Can machine learning algorithms lead to more accurate nowcasts of Singapore's GDP?

Abstract

This paper investigates if machine learning (ML) models are able to produce accurate real-time nowcasts of Singapore’s GDP. Adopting dynamic factors with the Kalman smoother approach to simulate a realistic nowcasting exercise that incorporates the publication lags of variables, we evaluate the nowcasting accuracy of a suite of ML models, which includes the Elastic Net, Random Forest (RF), Gradient Boosted Trees (GBT), and Support Vector Regression (SVR), alongside simple ensembles. The models are assessed against the autoregressive model (AR) and the dynamic factor model (DFM) which act as benchmarks. In addition, we investigate if the ML models’ variable importance metrics are able to point policymakers towards the underlying economic indicators that would be useful to monitor, to address the common critique of ML models in lacking interpretability. We find that all four ML models were superior to the AR benchmark, but only the linear ML models (SVR & Elastic Net) outperformed the DFM benchmark and were robust in both calm and volatile periods. On the other hand, the tree-based models (RF & GBT) performed well only in calmer periods, which revealed their limitations in extrapolation. The simple ensemble models were also competitive and provided accuracy gains over most ML models individually. We also find that the ML models were able to appropriately assign high importance to variables in a way that conforms with Singapore’s economic structure and the business cycle literature. In sum, we find that the use of ML models helps to improve nowcasts of Singapore’s GDP.