Investigation of a data-driven hybrid machine learning model for nuclear proliferation risk prediction for early warning

Early detection of nuclear proliferation risk is inherently challenging due to the rarity and secrecy of the underlying escalation events. This study presents an investigation of a data-driven early-warning model based on historical data by transforming discrete proliferation stages into a continuous 0-to-1 risk score. The model uses political-economic and HS-code trade variables for 148 countries from 1939-2012. The model architecture (i) applies supervised learning regression (LightGBM regressors) in rolling windows to track annual stage scores based on Bleek and Narang taxonomies, (ii) detects residual anomalies using unsupervised learning via Isolation Forests, and (iii) fuses these signals in a meta-classifier to generate interpretable yearly alarm probabilities. The final model achieves an event-F1 score of 0.65 and ROC-AUC of 0.99, with a mean warning lead time of +1.14 years against the historical proliferation events. Quantifying the contributions of inputs to predictions using SHAP analysis reveals a post-1995 shift in proliferation drivers toward uranium centrifuge-related trade patterns. This underscores the need to examine the growing role of global supply chains. The system model based on an explainable, tree-based ensemble machine learning technique provides a transparent, lead-time-positive alternative to "black-box" deep learning and presents the possibility of developing a methodology for integrating multi-modal AI and dynamic data streams into an early warning tool to support nonproliferation intelligence.