Working paper
Forecasting with High-Frequency Predictors: a Combination of Shrinkage Estimation and Factor Analysis
Abstract:
This paper proposes the combination shrinkage estimation (\textit{LASSO} or \textit{EN}) and factor analysis to deal with the high-dimensionality problem in forecasting with high-frequency predictors. Our empirical study on GDP forecast during out-of-sample periods $2001Q1 \sim 2008Q4$ verifies the predictability of monthly macroeconomic and daily financial variables in both short- and long-run. Based on the same information, the the proposed models, especially based on \textit{EN}, do better in information selection and utilization, thus producing on average better forecasts than traditional \textit{MIDAS} models. Moreover, \textit{EN} turns out to be a better selection tool than \textit{LASSO} in case when predictors are highly-correlated.
This paper proposes the combination shrinkage estimation (\textit{LASSO} or \textit{EN}) and factor analysis to deal with the high-dimensionality problem in forecasting with high-frequency predictors. Our empirical study on GDP forecast during out-of-sample periods $2001Q1 \sim 2008Q4$ verifies the predictability of monthly macroeconomic and daily financial variables in both short- and long-run. Based on the same information, the the proposed models, especially based on \textit{EN}, do better in information selection and utilization, thus producing on average better forecasts than traditional \textit{MIDAS} models. Moreover, \textit{EN} turns out to be a better selection tool than \textit{LASSO} in case when predictors are highly-correlated.
Optimal Density Forecasts of the Equity Risk Premium (third year paper)
Abstract:
This paper evaluates the performance of density forecasts derived from univariate and multivariate econometric models, based on either point (mean) forecasts or the underlying explanatory variables which generate those point forecasts. Due to the "curse of dimensionality" issue in macroeconomic forecasting literature, we form our first conjecture that univariate models would outperform multivariate models, given a relative short data set. Moreover,
since point forecasts are obtained from the underlying explanatory variables containing private information, we expect that point forecasts can serve as a good substitute for explanatory variables in density forecast construction. The
application on equity risk premium forecasting verifies these two hypotheses.
This paper evaluates the performance of density forecasts derived from univariate and multivariate econometric models, based on either point (mean) forecasts or the underlying explanatory variables which generate those point forecasts. Due to the "curse of dimensionality" issue in macroeconomic forecasting literature, we form our first conjecture that univariate models would outperform multivariate models, given a relative short data set. Moreover,
since point forecasts are obtained from the underlying explanatory variables containing private information, we expect that point forecasts can serve as a good substitute for explanatory variables in density forecast construction. The
application on equity risk premium forecasting verifies these two hypotheses.