Forecasting Ethiopian coffee export for risk management: an interpretable machine learning model

Coffee exports are a major pillar of Ethiopia’s export revenue, directly linked to livelihoods of many farming households and macroeconomic stability. However, coffee export revenue is simultaneously affected by changes in prices and exchange rates, making it highly volatile and complicating foreign exchange planning and risk management. Therefore, developing an accurate and interpretable forecasting method is vital for strengthening risk prevention for agricultural exports and foreign exchange management. Using monthly data from July 2012 to June 2025 from multiple sources, this study constructs a decomposition‑reconstruction hybrid model based on X- 11 decomposition, with coffee export revenue as the dependent variable and average unit price, exchange rate, and coffee productivity as main explanatory variables. Linear regression forecasts the trend component, while gradient boosting trees learn the residual component. Compared with multiple benchmark models, the proposed model performs better overall in RMSE, MAPE, and directional accuracy. Furthermore, monthly forecasts are incorporated as warning information into a scenario backtest. The results show that this forecast information helps reduce smoothed income volatility. Overall, the scenario backtest analytical process not only improves monthly forecasting performance but also provides actionable empirical evidence for monthly foreign exchange planning, and buffer- fund adjustments by the government.