Statistical Methods Pass Nervous Networks in Time Series Anomali Detection

Time series plays a critical role in areas such as anomaly detection, finance, health and industry 4.0. A new arXiv study reveals a surprise competitor against deep nerve networks that dominate for a long time in this field: simple statistical methods.

The approach developed by researchers consists of these steps:

• Deleting statistical properties with floating window (sliding window) on Univariate time series
• Score each window using robust (robust) MAD model Tags
• Select the optimal feature subset for each domain on the tuning data set
• Gain 0.529 VUS-PR score in TSB-AD-U benchmark

The results are quite stunning. This simple pipeline determined a new benchmarking point, leaving the best nervous network methods over 0.45 scores and 0.44 scores of statistical methods. The study makes findings more reliable due to being realized under a strict selection protocol.

This result gives an important lesson to AI researchers. The assumption that complex architectures are always superior to simple methods is not always true. Especially in problems with good defined statistical properties such as time series analysis, transparent and calculation cheap methods can offer competitive performance.