, Ghysels (2012), and Schorfheide and Song (2013) and McCracken and Sekhposyan (2012), both of which developed mixed frequency Bayesian VAR models, and Marcellino et al. (2012), which introduced a small scale factor model that allows for stochastic volatility in the Sulfatinib biological activity common and idiosyncratic components, and provided density forecasts. Relative to the existing partial model and full system approaches, the innovations in our approach include the use of Bayesian shrinkage and the inclusion of stochastic volatility. Bayesian shrinkage often improves the accuracy of forecasts from time series models, and it permits us to include a potentially large set of indicators, which some evidence (e.g. De Mol et al. (2008)) suggests should permit our model to achieve forecast accuracy comparable with that of factor models (full system methods). The use of direct-type estimation means that we do not need to model explicitly the conditioning variables. Moreover, with the univariate forecasting equation of our approach, we can easily allow for stochastic volatility, a feature that is important to the accuracy of density forecasts, mostly neglected so far in the nowcasting literature (we can also easily allow time varying regression coefficients). Finally, parameter time variation in the coefficients of the conditional mean can be also allowed, though we did not find it useful in the empirical application and therefore we do not discuss it explicitly (see Carriero et al. (2013) for details). The ability to include tractably a large set of indicators and stochastic volatility in the model gives our approach some advantages over other approaches in the partial model and full system classes. For example, with MIDAS methods, it is computationally difficult in general to consider more than a few indicators, whereas with factor model methods it is computationally difficult to include stochastic volatility unless only a small set of variables is used. However, admittedly, these other approaches can be seen as having some advantages over our proposed approach. For example, with MIDAS methods it is feasible to handle a large mismatch in frequencies, e.g. daily uarterly, whereas handling this with our model could be challenging (though still feasible), as it would generate a large number of additional regressors. Moreover, it could be feasible to design MIDAS models where the same LarotrectinibMedChemExpress LOXO-101 non-linear polynomial applies to different indicators, which would reduce the extent of the non-linearity and make the model more easily estimable even with several indicators. In practice, this would be an alternative to our use of Bayesian shrinkage to reduce the curse of dimensionality. The paper is structured as follows. Because data choices and availability play into our model specification choices, we first present the data in Section 2. Section 3 details our model and estimation method, and Section 4 introduces competing nowcasts. We then present results in Section 5. Finally, we provide some concluding remarks in Section 6.A. Carriero, T. E. Clark and M. MarcellinoThe data that are analysed in the paper and the programs that were used to analyse them can be obtained from http://wileyonlinelibrary.com/journal/rss-datasets 2. DataWe focus on current quarter forecasting of real GDP (or gross national product (GNP) for some of the sample) in realtime. Quarterly realtime data on GDP or GNP are taken from the Federal Reserve Bank of Philadelphia’s `Real-time data set for macroeconomists’., Ghysels (2012), and Schorfheide and Song (2013) and McCracken and Sekhposyan (2012), both of which developed mixed frequency Bayesian VAR models, and Marcellino et al. (2012), which introduced a small scale factor model that allows for stochastic volatility in the common and idiosyncratic components, and provided density forecasts. Relative to the existing partial model and full system approaches, the innovations in our approach include the use of Bayesian shrinkage and the inclusion of stochastic volatility. Bayesian shrinkage often improves the accuracy of forecasts from time series models, and it permits us to include a potentially large set of indicators, which some evidence (e.g. De Mol et al. (2008)) suggests should permit our model to achieve forecast accuracy comparable with that of factor models (full system methods). The use of direct-type estimation means that we do not need to model explicitly the conditioning variables. Moreover, with the univariate forecasting equation of our approach, we can easily allow for stochastic volatility, a feature that is important to the accuracy of density forecasts, mostly neglected so far in the nowcasting literature (we can also easily allow time varying regression coefficients). Finally, parameter time variation in the coefficients of the conditional mean can be also allowed, though we did not find it useful in the empirical application and therefore we do not discuss it explicitly (see Carriero et al. (2013) for details). The ability to include tractably a large set of indicators and stochastic volatility in the model gives our approach some advantages over other approaches in the partial model and full system classes. For example, with MIDAS methods, it is computationally difficult in general to consider more than a few indicators, whereas with factor model methods it is computationally difficult to include stochastic volatility unless only a small set of variables is used. However, admittedly, these other approaches can be seen as having some advantages over our proposed approach. For example, with MIDAS methods it is feasible to handle a large mismatch in frequencies, e.g. daily uarterly, whereas handling this with our model could be challenging (though still feasible), as it would generate a large number of additional regressors. Moreover, it could be feasible to design MIDAS models where the same non-linear polynomial applies to different indicators, which would reduce the extent of the non-linearity and make the model more easily estimable even with several indicators. In practice, this would be an alternative to our use of Bayesian shrinkage to reduce the curse of dimensionality. The paper is structured as follows. Because data choices and availability play into our model specification choices, we first present the data in Section 2. Section 3 details our model and estimation method, and Section 4 introduces competing nowcasts. We then present results in Section 5. Finally, we provide some concluding remarks in Section 6.A. Carriero, T. E. Clark and M. MarcellinoThe data that are analysed in the paper and the programs that were used to analyse them can be obtained from http://wileyonlinelibrary.com/journal/rss-datasets 2. DataWe focus on current quarter forecasting of real GDP (or gross national product (GNP) for some of the sample) in realtime. Quarterly realtime data on GDP or GNP are taken from the Federal Reserve Bank of Philadelphia’s `Real-time data set for macroeconomists’.