SINGAPORE (July 2): Gross domestic product (GDP) numbers may be likely to emerge on the upside for 2Q20 relative to consensus forecasts, says Oxford Economics senior economist Tamara Vasiljev.

In a report dated 30 June, Vasiljev says the upside may be seen especially for the Eurozone, UK, and US economies. China and Japan’s economies are likely to remain closer to consensus forecasts.

Despite the surprise, uncertainties for 2Q’s GDP numbers remain high. This is due to a recession arising from the global outbreak that’s never been seen before.

However, Vasiljev warns against celebrating too early.

"Sluggishness in the economic variables that the algorithm captures is likely to slow the Q3 bounce-back, leaving year-end forecasts close to where we see them now,” she says.

“Come July, if GDP data releases positively surprise you, rein in your optimism: The worst might be over, but the way back probably won’t be as fast as we might have hoped when the Global Coronavirus Recession began,” she adds.

One of the key differences in this Covid-19-induced recession is due to its accelerated pace.

“While worsening sentiment would normally take months to produce a proportionate decline in real activity, the two have now tumbled almost at the same time,” says Vasiljev.

“The pandemic has introduced substantial compression of usual lags in economic relationships. Where once it took several months for worsening of sentiment to lead to worsening of real activity, the collapse of both sentiment and real activity was practically instant this time around,” she adds, citing the global lockdowns as part of the reason behind the compression of usual lags.

Unprecedented policy support from global governments has also rendered many financial indicators such as government bond yields and spreads, as well as stock market indexes “almost useless”.

“In these circumstances, we looked for an alternative source of insights and opted for the machine learning (ML) algorithms applied to high-frequency data to produce GDP nowcasts (GDP forecasts in real time),” she says.

“We used the ‘long short-term memory’ algorithm (LSTM) originally designed for natural language processing but also particularly suitable for economic data in which time interdependencies play an important role,” she adds.