A Bayesian Approach to Variable Selection in Logistic Regression with Application to Predicting Earnings Direction from Accounting Information

This paper presents a Bayesian technique for the estimation of a logistic regression model including variable selection. The model is used, as in Ou and Penman (1989), to predict the direction of company earnings, one year ahead of time, from a large set of accounting variables from financial statements. We present a Markov chain Monte Carlo sampling scheme, that includes the variable selection technique of Smith and Kohn (1996) and the non-Gaussian estimation method of Mira and Tierney (1997), to estimate the model. The technique is applied to companies in the United States, United Kingdom and Australia. This extends the analysis of Ou and Penman (1989) who studied United States companies only. The results obtained compare favourably to the technique used in Ou and Penamn (1989) for all three regions.