The UCM Procedure

By default, the UCM procedure computes the standard errors of the series and component forecasts (both the filtered and smoothed estimates) by assuming that the estimated parameters are in fact the true parameters. Rodriguez and Ruiz (2010) describe a bootstrap-based procedure to compute the standard errors of the series and component forecasts that takes into account the uncertainty of parameter estimation. As an experimental feature in this release, you can request the computation of standard errors based on this bootstrap-based procedure by specifying the BOOTSTRAP option in the FORECAST statement. Subsequently, the confidence intervals for the series and component forecasts are based on these bootstrap standard errors. The algorithm that PROC UCM uses closely follows the first procedure described in Section 3 of Rodriguez and Ruiz (2010). Note that this bootstrap algorithm is computationally expensive. The computational burden increases with the number of bootstrap replications and is comparable to the computational burden of fitting the specified model as many times as the number of replications. Fortunately, these replications can be executed in parallel, and the UCM procedure can use multiple cores and multiple grid nodes (if they are available) to complete these calculations faster. For a single machine with multiple cores, the procedure automatically detects and uses all the cores. If a grid environment with multiple machines is available (with the appropriate SAS license), you must use the PERFORMANCE statement to supply the necessary information to the UCM procedure.