary results from the study.

Keywords: Time series, financial time series, data mining, oracle property, LASSO, adaptive LASSO, doubly adaptive LASSO, positive LASSO, PAC-weighted adaptive LASSO, PAC-weighted adaptive positive LASSO, PLAC-weighted adaptive LASSO, autoregressive, AR(P), autoregressive conditional heteroscedastic, ARCH(q), vector autoregressive, multivariate autoregressive, VAR(p), vector ARCH, multivariate ARCH, VARCH(q), analytical score, analytical Hessian, quadratic approximation, surrogate to approximate likelihood, S\&P 500, Nikkei.


Share

COinS