“. . . some of our statistics are too late to be as useful as they ought to be. We are always, as it were, looking up a train in last years Bradshaw [timetable]”
Harold MacMillan, UK Chancellor of the Exchequer, 1956.
NOWCASTING: AN OVERVIEW
Several excellent surveys of nowcasting (or closely related topics such as short-term forecasting) have recently appeared. These include Banbura, Giannone and Reichlin (2011), Banbura, Giannone. Modugno and Reichlin
(2013), Camacho, Perez-Quiros and Poncela (2013) and Foroni and Marcellino (2013). There are several approaches to nowcasting in the literature and the reader interested in full details of the main existing approaches (including bridge equations, factor models, mixed frequency VARs and MIDAS) is referred to these papers. Here we outline the general concepts underlying nowcasting before describing the particular set of methods that we use in this work.
At the most general level, nowcasting methods (like many forecasting methods) seek to find explanatory variables/predictors which are useful for predicting the dependent variable to be nowcast. Nowcasts are based on an econometric model linking the predictors to the dependent variable. For GDP growth there are a myriad of such predictors. For instance, Banbura, Giannone, Modugno and Reichlin (2013) use 23 predictors in their nowcasting model of US GDP growth including both hard variables such as industrial production and soft variables such as surveys of businesses.
Important econometric issues arise when nowcasting due to the fact that nowcasters want their predictors to be as timely as possible. For instance, when nowcasting 2013Q4 GDP growth having a predictor for which data becomes available in October or November, 2013 is very useful. A predictor which is not available until April 2014 (when the initial estimate of 2013Q4 GDP is released) is virtually useless. Furthermore, nowcasters typically update their nowcasts throughout the quarter as new information becomes available. The desire for timeliness and updating of nowcasts leads to two econometric issues which are treated in different ways by the different nowcasting approaches. These are: i) the dependent and explanatory variables have different frequencies and ii) the nowcasters’ data set typically has a ragged edge .
The mixed frequency issue arises since GDP is observed quarterly whereas many potential predictors for GDP (e.g. industrial production, some labour force statistics and PMI surveys) are available at a monthly level. In this paper, we will use Mixed Data Sampling (MIDAS) methods (described below) to address the mixed frequency issue, but several other methods exist (see, in particular, Foroni and Marcellino, 2013 for a survey of the various econometric methods used with mixed frequency data).
The ragged edge problem refers to the fact that the variables in the nowcasters’ data set typically have different release dates and, thus, at the end of the sample missing observations will exist for some of them.
Consider, for instance, nowcasting 2013Q4 Scottish GDP growth on 13 January, 2014. On this date, the value of December’s Bank of Scotland’s PMI was released and the nowcaster would wish to update the 2013Q4 nowcast. But on 13 January, data on Scotland’s labour market in December will not be available since the Bank of Scotland’s Report on Jobs is not released until 20 January. Hence, the December 2013 value for labour force variables will be missing on 13 January, 2014. Again, there are several ways of addressing this ragged edge problem, but we will use address them using MIDAS methods.