An the AR model. The numbers in parentheses are the p-values of two-sided t-statistics for equal MSE. The two panels of Table 2 refer to the periods 1985, quarter 1?011, quarter 3, and 1985, quarter 1?008, quarter 2.A. Carriero, T. E. Clark and M. MarcellinoWe can draw six main conclusions from the RMSE results in Table 2. First, as might be expected, the accuracy of forecasts from the BMF and BMFSV models improves as more data on the quarter become available, and we move from month 1 to 2, 2 to 3, and 3 to month 1 of the next quarter. The gains look a little bigger with the move from month 2 to month 3 than from month 3 to month 1 of the next quarter. As a consequence, the accuracy of the nowcasting models relative to the AR baseline increases with the addition of more information on the quarter. For both the small and the large BMF and BMFSV models, the gain in RMSE in the full sample results rises from about 7 in month 1 to 18 in month 3 and 23?5 in month 1 of the next quarter. Somewhat surprisingly, the Diebold ariano est test does not often imply that the gains are statistically significant in the full sample, but it does imply more significance in the sample that ends before the depths of the crisis. Second, in the sample that ends in mid-2008 and thereby avoids the huge ARQ-092 site forecast errors of the severe recession, our BMF and BMFSV nowcasting models are often as accurate as or even a little more accurate (although not significantly so) than those of the BC survey, particularly in months 2 and 3 of quarter t. However, from month 3 of quarter t to month 1 of quarter t + 1, the BC forecasts improve in accuracy more so than do the model forecasts. As a result, in month 1 of quarter t + 1, the nowcasting models are generally less accurate than the BC results although not dramatically so. To shed some further light on the performance of the nowcasting models and BC forecast over time, Fig. 1 compares actual quarterly GDP growth (annualized) to point forecasts from the BC survey and our large BMFSV nowcasting model. The chart makes clear the improvement in accuracy that occurs with the addition of more data on the quarter–improvement that seems most noticeable around recessions (1990?991, 2001 and 2007?009). It also shows that, over some periods of time, the model is more accurate than the BC forecast, whereas, in others, the BC forecast is more accurate than the model. One period in which the BC forecast fares better is the most Y-27632 price recent recession, when it did a better job of picking up and projecting unprecedented declines in GDP growth. Accordingly, the third main conclusion from the RMSE results is that, in the full sample, the nowcasting models are somewhat less accurate than the BC forecast, perhaps because of the relative performance in the depths of the crisis, though the differences are not statistically significant. The challenge of beating a survey forecast with good nowcasting models is also evident in such studies as Banbura et al. (2013), who developed a mixed frequency factor modelbased forecast that is comparable with, but not quite as good as, the SPF in forecasts for 1995?2010. In light of the evidence in Chauvet and Potter (2013) that the advantage of some time series models over an AR model baseline stems largely from periods of recession, not during economic expansions (or normal times), we have checked the forecast performance of our models during just economic expansions (dropping out observations falling during Nat.An the AR model. The numbers in parentheses are the p-values of two-sided t-statistics for equal MSE. The two panels of Table 2 refer to the periods 1985, quarter 1?011, quarter 3, and 1985, quarter 1?008, quarter 2.A. Carriero, T. E. Clark and M. MarcellinoWe can draw six main conclusions from the RMSE results in Table 2. First, as might be expected, the accuracy of forecasts from the BMF and BMFSV models improves as more data on the quarter become available, and we move from month 1 to 2, 2 to 3, and 3 to month 1 of the next quarter. The gains look a little bigger with the move from month 2 to month 3 than from month 3 to month 1 of the next quarter. As a consequence, the accuracy of the nowcasting models relative to the AR baseline increases with the addition of more information on the quarter. For both the small and the large BMF and BMFSV models, the gain in RMSE in the full sample results rises from about 7 in month 1 to 18 in month 3 and 23?5 in month 1 of the next quarter. Somewhat surprisingly, the Diebold ariano est test does not often imply that the gains are statistically significant in the full sample, but it does imply more significance in the sample that ends before the depths of the crisis. Second, in the sample that ends in mid-2008 and thereby avoids the huge forecast errors of the severe recession, our BMF and BMFSV nowcasting models are often as accurate as or even a little more accurate (although not significantly so) than those of the BC survey, particularly in months 2 and 3 of quarter t. However, from month 3 of quarter t to month 1 of quarter t + 1, the BC forecasts improve in accuracy more so than do the model forecasts. As a result, in month 1 of quarter t + 1, the nowcasting models are generally less accurate than the BC results although not dramatically so. To shed some further light on the performance of the nowcasting models and BC forecast over time, Fig. 1 compares actual quarterly GDP growth (annualized) to point forecasts from the BC survey and our large BMFSV nowcasting model. The chart makes clear the improvement in accuracy that occurs with the addition of more data on the quarter–improvement that seems most noticeable around recessions (1990?991, 2001 and 2007?009). It also shows that, over some periods of time, the model is more accurate than the BC forecast, whereas, in others, the BC forecast is more accurate than the model. One period in which the BC forecast fares better is the most recent recession, when it did a better job of picking up and projecting unprecedented declines in GDP growth. Accordingly, the third main conclusion from the RMSE results is that, in the full sample, the nowcasting models are somewhat less accurate than the BC forecast, perhaps because of the relative performance in the depths of the crisis, though the differences are not statistically significant. The challenge of beating a survey forecast with good nowcasting models is also evident in such studies as Banbura et al. (2013), who developed a mixed frequency factor modelbased forecast that is comparable with, but not quite as good as, the SPF in forecasts for 1995?2010. In light of the evidence in Chauvet and Potter (2013) that the advantage of some time series models over an AR model baseline stems largely from periods of recession, not during economic expansions (or normal times), we have checked the forecast performance of our models during just economic expansions (dropping out observations falling during Nat.