Predicting Local Returns with Macroeconomic Variables

Predictability of the return on the market portfolio is a well established fact. This study shows that predictability is a more general phenomenon and it extends to return indices of the U.S. states. At the state level, the consumption trend deviation of Lettau and Ludvigson, and the collateral ratio of Lustig and Van Nieuwerburgh can predict short-term and long-term state-level returns. The state-level unemployment rate is also a significant predictor, as suggested by the incomplete markets model of Constantinides and Duffie. JEL codes: C34, E21, G12 ∗The paper was benefited by comments by Elena Andreou, Andros Kourtellos, Alok Kumar and the seminar participants at the Department of Economics at the University of Cyprus. I am responsible for all remaining errors. †E-mail: [email protected]. Mail: Department of Finance, 102 Mendoza College of Business, Notre Dame, IN 46556. Telephone: (574) 631-9322. Predicting Local Returns with Macroeconomic Variables 2 This paper investigates stock market predictability at the U.S. state level. As Cochrane (1999) points out, predictability of the aggregate stock market return is one of the new facts in finance. This aggregate return is the return on a market portfolio. A natural question is whether predictability of returns extends from the market portfolio to other less aggregated portfolios. If predictability is a robust new fact in finance, it should emerge in other portfolios that include a subset of the assets from the market portfolio. This is the goal of the current paper: investigate if predictability is present using returns of state-level portfolios. This is the first paper to test for predictability at the state level. In order to predict state-level returns, I use marcoeconomic variables. Recent papers argue that using financial ratios, like the dividend-price ratio, to predict returns poses many econometric challenges.1 Financial ratios are endogenous because they include the stock price, which also affects returns. The endogeneity problem is mitigated when the researcher exploits our understanding of the macro economy to find predictors, which do not include the stock price. In particular, I use the cay of Lettau and Ludvigson (2001), the hy of Lustig and Van Nieuwerburgh (2004), and the state-level unemployment rate. The importance of these three predictors can be explained by consumption-based asset pricing models. Lettau and Ludvigson (2001) show that the deviation of U.S. consumption from its long-run trend with U.S. wealth can predict the market return. This deviation called cay captures the expectations of individuals about the future path of returns. A positive cay implies that consumption today is above its long-run level. Then, it has to be the case that on average people expect the returns on their wealth to be higher in the future. Lustig and Van Nieuwerburgh (2004) identify the collateral ratio hy, the log ratio of home wealth to income, as another macroeconomic predictor of the aggregate stock return. As hy decreases, the collateral becomes relatively more scarce, the consumption growth of an individual becomes more perceptible to idiosycratic income shocks, the variance of consumption growth increases, and individuals ask a higher return at the market level. The significance of the collateral mechanism is stronger in an incomplete market setting. Also, the collateral ratio can be interpreted as a measure of financial distress. 1The work of Campbell and Yogo (2004), Hansen and Tuypens (2004), Hjalmarsson (2004) and Mark and Sul (2004) are excellent discussions of the difficulties in uncovering stock-market predictability with aggregate U.S. data using financial ratios that include the stock price. Predicting Local Returns with Macroeconomic Variables 3 A novel feature of the study is to recognize that there is no full risk-sharing among the U.S. states, which indicates that the state-level unemployment rate can be a useful predictor. Asdrubali, Sorensen and Yosha (1996) show that only 75% of the shocks on gross state product are smoothed across the states. Similarly, Athanasoulis and Van Wincoop (2001) find that financial markets and federal fiscal policy reduce the uncertainty of state income only by 44%. Markets are therefore incomplete across the U.S. states. Constantinides and Duffie (1996) show that in an incomplete markets setting, where idiosycratic shocks are not diversifiable, agents ask for a hefty premium to hold stocks. People fear stocks because they tend to do badly during recessions, which are periods with high unemployment. Korniotis (2005) shows that an asset pricing model which includes incomplete markets at the state level can rationalize a higher equity premium, than the complete markets model. The local return of state i includes the stocks all the firms with their headquarters in state i for the period from 1982(Q1) to 1999(Q1). The local index of state i is affected by the conditions in the national stock market; its correlation with the market return is 0.82.2 To predict this component of local returns I use the cay of Lettau and Ludvigson (2001). The local return should also be affected by the