مجید ابتیاع

Credit card fraud remains a significant threat to financial institutions and the integrity of digital payment systems, posing challenges for both operational risk management and regulatory oversight. This paper presents a novel hybrid machine learning framework for credit card fraud detection that combines Principal Component Analysis (PCA) for feature extraction with a Support Vector Machine (SVM) based feature selection mechanism. The aim is to reduce dimensionality while retaining the most informative features, thereby improving detection performance on highly imbalanced transaction datasets. The approach is evaluated on a large credit card transactions dataset, where PCA is first used to transform the input variables into principal components capturing the majority of variance, and an SVM with recursive feature elimination is then employed to identify and retain the most relevant components. Experimental results demonstrate that the proposed PCA-SVM pipeline significantly outperforms baseline models lacking this hybrid feature engineering: for example, it achieves a higher fraud recall (detection rate) by several percentage points while maintaining high precision, leading to improved F1-scores and overall accuracy. These findings indicate that the hybrid method effectively mitigates class imbalance issues and eliminates redundant features, yielding a more compact and robust fraud detection model. By enhancing the identification of rare fraudulent transactions without excessive false alarms, our study contributes to central bank objectives in fraud risk management. The proposed framework can strengthen the resilience of digital payment infrastructures and support payment system oversight, ultimately helping to safeguard financial stability and public trust in electronic payment channels.

The recent surge of artificial-intelligence (AI) activity has sparked concern that large-scale model training, cloud inference, and data-centre expansion could accelerate national energy demand. We marshal a 21-year annual panel for the United States (2004–2024) that couples multiple AI proxies—technology-stock valuations and a ChatGPT-era dummy—with four aggregate energy series (fossil fuels, nuclear, renewables, total primary energy). A five-stage empirical protocol implemented in Python combines Engle–Granger cointegration testing, higher-order ADF stationarity checks, linear and nonlinear dependence diagnostics (Pearson, Dynamic Time Warping, mutual information), multicollinearity screening (variance-inflation factors), and out-of-sample forecasting with linear regression, decision trees, random forests, and support-vector machines augmented by SHAP explainability. Across all tests we find no evidence that AI developments imprint on national energy use: AI variables cointegrate only with one another, their short-run correlations with energy vanish once trends are removed, their mutual-information scores remain near zero, and their inclusion never improves predictive accuracy beyond a parsimonious macro model driven by GDP, inflation, and population. SHAP rankings confirm that AI features carry negligible explanatory weight relative to conventional fundamentals. We conclude that, to date, AI’s macro-level energy footprint is statistically invisible—any electricity it consumes is either too small or offset by efficiency gains within the wider economy. Policymakers should therefore continue to anchor long-range energy scenarios to established economic drivers while monitoring localised data-centre hotspots that national aggregates obscure.