Private information of the Fed and predictability of stock returns

Abstract

This article investigates whether the Federal Reserve's (Fed's) private Gross Domestic Product (GDP) growth forecasts, as reported in the Greenbook of the Fed, contain information about future real and excess stock returns. I implement long-horizon regressions to analyse the predictive power of the Fed's GDP growth forecasts. The regressions conclude that the Fed's GDP growth forecasts can be used to predict long- and short-term stock returns. The size of the coefficient of the Fed's orthogonal GDP growth forecast indicates that 1% increase in the Fed's forecast predicts 2–4% decrease in real and excess stock returns. The regressions considering the size effect suggest that the predictive power of the Fed's GDP growth forecasts increases as the size of the portfolio decreases. A comparison of the Fed's forecasts and the commercial forecasts shows that the Fed's GDP growth forecasts contain information that does not exist in the commercial forecasts. I investigate the sources of the Fed's superior private information and predictive power. Analysis suggests that the source of the predictive power of the Fed's GDP growth forecasts is the private information about future surprise monetary policy actions embedded in them.

Notes

¹ Faust et al. (2004) find little evidence that Fed policy surprises convey superior information about the state of the economy. There are several empirical studies that provide reliable and consistent evidence that stock returns strongly react to monetary policy changes. Faust et al. provide the following explanation for the relation between stock returns and monetary policy:

… it may be that the policy surprise conveys information not about the state of the economy, but rather about the future course of policy, for which the FOMC has a natural advantage.

This article provides empirical evidence for the above explanation. I conduct further analysis about the information content of the expectations of the Fed. The analysis in Section VI shows that the Fed's GDP growth forecasts contain private information about future surprise actions of the Fed. Thus, I believe that the results of this article are in line with the findings of Faust et al. (2004) and Faust and Wright (2006).

² The Fed recently started to release GDP and inflation forecasts approximately 2 weeks after the Federal Open Market Committee (FOMC) meetings. For example, Table 1 in the minutes of the 28–29 October 2008 meeting displays the projections of the Fed for 2008, 2009, 2010 and 2011. The minutes were released on 19 November 2008.

³ Greenbook forecasts also include the forecasts of future inflation. Predictive regressions conclude that the Fed's inflation forecasts have predictive power. Predictive power of inflation forecasts are not analysed in this article because the correlation between the Fed's inflation forecasts and the commercial (Survey of Professional Forecasters (SPF)) inflation forecasts are very high suggesting that inflation forecasts of the Fed do not contain too much private information.

⁴ The analysis in this article can be seen as the first step to identify the asymmetric information between the Fed and the public as one of the reasons of effects of changes in monetary policy on stock returns. Faust and Svensson (2001) and Gurkaynak et al. (2005) construct theoretical models where the Central Bank possesses private information about the state of the economy. Tas (2006) constructs a learning model to analyse this hypothesis and the simulations show that the asymmetric information between the Fed and the public and the learning dynamics is an integral part of the effects of monetary policy on stock returns.

⁵ Gertler and Gilchrist (1994) argue that a monetary tightening, by increasing interest rates, worsen cash flow net of interest and thus firms' balance sheet positions. This decline in net worth can reduce a firm's ability to borrow. These credit constraints affect small firms, which are less well-collateralized, more than large firms. Therefore, stock returns of small firms are affected more by monetary policy actions since monetary policy alters firms' access to credit. Cooley and Quadrini (2006) show that when the economy is hit by monetary shocks, the response of small and large firms differs substantially, with small firms responding more than big firms. The higher sensitivity of small firms to monetary policy shocks derives from the fact that small firms take on more debt.

⁶ As mentioned at the Federal Bank of Philadelphia's web site, the Research Staff at the Board of Governors prepares projections using assumptions about monetary policy. Thus, the predictive power of the Fed's forecasts can be a result of the private information of the Research Staff about future actions of the Fed.

⁷ Fed's GDP growth forecasts have significant predictive power for nominal portfolio returns. The results of nominal portfolio returns are not displayed here to save space. The regression results for nominal portfolio returns are available from the author upon request.

⁸ These control variables are selected because they are shown in the literature to predict stock returns. The addition of these control variables makes sure that the predictive power of the Fed's GDP growth forecasts is caused by private information of the Fed, not by common factors (known by the public to predict stock returns).

⁹ Predictive power of the relative bill rate was shown before by Campbell (1991) and Hodrick (1992).

¹⁰ They also identify dividend yield as a good predictor, but Ludvingson and Lettau (2001) show that dividend yield does not have predictive power when cay and rrel are added into the regression. So, we do not use dividend yield as a control variable.

¹¹ http://www.econ.nyu.edu/user/ludvigsons/.

¹² I would like to thank George Aragon for suggesting orthogonalized forecasts.

¹³ http://www.phil.frb.org/econ/forecast/index.html.

¹⁴ Before 1992, the Greenbooks contain Gross National Product (GNP) forecasts instead of GDP forecasts.

¹⁵ Several studies like Fair and Shiller (1989) and Romer and Romer (2000) use different forecasts of the same variable as explanatory variables. Previous versions of this article contain analysis that uses gdpqtr1g and spfgdp1g as explanatory variables simultaneously. The results of previous versions are very similar with this article. gdpqtr1g is significant and spfgdp1g is not significant in that analysis suggesting that Greenbook forecasts contain all the information in the SPF forecasts and also additional private information. The orthogonalization methodology used in this article is more robust. The tables of that analysis are not displayed in this article to save space. The results of previous versions are available from the author upon request.

¹⁶ The growth rate is calculated as the percentage change between the current quarter GDP forecast and one quarter ahead GDP forecast.

¹⁷ This methodology has been extensively used in the predictability literature by studies such as Patelis (1997) and Ludvingson and Lettau (2001).

¹⁸ I would like to thank Arthur Lewbel for suggesting this methodology.

¹⁹ I would like to thank Steve Bond for drawing my attention to this possibility.

²⁰ Pagan (1984) investigates the properties of Two Stage Least Squares (2SLS) and Instrumental Variable (IV) regressions in the case of generated regressors. The regressions I am running in this article are the same type of model 2 that Pagan has analysed in his paper. Considering that the GDP growth forecasts are constructed using regressions, the method I am running is a 2-step estimator (2SLS). Pagan (1984) concludes that the 2-step estimator with predictor-generated regressors is perfectly efficient and yields correct inferences for H ₀: β = 0.

²¹ This section does not investigate value-weighted portfolio returns because size portfolios are investigated separately.

²² The results for the other size portfolios are available from the author upon request.

²³ I would like to thank Kenneth N. Kuttner for providing me the dataset.

²⁴ The Durbin–Watson (DW) statistics of the second regression specifications (b) in Panels A and C are 0.65 and 0.3, respectively.