Opening the Black Box: Nowcasting Singapore's GDP Growth and its Explainability

Multiple AI models outperformed traditional forecasting methods by 40-60%: The study found that machine learning approaches (penalized regressions, dimensionality-reduction models, and neural networks) significantly reduced forecast errors compared to standard economic benchmarks, demonstrating the practical value of AI for real-time economic assessment.

Explainability tools were essential for understanding model decisions: The researchers employed both model-specific and XAI (explainable AI) based tools to interpret which economic indicators drove predictions, addressing the "black box" problem and making AI recommendations actionable for policymakers who need to understand why forecasts are made.

Combining multiple models improved reliability: Rather than relying on a single AI model, the study used ensemble methods that weighted different models' contributions over time, providing more robust and interpretable predictions—a practical approach for high-stakes economic decisions where transparency about model uncertainty matters.

Rigorous validation procedures ensured trustworthiness: The framework used expanding-window testing, bootstrap confidence intervals, and statistical significance tests (Giacomilla-White tests) to validate results, demonstrating that responsible AI deployment requires multiple layers of verification beyond raw accuracy metrics.

Real-time economic monitoring requires handling uncertainty: The study's focus on prediction intervals and confidence bands reflects that responsible AI for policy decisions must quantify uncertainty, not just provide point estimates—critical for small, open economies vulnerable to external shocks.