Professor Norman R. Swanson, Rutgers University.

'Machine learning with big data: new empirical and theoretical findings'
January 14, 4.00pm-5.30pm
Zoom online seminar

Register now

Abstract

1. Forecasting Volatility Using Double Shrinkage Methods

In this paper, we propose and evaluate a shrinkage based methodology that is designed to improve the accuracy of volatility forecasts. Our approach is based on a two-step shrinkage procedure for extracting latent common volatility factors from a large dimensional and high-frequency dataset. In the first step, we apply either LASSO or elastic net shrinkage on estimated integrated volatilities, in order to select a finer set of assets that are informative about the target asset. In the second step, we utilise (sparse) principal component analysis on the selected assets, in order to estimate latent return factors, which in turn are utilised to construct latent volatility factors. All of our proposed factor-augmented models result in substantial predictive gains, as measured by both out-of-sample R square and mean absolute forecasting errors, and via the application of predictive accuracy tests. Forecasting gains of our proposed methods are observed in various scenarios with respect to different scales of estimates, different data sampling frequencies and different forecasting sub-periods. Additionally, our empirical findings suggest that the first step of our procedure, which utilises targeting shrinkage, plays a crucial role in the success of our methods, and the second step of our procedure relies on the underlying continuous-time factor structure for optimal predictive performance.

2. Consistent Estimation, Variable Selection, and Supervised Forecasting in Factor Augmented VAR Models

We develop estimation and variable selection methods for cases where the usual pervasiveness assumption does not hold in factor models with large numbers of variables. In order to motivate our methodology, we first establish that consistency of factor estimators is not achieved in natural settings where some variables in a large-scale dataset do not load on the factors in said model. We also provide results based on primitive assumptions allowing for consistent estimation of conditional mean models (for the purpose of forecasting) using FVAR models, unless our more general pervasiveness assumptions.

Biography

Norman R. Swanson is Professor of Economics at Rutgers University. His primary research interests include time series forecasting, big data, econometric methodology, financial econometrics and macro econometrics. He is a fellow of the Journal of Econometrics and the International Association of Applied Econometrics, and holds or has held various editorial positions at journals ranging from International Journal of Forecasting to Journal of Econometrics. He is a member of various professional organisations, including the Econometric Society, the American Statistical Association, the American Economic Association, the Canadian Economic Association, and the Society for Financial Econometricst. He has published over 100 peer reviewed articles in leading economics and and statistics journals ranging from Econometrica and Journal of Econometrics to Review of Economics and Statistics, Journal of Business and Economic Statistics, and the Journal of the American Statistical Association.


Further information: business-events@glasgow.ac.uk

First published: 7 January 2022

<< 2022