Preserving Performance in Anomaly Detection Models for Real-Time Univariate Streams