dt-obs-predictive-analytics

Use the timeseries-forecast tool with a DQL query like timeseries avg(dt.host.cpu.usage) grouped by host at interval 1d, set forecastHorizon to 30, and provide at least 60 days of training history. The result includes point forecasts and confidence bands per host.

Novelty detection answers whether a metric's behavior changed and when, without needing a threshold. Anomaly detectors like adaptive or seasonal-baseline confirm whether a metric is currently violating an expected range and for how long.

No, the Dynatrace Forecast Analyzer supports univariate forecasting only, predicting one metric from its own history. Multivariate forecasting requires external tools such as Python, R, or Azure AutoML.

The analyzer requires at least 14 non-null values in the last third of the training window. Fix this by filtering to entities with continuous data or shortening the training window to exclude inactive periods.

Follow the 2:1 rule: training history must be at least twice the forecast horizon, so use 60 days of data for a 30-day forecast. Use interval 1d with forecastHorizon 30, since the horizon is counted in steps.

Use static-threshold-analyzer when the acceptable limit is known and fixed, such as CPU above 90 percent. Use adaptive-anomaly-detector when the normal range must be learned from the signal's own distribution.