The Now-Casting model for quarterly GDP uses a large and heterogeneous set of predictors, including both 'hard' and 'soft' data (e.g., everything from unemployment statistics to consumer surveys). We use all of the data that is monitored by market participants plus any other potentially relevant series to extract a signal about the state of the economy. The estimation procedure exploits the fact that these data series, although numerous, co-move quite strongly so that their behaviour can be captured by few factors.
All Now-Casting output series are generated by a dynamic factor model developed by our founders. This type of model was chosen in order to cope with the so-called 'curse of dimensionality' (large numbers of correlated series) as it requires us to estimate only a limited number of parameters for a large dataset. The model assigns weights to the series, optimally exploiting the dynamic relationships among them.
The now-cast can be interpreted as that component of GDP growth which is highly correlated with all of the input data series: it disregards idiosyncratic information such as the weather, but it captures common signals given by all macroeconomic data releases including surveys.
The model is also designed to capture the essential characteristics of the now-casting problem in real time: that is, updating the estimates in relation to the flow of data releases throughout the day, the week and the month. This implies that the estimates are computed with missing observations in some of the series at period ends due to varying publication lags ("jagged edged" data sets). The methodology for now-casting incorporates a technical solution to this problem, as well as for the problem that the data are of mixed frequency, and that some input data are missing because of the limited availability of long time series for some variables in some countries. Specifically, we cast the model in state space, estimate the parameters by 'maximum likelihood', and apply the Kalman filter with a smoother to cope with jagged edged and mixed frequency data series.
The model allows us to compute a joint forecast of predictors and target series and, at each release, to calculate the surprise component of the published data release (what we call the 'news'). The revision of the now-cast of quarterly GDP growth can then be described as the product of the weight of each series (estimated on historical data) and the news for each release. This gives a transparent means of reading the flow of data releases.