A1605
Title: Improving index tracking and portfolio performance with regularisation
Authors: Dietmar Maringer - University of Basel (Switzerland) [presenting]
Sandra Paterlini - University of Trento (Italy)
Abstract: Historical data are often used for portfolio optimization when parametric distribution assumptions fail to capture key properties of asset returns. This, however, comes with the practical problem of choosing training window length: Long time series might include outdated observations that are no longer representing the immediate future, while short time series with more recent observations only are more prone to all sorts of inaccuracies typical for small sample sizes and lead to overfitting, in particular when the number of assets is large. Recent studies found that regularization is beneficial in the context of finding efficient portfolios. However, little guidance is usually given on calibration issues, and passive management techniques are rarely addressed. The aim is to fill these gaps. First, the elastic net, combining Lasso and ridge, is presented along different versions of index tracking and enhanced indexing. Next, an empirical study is performed analysing how different calibrations affect the reliability of optimized portfolios, considering different objectives, in-sample training window lengths, and out-of-sample investment horizons. We find that regularization can indeed improve the out-of-sample tracking error and other criteria, but the amount of improvement depends substantially on the calibration and the additional objectives considered on top of the tracking error.
