RDP 2024-04: Nowcasting Quarterly GDP Growth during the COVID-19 Crisis Using a Monthly Activity Indicator 4. Conclusion

We have made two important contributions to the factor monitoring and prediction literature related to Australia. First, we developed a monthly activity indicator with a very long history using a ‘supervised’ DFM model, with the explicit goal of providing policymakers with a timely snapshot on prevailing economic conditions. The time span covered by our MAI (45 years) is unmatched by previous work. Second, we have exploited the higher-frequency information imbedded in the MAI in a comprehensive nowcasting exercise covering a 35-year period and show statistically significant outperformance compared to standard benchmark models is possible. In this regard, our work is the first to apply a mixed frequency framework in a systematic manner.^[60] Further, we show that outperformance was greatest during the COVID-19 period, emphasising the benefit of using monthly data.

One curiosity related to our results is unlike many other works in which the prediction of GDP growth becomes more accurate as more data on the quarter comes to hand, our results show the opposite and get less accurate. We speculate this is related to increased parameter uncertainty due to estimating progressively larger models. Further, the COVID-19 crisis caused very large outliers (otherwise known as ‘leverage points’ due to their effect on the estimated regression fit) which can have a substantial effect on parameter estimation.

Despite this, our results do have some encouraging news for policymakers. By using MIDAS-based models incorporating the (timely) MAI, we show it is possible to predict Australian quarterly GDP growth more accurately during crisis periods (such as during the COVID-19 crisis) – a situation when accuracy is needed most. This comes about because the higher-frequency information contained in the MAI means the MIDAS models are quicker to detect abrupt changes, thereby giving policymakers more time to react.

One potential limitation of our work is that we do not redo the pre-selection step to determine the targeted predictor dataset at each time point in the out-of-sample prediction evaluation exercise. Instead, the ranking is done only once and using the full sample. This could bring some issues with our results; however, we are restricted by the unbalanced nature of our dataset which starts with only 17 series and is not as large as in other studies which have also mostly considered balanced datasets. But since we are not seeking to compare predictive accuracy of factor(s) extracted from a targeted predictor dataset to non-targeted predictor datasets this is probably less of a concern. Nonetheless, in future iterations, the targeted predictor dataset should be reviewed and updated as required to ensure it continues to contain only series that are informative about quarterly GDP growth.

In future work we intend to investigate three extensions: non-traditional data; sparsity; and nonlinearity. During the COVID-19 crisis greater use was made of non-traditional data such as mobility. We have not included any of these types of data in our extended dataset, although previous work suggests there might be merit for doing so (e.g. Choi and Varian (2012), who show internet searches can have predictive content). The challenge with some newer non-traditional datasets relates to their relatively short histories, making them harder to incorporate into analysis such as ours. In our work we used a two-step approach to identify the targeted predictor dataset and then extracted the factor afterwards. Recent work from Mosley, Chan and Gibberd (2024) suggests it might be possible to combine these two steps into one by incorporating sparsity using a form of regularisation into the estimation step. Lastly, the DFM is necessarily a linear model. Perhaps there are benefits to nowcasting from considering nonlinear specifications instead.

Footnote

Anthonisz (2021) is the only study using Australian data that is comparable to our work; however, his focus was nowcasting year-ended (annual) GDP growth. [60]