Thank you for your feedback.
Form temporarily unavailable. Please try again or contact to submit your comments.

Detect anomalies using predictive models

Log in to subscribe to topics and get notified when content changes.

Detect anomalies using predictive models

Submit MetricBase data to train a model, which you can use with a model trigger. Use models to detect data that varies from normal values.

Before you begin

Role required: admin

About this task

MetricBase supports the probabilistic exponentially weighted moving average (PEWMA) model. You need to run data through the model to train it. Training is the process of entering a lot of normal data into the model so it can determine the model parameters that best fit the normal data. In this way, the model can distinguish anomalous data. When you choose the data to train your model, make sure to eliminate all anomalous data.


  1. Navigate to MetricBase > MetricBase Models and click New.
  2. On the form, fill in the fields.
    Table 1. Model New Record form
    Field Description
    Model name Name of the model. The name can be any combination of alphanumeric characters. This model name is not the same as the PEWMA model type. In general, the name relates to the value in Group by.
    Table name Name of the table that contains the training data.
    Metric Name of the metric that you use to train the model. The metric must be in the table.
    Created Date that you trained the model.
    Filter Filters that you use to exclude some of the data in the dataset.
    Group by Field that is the basis of the predictive model. For example, if you select altitude, you generate a model based on the altitude metric, which establishes what is normal for altitude. To create multiple models, select a field, and click Submit and Train. To create a different model, select a different field in Group by, and click Submit and Train.
    Training Dataset Start Date First date in the dataset to start with. The data in the table starting with this date is used to train the model.
    Training Dataset End Date Last date in the dataset that you want to use. The data in the table from the start date up to this date is used to train the model.
    Valid until Date that serves as a reminder to consider retraining the model. If the model is performing well, there's no need to retrain it. The model can continue working past this date.
    Active Option to use the trained model.
  3. Click Submit and Train.
    MetricBase trains the model. When complete, the model appears on the MetricBase Model Instances tab.
  4. Click the model name.
    The modeling data appears as does the model string with the parameters optimized by the training.
    Trained model data
  5. (Optional) Change the model parameters and click Update.
  6. (Optional) Change the model parameters and click Update.
    The graph does not update. By clicking Update, you only save the revised model string.
    The PEWMA parameters are:
    • Alpha—A greater value means that more data is used to calculate the weighted average. Also, the weighted average line is smoother.
    • Beta—A greater number means that the model is more sensitive to variations from the mean. The following images show differing beta values. The second image has a lower beta and a higher tolerance for variations from the predicted value.
    Figure 1. High beta
    Higher beta
    Figure 2. Low beta
    Low beta
  7. (Optional) Click Delete to delete the model.
  8. Click the MetricBase model triggers tab to create a Flow Designer trigger for this model. For more information, see Create a model trigger.