332K Orders Later: How Ensemble ML Cut False Positives by 35%

A 25-day production experiment processed 332K orders to compare a single Isolation Forest model against a 3-model ensemble (Isolation Forest, LSTM, Autoencoder) for data quality monitoring. The ensemble reduced false positives by 35% and caught 30% more real anomalies, with only a slight increase in inference time. Results show ensemble ML is worth the complexity for multi-type anomaly detection.