Forecasting Stock Prices Using a Hierarchical Bayesian Approach

The Ohlson model is evaluated using quarterly data from stocks in the Dow Jones Index. A hierarchical Bayesian approach is developed to simultaneously estimate the unknown coefficients in the time series regression model for each company by pooling information across firms. Both estimation and prediction are carried out by the Markov chain Monte Carlo (MCMC) method. Our empirical results show that our forecast based on the hierarchical Bayes method is generally adequate for future prediction, and improves upon the classical method. Copyright © 2005 John Wiley & Sons, Ltd. key words autoregression; hierarchical mixture priors; MCMC; prediction; simultaneous estimation INTRODUCTION Recent development in the security valuation literature has provided a model that relates the stock price to its book value and expected future earnings. It includes the work of Bernard (1995), Feltham and Ohlson (1995), Lang and Lundholm (1996) and Ohlson (1991, 1995). These studies develop a logically consistent framework for thinking about equity valuation using accounting data. The primary objectives of this paper are to empirically evaluate the adequacy of the security valuation model and to use it to forecast stock prices. The security valuation model has been developed based on a single firm. The empirical literature in both accounting and finance is based primarily on classical statistical techniques. In this paper, we apply an innovative statistical method, a hierarchical Bayesian (HB) approach that allows improved estimation of the regression coefficients by sharing information across firms. Using 14 years of quarterly stock price data, accounting book values and expected future earnings for 28 companies included in the Dow Jones Industrial Average, we show that the forecast based on the HB model is consistently superior to those obtained using the classical approach. The original Ohlson model proposes that the stock price is a linear function of the company’s book value per share and expected excess earnings per share for the following four periods with normally distributed innovation terms. Each company has its own coefficients; we use bi = (bi1, . . . , Journal of Forecasting J. Forecast. 24, 39–59 (2005) Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/for.933 * Correspondence to: Lynn Kuo, Department of Statistics, University of Connecticut, Storrs, CT 06269-3120, USA. E-mail: [email protected] 40 J. Ying, L. Kuo and G. S. Seow Copyright © 2005 John Wiley & Sons, Ltd. J. Forecast. 24, 39–59 (2005) bi6)¢ = (bi,1, . . . , bi,6)¢ to denote the regression coefficients of the intercept, book value, each of the expected excess earnings for the following four periods for the ith company with i = 1, . . . , n. The model can be described as follows for all t = 0, . . . , T;