Spatial dependence in stock returns - Local normalization and VaR forecasts

Abstract

We analyze a recently proposed spatial autoregressive model for stock returns and compare it to a one-factor model and the sample covariance matrix. The influence of refinements to these covariance estimation methods is studied. We employ power mapping as a noise reduction technique for the correlations. Further, we address the empirically observed non-stationary behavior of stock returns. Local normalization strips the time series of changing trends and fluctuating volatilities. As an alternative method, we consider a GARCH fit. In the context of portfolio optimization, we find that the spatial model has the best match between the estimated and realized risk measures.