Released on 16/01/2013:
• | The Kalman filter recursions of the state covariance matrix can now be computed via Chandrasekhar recursions rather than via the Riccati difference equation. For models with many state variables relative to the number of observed variables, the Chandrasekhar recursions can result in considerable computational gains. To get some idea of potential gains, the NAWM (Christoffel, Coenen and Warne, 2008) has 82 state variables, 18 observed variables, 18 structural shocks, 4 unique measurement errors, and 45 unknown parameters. For a computer running Windows XP SP3 and Matlab R2011b (single threaded) on an x86 processor, family 6 model 42 stepping 7 at 3092 MHz with 4GB of physical memory, it takes about 5 hours and 4 minutes (or 304 minutes) to obtain 250,000 draws with the random walk Metropolis algorithm, the AiM model solver, and the standard Kalman filter routine. On the same system, but with the Chandrasekhar recursions instead of the standard Kalman filter it takes about 2 hours and 48 minutes (or 168 minutes) to perform the same calculations. This is a speed gain of roughly 45 percent. Hence, the time savings can be substantial. |
• | The Sims, Waggoner and Zha's (2008) version of the modified harmonic mean estimator of the log marginal likelihood, where the weighting function is based on a truncated elliptical distribution, is now supported. |
• | All marginal likelihood estimators now come with numerical standard errors of the estimate based on the Newey and West (1987) method. |
Page url: http://www.texlips.net/yada/help/index.html?3_30.htm