<p>Non-technical losses (NTL) remain a major source of economic risk for electricity distribution utilities. Although extensive research has focused on classification-based theft detection, such approaches typically provide binary decisions and offer limited support for quantifying the magnitude of consumption irregularities, which is essential for inspection prioritization and financial planning. This study proposes a regression-driven framework to estimate the severity of deviation from expected demand behavior using the Cumulative Abnormal Deviation (CAD) index. Instead of attempting theft verification, the method provides a continuous risk indicator derived from the difference between measured consumption and a baseline demand model. The framework combines Extreme Gradient Boosting with Recursive Feature Elimination to construct a compact predictor while preserving feature interpretability. Experiments conducted on a publicly available smart-meter dataset demonstrate that the optimized XGBoost-RFE model achieves high predictive consistency (R² = 0.9869, RMSE = 2.4062, MAE = 1.1285) and outperforms ensemble and deep learning benchmarks under identical data conditions. Residual-based prediction intervals show stable uncertainty behavior, with 93.83% of observations captured within the nominal 95% confidence bounds. Feature-attribution analysis indicates that load-related variables dominate deviation formation, supporting the operational plausibility of the model outputs. The results suggest that reliable estimation of deviation severity can be achieved under the evaluated experimental conditions without direct access to inspection outcomes. Nevertheless, the framework should be interpreted as a decision-support and prioritization tool, not as proof of theft. The proposed methodology provides a statistically grounded step toward quantitative, explainable, and practically applicable NTL risk assessment in modern distribution systems.</p>