Baselines and forecasts

Baselines provide smoothed estimates of time series data along with lower and upper bounds for the data over time. They can also be used as forecasts, using historical data to predict the behaviour of future data.

This functionality can also be used for anomaly detection by highlighting timestamps where the true value strayed outside of the lower and upper predicted bounds.

Baselines can be added to panels by using the SceneBaseliner component of scenes-ml, which will add a control to enable/disable the calculation, adjust the prediction intervals, discover seasonalities, and pin the results.

Usage

The code example below demonstrates how to add baselines to a time series panel.

import { SceneBaseliner } from '@grafana/scenes-ml';

// Default values are shown here, all are optional.
const baseliner = new SceneBaseliner({
  interval: 0.95,
  discoverSeasonalities: false,
  pinned: false,
});
const panel = PanelBuilders.timeseries().setHeaderActions([baseliner]).build();

note

Make sure you only add baselines to time series panels, as they rarely make sense for other panel types.

Pinning results

By default, baselines are recalculated on every state change, i.e. whenever the time range, query or interval changes. This isn't always desirable: for example, the user may want to zoom out and view the current forecasts in a future time range.

Enabling the pinned setting will freeze the current results, so they won't be recalculated as the time range or other settings are changed.

Technical details

scenes-ml uses the MSTL algorithm to produce in-sample and out-of-sample forecasts. This algorithm decomposes the data into trend, seasonality and residuals, then uses an an ETS algorithm to model the trend series.

By default, the algorithm assumes hourly, daily, weekly and yearly seasonality (if the data spans at least two of the given season length, i.e. at least two hours for hourly or at least two days for daily).

If the discover seasonalities setting is enabled, the baseliner will first attempt to detect any non-standard seasonality in the data using a periodogram and account for these seasonalities when modeling the data.