Signal Decay inFinancial Prediction : Time-Decay Feature Engineering forFraud Detection and Customer Churn

Sammanfattning

Predictive systems in financial applications often rely on static summary features derived fromhistorical transaction data. Although such features are simple and interpretable, they can overlook thetemporal character of financial behavior, where recent events may carry a stronger predictive valuethan older events. This thesis investigates whether time-decay feature engineering can improve pre-dictive performance in longitudinal financial data, with a focus on fraud detection and customer churnprediction.The study was conducted as a quantitative computational experiment using two synthetic financialtransaction datasets. Static baseline features were compared with time-weighted features constructedusing exponential and linear decay functions. Several decay rates were evaluated through a parametersweep to examine how the decay parameter λ affected model performance. The experiments were eval-uated using AUC, F1-score, precision, recall, precision-recall analysis, and feature-category ablation.The results show that time-decay feature engineering can improve predictive performance, but the ef-fect depends on the dataset, prediction task, decay function, and selected decay rate. Exponential decaygenerally produced more stable results than linear decay, particularly for fraud detection. Moderate orlow deterioration rates often performed best, suggesting that recent events are important while olderhistorical information retains predictive value. The churn prediction experiments showed more mixedresults, partly because churn had to be defined as a proxy target based on future inactivity rather thanby an explicit business-defined churn label.Overall, the data show that time decay feature engineering is a useful and interpretable approach forembedding temporal dynamics into financial predictive modeling. The results, however, also show thatdecay parameters should not be treated as universal. Instead, they should be tuned according to thedataset, feature category and prediction objective. Since the study was based on synthetic datasets, thefindings should be interpreted as methodological evidence rather than direct operational conclusionsfor real financial institutions.