local macroeconomic conditions in state i. This component of local returns can be captured by the state-level collateral ratio and by the state-level unemployment rate. The prediction regressions also include a state-level dividend-price ratio since it is one of the most successful financial predictors. The empirical analysis estimates one-quarter ahead forecasting regressions for the local return r, the excess of the local return over the Treasury Bill return, r − rf , and the idiosycratic component of the local return. The last one is defined as the excess of the local return over the market return, r − rm. As in Fama and French (1988), regressions on long-term returns are also considered. Two types of such regressions are estimated: one on the long-term return over the Treasury Bill return, rK − rKf , and one on the long-term idiosycratic component of the state return, rK − rKm. The paper demonstrates that state-level returns are predictable. In particular, the one-quarter-ahead forecasting regressions reveal that the dividend-price ratio becomes insignificant in the presence of the cay residual and the unemployment rate. The trend deviation cay predicts the market component of the local return since it is significant in the regressions for r and r−rf , but it is insignificant in the regression 2This number is reported in Table 1, Panel A. Predicting Local Returns with Macroeconomic Variables 4 for r − rm. The collateral ratio is also an important predictor. It is most significant in predicting the idiosycratic component of the local returns, because it can capture financial distress at the state level. Finally, the state unemployment rate is statistically significant in all three regressions. The long-run regressions reveal that state returns are the most predictable at the 3 year horizon. Similar to the one-quarter-ahead forecasting regressions, the dividend-price ratio is insignificant because most of its information is captured by the residual cay and the state unemployment rate. It is found that the trend deviation cay can only predict rK − rKf , and the unemployment rate is an important predictor only for rK − rKm. The collateral ratio though is statistically significant for both the rK − rKf and rK − rKm returns. The rest of the paper is organized as follows. Section 1 includes more details on the predicting variables. Section 2 describes the data set. Section 3 presents the one-quarter-ahead forecasting regressions. Section 4 includes the results for the regressions with long-term returns. Finally, Section 5 concludes the discussion. 1. The Forecasting Variables The main goal of the paper is to investigate predictability at the state level using marcoeconomic variables. Nevertheless, the dividend-price ratio is included in the analysis because it is one of the most successful financial ratios in predicting returns. Its presence in the forecasting regression will reveal if the macroeconomic variables are good predictors over and above the predictability that might come from the dividend-price ratio. The dividend-price ratio In U.S. data one of the most successful variables in forecasting long-horizon returns is the log dividendprice ratio introduced by Campbell and Shiller (1988). Their analysis uses the definition of the stock return rt+1, rt+1 ≡ log (Pt+1 +Dt+1)− log (Pt) , where P and D denote the price and the dividend of the stock. By log-linearizing rt+1, they derive the Predicting Local Returns with Macroeconomic Variables 5 following accounting identity between the log dividend-price ratio and future returns and dividends: dt − pt = constant +Et ( ∞ X j=1 ρ (−∆dt+1+j + rt+1+j) ) , (1) where d = logD, p = logP , ρ is the discount rate evaluated at the steady state and it equals 1/ (1 + exp(d− p)), and ∆d is the growth of dividends. Given the dividend growth, if future returns are expected to grow, then the dividend-price ratio should increase today. Therefore, there should be a positive relation between dt − pt and future returns rt+1+j. The dividend-price ratio suffers from endogeneity problems because it includes the log of the stock prices, log (Pt), which is also part of the definition of the log return rt+1. The dividend-price ratio is stationary, even though it is highly persistent, and it can forecast longhorizon returns very well. As Cochrane (2005) notes, its importance builds with horizon as it captures decade-to-decade movements as well as business-cycle movements. Therefore, in long-horizon forecasting regressions, the R of the regression and the significance of the dividend-price ratio increase with the forecasting horizons.3 The consumption trend deviation cay Lettau and Ludvigson (2001) obtain the deviation of consumption from its long-run trend with total wealth using the budget constraint of a representative agent: Wt+1 = (1 +Rw,t+1) (Wt − Ct) , where W represents total wealth including both financial and human wealth, Rw is the return on total wealth, and C is consumption. They log-linearize this budget constraint around the steady state and solve it forward. Then, by taking expectations conditional on information known at period t, they show that the consumption-wealth ratio has the following form: