A Proposed Hybrid CNN–LSTM Framework for Forecasting Botswana’s Rough-Diamond Export Unit Values with Explainable AI

Abstract

Advanced forecasting models can represent complex nonlinear and temporal relationships, but their black-box nature may limit transparency and practical application in policy and strategic decision-making. This concern is particularly relevant to Botswana, where rough-diamond export performance has significant implications for fiscal planning, foreign-exchange earnings and macroeconomic stability. This paper proposes a unified and reproducible framework for forecasting Botswana’s rough-diamond export unit values, measured in United States dollars per carat, by integrating a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model with SHapley Additive exPlanations (SHAP). The proposed framework is designed to combine macroeconomic, diamond-market, country-level and global-supply indicators covering 2004–2025. Variables reported at different temporal frequencies will be harmonised to a monthly frequency while preserving temporal ordering, aggregation consistency and information availability at each forecast origin. The forecasting performance of the proposed CNN–LSTM model will be evaluated against naïve and appropriately specified econometric benchmarks, including Autoregressive Integrated Moving Average (ARIMA) and Generalised Autoregressive Conditional Heteroskedasticity (GARCH)-family models. Evaluation will employ walk-forward validation, Root Mean Squared Error, Mean Absolute Error, Mean Absolute Percentage Error and pairwise Diebold–Mariano tests, supported by robustness and sensitivity analyses. SHapley Additive exPlanations is pre-specified as a post-hoc interpretation stage within the analytical pipeline and will be used to quantify the predictive contributions of input variables without interpreting these associations as causal effects. As a proposed-framework paper, the study does not report completed empirical forecasting or SHAP results. Its contribution lies in establishing a methodologically coherent foundation for subsequent empirical evaluation of forecasting accuracy and interpretability in Botswana’s diamond-dependent economy.