Scholarlink

Anomaly detection in banking IT systems is vital for maintaining operational efficiency but remains complex due to temporal patterns in time-series data. Traditional approaches, such as static thresholding, often fall short, underscoring the need for advanced forecasting methods. This study assesses five forecasting models—Amazon Chronos, Temporal Fusion Transformer (TFT), Prophet, AutoARIMA, and XGBoost—using proprietary data from Yapı Kredi Technology, covering 20 banking services across 578 days. A rolling-window forecasting approach was employed, with performance evaluated through Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). Amazon Chronos excelled, delivering the lowest MAE (39.09 ms) and MAPE (16.24%), with its superiority validated by Friedman and Nemenyi statistical tests. Its probabilistic forecasts facilitated dynamic anomaly detection using a 95% confidence interval. These results emphasize the potential of cutting-edge models to improve banking system reliability, paving the way for future enhancements in optimization and real-time deployment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Time Series Forecasting for Anomaly Detection in Banking IT Infrastructure: A Comparative Study of Modern and Traditional Methods

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ali Oguz Dogru,
  • Salih Imece

摘要

Anomaly detection in banking IT systems is vital for maintaining operational efficiency but remains complex due to temporal patterns in time-series data. Traditional approaches, such as static thresholding, often fall short, underscoring the need for advanced forecasting methods. This study assesses five forecasting models—Amazon Chronos, Temporal Fusion Transformer (TFT), Prophet, AutoARIMA, and XGBoost—using proprietary data from Yapı Kredi Technology, covering 20 banking services across 578 days. A rolling-window forecasting approach was employed, with performance evaluated through Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). Amazon Chronos excelled, delivering the lowest MAE (39.09 ms) and MAPE (16.24%), with its superiority validated by Friedman and Nemenyi statistical tests. Its probabilistic forecasts facilitated dynamic anomaly detection using a 95% confidence interval. These results emphasize the potential of cutting-edge models to improve banking system reliability, paving the way for future enhancements in optimization and real-time deployment.