Abstract

Accurate forecasting of Gross Domestic Product (GDP) growth is vital for effective economic planning, particularly in structurally volatile economies like Kenya. This study investigates the effectiveness of machine learning (ML) models in forecasting Kenyas GDP growth and compares their performance with traditional econometric approaches. Using quarterly time-series data from 2000 to 2024focusing on exchange rates, public debt, and tourism inflowswe implement and evaluate ARIMA, ARIMAX, VAR, Random Forest, XGBoost, Long Short-Term Memory (LSTM), and a Hybrid Ensemble model. Our results show that machine learning models, particularly Random Forest, outperform classical models, achieving the lowest RMSE (0.4238) and a modest positive R2 (0.0827), whereas traditional models yielded negative R2 values. SHAP (SHapley Additive exPlanations) analysis identifies external debt (lag 2) as the most influential predictor, followed by tourism inflows and exchange rate dynamics. Although the Hybrid Ensemble outperformed standalone time-series models, it did not surpass Random Forest, suggesting that ensemble strategies may require dynamic weighting or meta-learning enhancements. These findings underscore the potential of ML methods in macroeconomic forecasting by capturing non-linear relationships and improving model interpretability. Policy implications include the need for forward-looking debt management frameworks, tourism sector reforms, and balanced exchange rate strategies. We advocate for integrating interpretable ML tools into Kenyas national statistical and policy infrastructure to support adaptive and evidence-based decision-making. Future research should explore high-frequency and alternative data sources, advanced ensemble techniques, and cross-country comparisons to assess model generalizability across emerging markets.