Title: Learning Financial Network with Focally Sparse Structure

Abstract: This paper studies the estimation of network connectedness with focally sparse structure. We try to uncover the network effect with a flexible sparse deviation from a predetermined adjacency matrix. To be more specific, the sparse deviation structure can be regarded as latent or misspecified linkages. To obtain high-quality estimator for parameters of interest, we propose to use a double regularized high-dimensional generalized method of moments (GMM) framework. Moreover, this framework also facilitates us to conduct the inference. Theoretical results on consistency and asymptotic normality are provided with accounting for general spatial and temporal dependency of the underlying data generating processes. Simulations demonstrate good performance of our proposed procedure. Finally, we apply the methodology to study the spatial network effect of stock returns.