A procedure that has been used at the Swiss National Bank for selecting vector-autoregressive (VAR) models in order to forecast Swiss consumer price inflation is presented. In order to examine and improve the quality of the procedure, it is submitted to several modifications and the results are compared with one another. Combining forecasts substantially improves the quality of the forecasts. Models specified with respect to levels of variables are superior to those specified with respect to differences in variables. Bank loans and the monetary aggregate M3 are the most important variables for inflation forecasting. The optimized procedure reduces the root mean squared error (RMSE) of the inflation forecast to one third of the RMSE of a naive "no change" forecast over the period from 1987 to 2005.