The dividend yield, corporate spread and term spread have been standard features of prior studies investigating the contribution of lagged macroeconomic information in pricing equities.1 The price of oil is also of interest to us as a factor, given the strong role oil shocks have played at turning points of postwar business cycles. An investigation by Hamilton (2000), allowing for a non-linear response function between GDP and oil disruptions, finds a clear link between petroleum supply disruptions and lower GDP. In an explicit examination of the impact of oil prices on equity markets, Jones and Kaul (1996) find that timely oil price information that precedes other economic series has a significant effect on real stock returns.
Although our factor model follows no specific theoretical model, it is well grounded in the ICAPM of Merton (1973) given that all our chosen variables are anticipated to forecast changes in future wealth and consumption flows. As will be discussed in Section 4, the lagged macroeconomic factor model also appears far better at pricing risk than the benchmark two parameter CAPM.
Pesaran and Timmermann (2002) argue that despite the empirical evidence indicating a time varying relationship between state variables and returns, research concentrating on the prediction of stock returns, still, to a large extent, employs models with time invariant parameters. Pesaran and Timmermann employ a simple reversed CUSUMQ test to identify structural change points and proceed to estimate a model relating S&P 500 monthly returns to a pre-specified set of lagged macroeconomic variables using only data after the most recent break has occurred. They find that the forecasting ability of their model greatly improves upon a comparable static specification, as well as a variety of alternative structural change models. Our proposed conditional model for predicting sector returns based on lagged fundamentals, although similar in spirit to theirs, explicitly accounts for time variation in the exposures to the state variables as well as the error variance in a dynamic Bayesian regression framework.
1 Chen, Roll and Ross (1986) use the level of the term spread in a cross sectional analysis of economic factors and stock returns. Lo and MacKinlay (1997) use the dividend yield, term spread and default spread in a predictive model