Seung Mo Choi and Tara Iyer
Economists are increasingly turning to new technologies that help track indicators such as growth and inflation in real time to sharpen their forecasts and offer better input for policymakers.
Nowcasting, or forecasting of the present, is especially promising for developing economies where statistical authorities may not release indicators frequently. At the IMF, we have developed an approach that pairs high-frequency data with machine learning, a kind of artificial intelligence, to provide economic-growth nowcasts and help policymakers make better decisions.
Some developing economies release key indicators such as gross domestic product with lengthy delays. That adds to challenges for setting policy in times of rapid change, such as the onset of the pandemic or outbreak of war in Ukraine, because decisions must be made without crucial data. At times like these, nowcasting can offer projections that approximate economic activity a lot more quickly than official GDP data.
As the Chart of the Week shows, the nowcasting framework offered invaluable information for Botswana when pandemic closures hit in the second quarter of 2020. At the time, it projected output would contract by about 20 percent from a year earlier. That September, when the government published GDP figures, they showed an even sharper 24 percent contraction. This wasn’t far off from the framework projections, which moved in step with the actual data and accurately predicted the turning point.
In a new paper, we seek to narrow the gap between data availability and policymaking in sub-Saharan Africa by developing a framework to track real-time economic activity. Our nowcasting framework extracts signals from high-frequency figures available before official GDP is out.
The tool generates nowcasts by incorporating a broad range of increasingly popular machine-learning techniques that use an array of high-frequency economic indicators historically related to change in GDP. Tourist arrivals, for example, are more reliable predictors for tourism-dependent countries. For oil-exporting countries, such as Nigeria, GDP tends to move with crude oil prices.
Other nontraditional data inputs can include satellite imagery of nighttime lights, which tend to glow with greater intensity as economic activity increases, and of shipping vessels, used to track trade volumes and disruptions. Web searches can also help more accurately forecast tourist arrivals, as one IMF working paper showed in 2020.
Our research shows machine-learning algorithms are often more accurate than traditional econometric methods, especially for predicting turning points such as when an economy begins to bounce back to life after a recession or crisis.
The nowcasting predictors we used for our work on Botswana included currency values, stock prices, inflation, imports, consumer lending, electricity generation, tax revenue, and Google search volume for the country’s name (an indicator of future tourist visits). And because it’s one of the world’s largest diamond exporters, our framework also incorporates prices for the stones as well as economic growth for two of the largest destination countries for their shipments.
The Bank of Botswana is now producing its own nowcasts, after participating in an IMF workshop. Related courses on developing high-frequency indicators will be offered to help authorities around the world better track economic activity.