Title: Inference and time series analysis with artificial neural networks
Authors: Gerhard Fechteler - Universität Konstanz (Germany) [presenting]
Abstract: Multilayer Perceptrons (MLP) have become a popular tool for nonlinear data analysis and proved to be useful for forecasting economic variables, outperforming not only linear models, but also other machine learning frameworks. Firstly, the asymptotic normality of the parameters of an MLP is discussed and an algorithm for estimating the asymptotic covariance even under misspecification of the network structure is provided. Moreover, also the marginal effects from an MLP are shown to be normally distributed. An algorithm to estimate the asymptotic covariance matrix is presented, such that confidence bounds for the marginal effects can be constructed. Based on the estimated distribution of the marginal effects, a local Granger causality test is proposed. It allows us to detect causal relationships between time series variables that are only present in local regions in the parameter space. To study high-dimensional time series data, a generalization of dynamic factor models (DFM) is proposed, which relaxes two critical assumptions of DFMs that are estimated via principal component analysis: linear dependency of the variables on the factors and orthogonality of the factors. The proposed framework is therefore less prone to misspecification.