Andrea Carriero, Todd E. Clark, and Massimiliano Marcellino, "Bayesian VARs: Specification Choices and Forecast Accuracy," Journal of Applied Econometrics, Vol. 30, No. 1, 2015, pp. 46-73. This file details the (1) data and (2) RATS program files used to produce the results in the paper. The programs can be run in whatever directory this file has been unzipped (the data files will be in the same directory, which is what the provided programs require). Contact information: todd.clark@clev.frb.org, (216)579-2015. ********** DATA FILES: ********** The following files contain the various time series of data used in the article (all at the quarterly frequency), with rows corresponding to dates and columns to variables. The first column in each file provides the dates of the observation. The first row lists the variables. file #obs #var US.txt 447 18 CAN.txt 472 9 FRA.txt 473 9 UK.txt 423 9 For those who may it find it more convenient, the archive includes corresponding Excel spreadsheets (same file names but with extensions of .xls instead of .txt) with the same contents and layout. The RATS programs call the data from the Excel files. The file US.txt (or US.xls) provides the monthly time series for the U.S. used to produce all of the U.S.-based results in the paper. The other files provide the data used to produce results for other countries (these results are summarized in the paper but not included in the figures). The text files, which are ASCII files in DOS format, are zipped in ccm-data-txt.zip. The XLS files are zipped in ccm-data-xls.zip ************* PROGRAM FILES: ************* The program files consist of files ending in .prg that read in and transform the data, set up aspects of the model, call procedures that do most of the model estimation and forecast processing, and then compile summary statistics of the results for the given model. The procedure files have names ending with .src. The program files, which are ASCII files in DOS format, are zipped in ccm-programs.zip. Unix/Linux users should use "unzip -a". BVARnormalWishart.src Forms estimates and forecasts of VAR under Normal-Wishart prior and posterior, using Minnesota-type prior as specified in such sources as Kadiyala and Karlsson (1997). BVARnormalWishart.dumobs.src Forms estimates and forecasts of VAR under Normal-Wishart prior and posterior, with prior captured by dummy observations. Prior includes Minnesota component, sums of coefficients, and initial observations. directBVARnormalWishart.src Using direct multi-step spec., forms estimates and forecasts of VAR under Normal-Wishart prior and posterior, using Minnesota-type prior as specified in such sources as Kadiyala and Karlsson (1997). BVAR.ridge.withsim.src Uses ridge regression approach to estimate VAR and forecast under Litterman prior. fcmoments.src Processes draws of forecasts to compute point forecasts and average log predictive scores. baselineVAR.prg Using BVARnormalWishart.dumobs.src, this program generates forecasts from the benchmark VAR model. figure1.opttightness.prg Using BVARnormalWishart.dumobs.src, this program generates forecasts from the VAR specification optimized to pick (at each forecast origin) the marginal-likelihood maximizing overall tightness (lambda1). Figure 1 shows results for this approach compared to the benchmark VAR. figure2.optlag.prg Using BVARnormalWishart.dumobs.src, this program generates forecasts from the VAR specification optimized to pick (at each forecast origin) the marginal-likelihood maximizing lag order. Figure 2 shows results for this approach compared to the benchmark VAR. figure3.opttightandlag.prg Using BVARnormalWishart.dumobs.src, this program generates forecasts from the VAR specification optimized to pick (at each forecast origin) the marginal-likelihood maximizing tightness and lag order. Figure 3 shows results for this approach compared to the benchmark VAR. figure4.growth.prg Using BVARnormalWishart.src, this program generates forecasts from the model with most variables in growth rates form. Figure 4 shows results for this approach compared to the benchmark VAR. figure5.pseudoiterated.prg Using BVARnormalWishart.dumobs.src, this program generates pseudo-iterated point forecasts from the baseline VAR specification without simulation, using just the posterior mean coefficients (no simulation). Figure 5 shows results for this approach compared to the benchmark VAR. figure6.direct.prg Using directBVARnormalWishart.src, this program generates forecasts from models specified in direct multi-step form, without simulation. Figure 6 shows results for this approach compared to the benchmark VAR. figure7.Litterman.prg Using BVAR.ridge.withsim.src, this program generates forecasts from models specified in levels with a Litterman prior, simulating forecasts on an equation by equation basis. Figure 6 shows results for this approach compared to the benchmark VAR. figure8.rolling.prg Using BVARnormalWishart.dumobs.src, this program generates forecasts from the benchmark VAR model, but with a rolling sample for estimation, with sample size held fixed at that used to generate the first forecast. Figure 8 shows results compared to the benchmark model estimated recursively. figure9.7variables.prg Using BVARnormalWishart.dumobs.src, this program generates forecasts from a model with just 7 variables, using the same prior as in benchmark VAR model. Figure 9 shows results from the smaller model compared to the benchmark larger model.