A0908
Title: From reactive to proactive volatility modeling with hemisphere neural networks
Authors: Karin Klieber - Oesterreichische Nationalbank (Austria) [presenting]
Philippe Goulet Coulombe - Université du Québec à Montréal (Canada)
Mikael Frenette - UQAM (Canada)
Abstract: Maximum likelihood estimation (MLE) is reinvigorated for macroeconomic density forecasting through a new neural network architecture with dedicated mean and variance hemispheres. The model features several key ingredients to make MLE work in the context of predicting short-time series with vastly overparameterized models. First, the hemispheres share a common core at the entrance of the network which accommodates various forms of time variation in the error variance. Second, volatility emphasis constraints and a blocked out-of-bag reality check are introduced to avoid overfitting in both conditional moments. Third, the algorithm handles large data sets both computationally and statistically. Ergo, the hemisphere neural network (HNN) provides proactive volatility forecasts based on leading indicators when it can, and reactive volatility based on the magnitude of previous prediction errors when it must. Point and density forecasts are evaluated with an extensive out-of-sample experiment and benchmark against a suite of models ranging from battle-hardened stochastic volatility specifications to more modern offerings like Bayesian additive trees and Amazon's DeepAR. In all cases, HNN fares well by providing timely mean/variance forecasts for all targets and horizons, and as such, provides an effective way to quantify uncertainty surrounding deep learning-based macroeconomic forecasts.
