Variable selection in discrete regression models applied to value investing
by
Richard Gerlach
University of Newcastle, Australia
Coauthors: Ron Bird (UTS, School of Finance and economics)

This paper presents a fully Bayesian technique for analysing and forecasting discrete count data using generalised linear regression models. This technique uses Markov chain Monte Carlo sampling and extends the variable selection method of Smith and Kohn (1996) as well as utilising the slice sampler of Mira and Tierney (2001). The method is advantageous as it can be applied in situations with a large number of potential variables to choose from. The technique is illustrated by an application to do with stock selection, based on one-year ahead forecast return performance, to enhance a simple value investment strategy. Value investing has been around since the 1930's and simply ranks stocks based on multiples such as book-to-market and price-to-earnings ratio. We enhance this strategy by selecting a significant subset of accounting variables (from a superset of 24 potential explanators), linked to stock performance by a logistic regression model, forecasting the direction of 'value' stocks (above or below the market return) and investing appropriately. Each year the investment portfolio is based on a new forecast model, estimated using the last five years of data. The results from this strategy are encouraging but not popular with classical economists.

Date received: April 4, 2002