How We Find the Best Algorithms for Dynamic Baseline Alerts
Blog post from New Relic
New Relic's development of Dynamic Baseline Alerts employs predictive analytics to set dynamic alert thresholds, offering an alternative to static thresholds by identifying significant deviations from normal behavior rather than precise predictions. The process involved defining "better" in terms of customer benefit, using both human and machine scoring, and settling on a Mean Absolute Scaled Error with specific adjustments to suit various data scales and anomaly detection. Additionally, the evaluation incorporated precision, recall, and F1 scores from machine learning to assess classification accuracy, alongside human evaluations to ensure practical applicability. These alerts, designed for metrics with cyclical patterns, aim to help users determine what constitutes normal behavior in their systems, with plans for broader availability and integration into New Relic's Digital Intelligence Platform.