‘Pooling’ models generates better output gap estimates – paper

Estimating the size of the output gap is highly challenging, given the drawbacks in models and data, but pooling models can help, researchers from the Central Bank of Colombia find.

Their paper, Measuring the unmeasurable, notes that different models tend to produce different estimates. Results also change over time owing to structural changes in the economy or revisions in the data. Lags in publication of national accounts data add further problems.

The authors – Karen Pulido-Mahecha, Sergio Restrepo-Ángel and Franky Juliano Galeano-Ramírez – add that developing economies face larger shocks and often bigger data revisions.

They propose a new way of evaluating models in real time. They focus on criteria such as stability over time, accuracy of the core inflation forecast, and the model’s ability to explain potential GDP responses to economic shocks.

They use their evaluation method to test different models. They find some, such as structural vector autoregressions, do a better job of forecasting inflation but show greater instability over time. By contrast, semi-structural models are more stable but less accurate.

With these results in hand, the authors propose a new way of nowcasting the output gap and potential GDP. This involves pooling models by using “cluster algorithms”, a form of machine learning. They find that the approach produces better real-time estimates for Colombia if a large weight is placed on a semi-structural model that includes adaptive expectations and labour market indicators.

“We believe that this approach offers a valuable tool for central banks to mitigate real-time estimation errors,” the authors say. “It is easy to implement and allows central banks to incorporate more meaningful real-time estimates into their monetary policy decisions.”