In this paper, I study the drop of real GDP volatility which has been observed in the United States during the postwar period. This paper thoroughly estimates how much sectoral shifts contributed to this phenomenon called the Great Moderation. In a short section, Stock and Watson (2003) find that this contribution is negligible, however, their data is disaggregated only up to 10 sectors. Blanchard and Simon (2001) come to the same result. Using a new estimation method and more disaggregated data, I find that sectoral shifts contributed between 15% and 30% to the great moderation. Moreover, I find that if in the year 1949 sectoral shares had been equal to what they were in 2005, then the conditional and unconditional standard deviation of GDP growth would have been, on average, 20-25% lower in the postwar period. Finally, I find that the shift out of durable goods production has significantly stabilized real GDP growth. As a methodological contribution, I show how to use the particle filter to estimate latent covariance matrices when they follow a Wishart autoregressive process of order one. I use this in order to get, for each observation period, an estimation of the covariance matrix of the sectoral growth rates. Since real GDP growth is the sum of these sectoral growth rates weighted by the sectoral shares, it is then straightforward to use these covariance matrices to express the conditional variance of GDP growth in each period as a function of sectoral shares. Computing the unconditional variance of GDP growth as a function of sectoral shares is a bit more involved, but also quite easy using Monte Carlo simulations. My methodology to estimate covariance matrices is preferable to alternatives like estimating a multivariate GARCH model or using a Nadaraya-Watson estimator for the following reasons: The multivariate GARCH model has undesirable properties for the Monte Carlo simulations and involves estimating a large number of parameters. The Nadaraya-Watson estimator, on the other hand, does not guarantee to give positive definite covariance matrices due to the limited number of observations available for estimating the relatively big covariance matrices.
Great Moderation Sectoral Shifts Stochastic Volatility Wishart Autoregressive Process Particle Filter ARCH-GARCH Bayesian Estimation