Nowcasting GDP with Machine Learning — the Reserve Bank of New Zealand's Real-Time Test Against ~600 Economic Indicators

The Reserve Bank of New Zealand - Te Pūtea Matua published Discussion Paper DP2019/03, 'Nowcasting GDP using machine learning algorithms: A real-time assessment', by Adam Richardson, Thomas van Florenstein Mulder and Tugrul Vehbi. The paper tested a range of machine-learning methods - including regularised regression, random forests, boosted trees and neural networks - against roughly 600 macroeconomic and financial indicators, using genuine real-time data vintages of New Zealand's quarterly GDP growth rather than final, revised figures.

To build the evidence base for the Bank's own forecasting practice by testing whether machine-learning nowcasting methods could improve on a standard autoregressive benchmark for estimating quarterly GDP growth in real time.

The strongest-performing algorithms (boosted trees, support-vector regression and neural networks) reduced average nowcast errors by approximately 20-30% relative to a standard autoregressive benchmark, with several machine-learning models producing lower root-mean-square errors than both the AR benchmark and a dynamic factor model comparator. The work was subsequently published in the peer-reviewed International Journal of Forecasting.

The paper concludes the methods 'show the potential to improve the official forecasts of the Reserve Bank of New Zealand.' The published record demonstrates a rigorously tested, real-time-validated improvement in one technical forecasting exercise; it stops short of confirming that the machine-learning nowcasts have been formally incorporated into the Bank's published Monetary Policy Statement forecasts on an ongoing basis.