Investigating the Impact of Transaction Temporal Patterns on Fraud Detection Models

Abstract

In the era of digital finance, the growth of transactions has increased the challenge of detecting fraudulent activities. Traditional systems are not enough and need to enhance the ability of capturing the patterns of financial fraud. This study investigates how important the temporal features are in fraud detection. Three algorithms, including Logistic Regression, Random Forest, and XGBoost, are used in this research. They are evaluated both with and without temporal features such as transaction hour and recent transaction counts. Experimental results show that integrating temporal features slightly improves overall accuracy of all used models. But they contribute little to the detection of fraudulent transactions. Static features like account and customer attributes, keep acting as the main predictors of fraud. Among all tested models, XGBoost has the highest and most stable performance. At the same time, Random Forest and Logistic Regression show similar patterns. Feature importance analysis further confirms that temporal features provide moderate but limited benefits. The statement that working out class imbalance problem is more critical when dealing with improving fraud detection could be drawn.