By Professor Christoph Schumacher
When the government releases their previous quarter GDP estimates, we always check to see how close our machine-learning predictions were. Prior to the pandemic hitting our shores and our economic fortunes, our forecasts have been very close to official figures, and in many cases spot on.
This time, however, our Q4, 2020 quarterly and annual growth predictions of around -4% and -3.7% respectively were considerably more pessimistic than the just-released values of -1.0% (quarterly) and -2.9% (annual). What could have caused this discrepancy? There are three points worth highlighting.
First, GDPLive is powered by a machine-learning ensemble algorithm that is very good in making predictions when there is sufficient past data for the algorithm to learn from. Things, however, get difficult if there isn’t enough training data. The lockdown happened in Q2, 2020 and until today, the government has only released Q2 and Q3 quarterly and annual GDP growth rates. GDPLive therefore only has two historic values to learn from since the economy was shut down — not enough for consistently good predictions. From a research point of view, it highlights areas were machine-learning needs further development as more traditional human-based forecasting models might still have the upper hand in predicting very sudden, unusual, data-poor events.
Second, during the lockdown, Kiwi’s consumption patterns changed almost overnight. One of the big changes was the large increase online spending brought about by retail closures and the need for social distancing. This trend is likely to continue even though we are, for most of the time, back to business-as-usual. GDPLive is partly powered by in-store sales data which means that a change in spending behaviour could negatively influence the prediction quality.
Third, the figures released by the government are estimates of GDP and often get revised at a later stage. These revisions can be substantial and are not that well-advertised. This means that we only know how good or bad our predictions are when the revisions have been made.
GDPLive is a world-first, machine-learning GDP predictor. Like us, it is trying to understand the impact of the events that unfolded in 2020. As more data becomes available, the more it can learn and the better the predictions will be.