A Neuro-Symbolic Edge Stack for Fragile Economies – Binary Cellular Neural Networks and Auto-Mined ASP Rules for Robust Econometrics in the Democratic Republic of Congo

Abstract

Forecasting macro- and microeconomic trends is especially difficult in data-scarce settings such as the Democratic Republic of the Congo (DRC). Standard deep learning models, LSTMs, temporal CNNs, and Transformers typically require clean, synchronized time-series data and extensive GPU training. Yet, they offer limited transparency to central bank analysts and policymakers. We propose dCNN-E(ASP), a fully analytical neuro-symbolic framework that combines a binary Cellular Neural Network reservoir with a self-growing Answer Set Programming rulebook. Training requires only two closed-form matrix inversions, enabling realtime inference on a $30 Raspberry Pi and robustness to missing or noisy data. Across three Congolese applications, headline inflation, hydropower grid balancing, and cross-border copper flows retrospective backtests achieve 12–20% lower MAPE than tuned benchmarks. The method also provides clause-level explanations (e.g., a fuel-price shock within 14 days triggers a maize-price surge) and includes tutorial prose and pseudocode for easy local replication without GPUs.