Forecasting birth trends in Ethiopia using time-series and machine-learning models: a secondary data analysis of EDHS surveys (2000-2019).

objectiveEthiopia, the second most populous country in Africa, faces significant demographic transitions, with fertility rates playing a central role in shaping economic and healthcare policies. Family planning programmes face challenges due to funding limitations. The recent suspension of the US Agency for International Development funding exacerbates these issues, highlighting the need for accurate birth forecasting to guide policy and resource allocation. This study applied time-series and advanced machine-learning models to forecast future birth trends in Ethiopia.

designSecondary data from the Ethiopian Demographic and Health Survey from 2000 to 2019 were used. After data preprocessing steps, including data conversion, filtering, aggregation and transformation, stationarity was checked using the Augmented Dickey-Fuller (ADF) test. Time-series decomposition was then performed, followed by time-series splitting. Seven forecasting models, including Autoregressive Integrated Moving Average, Prophet, Generalised Linear Models with Elastic Net Regularisation (GLMNET), Random Forest and Prophet-XGBoost, were built and compared. The models' performance was evaluated using key metrics such as root mean square error (RMSE), mean absolute error (MAE) and R-squared value.

resultsGLMNET emerged as the best model, explaining 77% of the variance with an RMSE of 119.01. Prophet-XGBoost performed reasonably well but struggled to capture the full complexity of the data, with a lower R-squared value of 0.32 and an RMSE of 146.87. Forecasts were made for both average monthly births and average births per woman over a 10-year horizon (2025-2034). The forecast for average monthly births indicated a gradual decline over the projection period. Meanwhile, the average births per woman showed an increasing trend but fluctuated over time, influenced by demographic shifts such as changes in fertility preferences, age structure and migration patterns.

conclusionsThis study demonstrates the effectiveness of combining time-series models and machine learning, with GLMNET and Prophet XGBoost emerging as the most effective. While average monthly births are expected to decline due to demographic transitions and migration, the average births per woman will remain high, reflecting persistent fertility preferences within certain subpopulations. These findings underscore the need for policies addressing both population trends and sociocultural factors.