Dynamic signal selection strategies

Abstract

This paper concerns the selection of a small number of predictors from a much larger set of potential stock signals. Selection is based on choosing predictors with high levels of dependency in stock returns. Eight different models of pairwise dependency are employed, including the Gaussian, t , Clayton, Frank and Gumbel copulas. In addition, predictive factors and returns are mapped to gamma, beta and standard bilateral gamma marginals with dependency constructed by using the magnitude of fractional common components from the same distributions. Each dependency model is used to select predictors. The predictions are used both directly and as required returns to
build a measure for the value of an invested dollar. Stocks are ranked by these two metrics daily. For the mean-reversion strategy, we take short positions in a quarter of the top-ranked stocks and long positions in a quarter of the bottom-ranked stocks. The positioning is reversed under momentum. Trading performance results are presented for a variety of dependencies and sectors over the period from June 2006 to January 2021. The selection procedures are observed to deliver a reasonable set of trading strategies.

Keywords: copulas, stock selection, factor selection, predictors, dependency models