A0617
Title: High-dimensional time-varying forecast combination under general loss functions
Authors: Kenwin Maung - Rutgers University (United States) [presenting]
Abstract: In economics and finance, regression-based estimation of forecast combinations often relies on symmetric and quadratic loss functions. However, forecasting performance is increasingly assessed using more flexible loss functions. In particular, asymmetric loss functions are widely used in policy and risk management. Furthermore, many combination approaches often overlook time variation in predictive relationships. A unified approach is presented to estimating time-varying forecast combination weights under a broad class of loss functions via nonparametric M-estimation. Crucially, this class includes potentially non-differentiable and popular losses such as the lin-lin or asymmetric squared loss. The asymptotic properties of the local linear estimator are examined when the number of forecasts is small, and the selection and estimation consistency of a penalized version in high-dimensional settings are analyzed. An empirical application of inflation forecasting with asymmetric losses during the tumultuous post-Covid period highlights significant gains from using more general loss functions in estimating combination weights beyond mean squared error loss.
