A1084
Title: HARNet: A convolutional neural network for realized volatility forecasting
Authors: Xandro Bayer - University of Vienna (Austria) [presenting]
Nikolaus Hautsch - University of Vienna (Austria)
Rafael Reisenhofer - University of Bremen (Germany)
Abstract: The HARNet model is designed to bridge the conceptual gap between established parametric time series approaches for realized volatility and state-of-the-art deep neural network (NN) models. HARNets allow for an explicit parameter initialization scheme such that before optimization, the predictions of a HARNet are identical to those of a HAR model. The approach facilitates an in-depth analysis of the performance of different HARNets compared to HAR baselines. The role of loss functions is analyzed, different HAR baselines, initialization, stability of optimization and the interpretability of optimized models. Based on the analysis, specific guidelines for optimizing HARNets are derived.
