Modelling and Forecasting Unemployment Trends in Kenya Using Advanced Machine Learning Techniques

This study conducts a comprehensive analysis exploring the relationship between key macroeconomic indicators and the unemployment rate, alongside evaluating the predictive accuracy of modern regression models. The correlation analysis examines the association between unemployment rate percentages and five macroeconomic variables: real Gross Domestic Product (GDP) growth, gross public debt as a percentage of GDP, population size, government revenue as a percentage of GDP, and government expenditure as a percentage of GDP. The results highlight significant correlations, particularly the strong positive relationship between unemployment rates and gross public debt (% of GDP) (0.8417), while real GDP growth shows a weak correlation (0.0783), indicating that debt levels may be a more crucial determinant of unemployment variations in this context. Additionally, a comparison of modern regression models, namely Support Vector Regression (SVR), Neural Network Regression, and Bayesian Regression, is conducted based on their performance metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), R-Squared, Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC). Among the models, Support Vector Regression outperforms the others, with the lowest MAE (0.0823), RMSE (0.0878), and the highest R-Squared value (0.9915), along with notably favourable AIC (-100.2072) and BIC (-69.4268) scores. Neural Network Regression also delivers competitive performance with a slightly higher MAE and RMSE but a similarly strong R-Squared (0.9887). In contrast, Bayesian Regression exhibits weaker predictive power with higher error metrics (MAE = 0.2579, RMSE = 0.3109) and a significantly lower R-Squared (0.8806), AIC (28.0408), and BIC (36.8352). These findings underscore the efficacy of SVR in predictive modelling for macroeconomic datasets, suggesting its suitability for unemployment rate forecasting.