A common problem policymakers, economists and forecasters face is the lack of essential economic data in real time. The information needed is often only available at lower frequencies and published with a considerable lag. This is the case with real gross domestic product (GDP), which is the single most relevant variable describing the path of the economy. GDP is used, together with inflation, to substantiate the direction of monetary policy. However the delay in GDP releases makes it difficult to predict the current state of the economy with accuracy. One way to improve this accuracy is to use higher frequency economic information more readily available in real time. Although some forecasters use judgement to incorporate higher frequency data into their forecasts, most models cannot incorporate this data because of three common challenges. First, higher frequency data are not released in a synchronous fashion, which means there tend to be gaps towards the ends of the sample. Second, information is not released at the same frequency; a quarterly projection model cannot use daily data. Finally, most traditional econometric models are unable to accommodate a large set of information.
This study addresses these forecasting challenges using a nowcasting model proposed by Giannone, Reichlin, and Small (2008). We adapt this model to the South African economy over the period 2005 to 2014. The framework uses daily, monthly, and quarterly data to forecast real GDP exploiting information in a data-rich environment. This leads to two major advantages: it incorporates a large information set and it permits disaggregation of the growth drivers to aid analysis of the forecast. For example, we can link recent movements in variables such as the Purchasing Managers’ Index (PMI) and consumer confidence directly to our expectation of real GDP growth providing a continually updated real-time view of the economy.
We compare the performance of our nowcasting model to consensus forecasts by Reuters and Bloomberg as well as to seven alternative models: a random walk model, two autoregressive (AR) models, two small-scale vector autoregressive (VAR) models, and two large-scale VAR models. The results show that the nowcasting model performs comparably to consensus forecasts despite not incorporating judgement as is the case with survey forecasts. Although our model does not include judgement, a valuable by-product of the nowcasting model is that the statistical methods can highlight potential anomalies. Thus through disaggregating growth drivers it can assist forecasters to use ex post judgement in assessing the severity of the anomaly. Importantly, the nowcast model also outperforms all other alternative models by a significant margin.
by 