Predicted By Check FAQ

Answers to common questions about how the prediction is compared, how the tolerance behaves, and how anomalies are reported, grouped by topic.

Behavior

What exactly does the check compare?

The value your expression predicts for the row against the field's actual value on that same row. The row passes when the difference between them is at most the tolerance.

Does the tolerance apply in both directions?

Yes. The comparison uses the absolute difference, so the band extends equally above and below the prediction, and a difference exactly equal to the tolerance passes.

Is the tolerance a percentage?

No, it is an absolute number in the field's own units and does not scale with the prediction. For a relative comparison, use Satisfies Expression.

How are empty values treated?

Unlike most field rules, they fail. An expression referencing an empty column returns nothing, and a row passes only when the difference is definitely within the band. Exclude those rows with a filter when a missing input is legitimate.

Can I use it on a date or timestamp field?

Yes, but convert both sides to a number inside the expression, for example to a Unix timestamp, and state the tolerance in the same unit. Use Validate to confirm before saving.

Anomaly Reporting

What do the anomaly messages look like?

Record Anomaly: The field '<field_name>' has value '<value>', which is not within the predicted range defined by <expression> +/- <tolerance>

Shape Anomaly: For the field '<field_name>', X.XXX% of N records (K) are not within the predicted range defined by <expression> +/- <tolerance>

When a filter is set, both messages end with [filter: <expression>].

Does Predicted By produce Record Anomalies, Shape Anomalies, or both?

At 100% coverage (the default), violating rows are reported as Record Anomalies. Below 100% coverage, a failed coverage assertion produces one Shape Anomaly for the dataset. A scan can also roll up a large number of Record Anomalies into one Shape Anomaly.

Does Custom Anomaly Description work for Predicted By?

Yes. Predicted By emits Record Anomalies, so the anomaly_message_field payload field (and the Custom Anomaly Description toggle in the UI) replaces the Record Anomaly message with the value of the named column on the violating row. When that column is null, missing, or empty, the standard Record template is used instead. The Shape Anomaly always uses the fixed template.

Configuration

Can I lower the coverage on a Predicted By check?

Yes. Coverage at 1.0 (100%) means every row must pass. Lowering coverage to 0.995 allows up to 0.5% of rows to fail without raising an anomaly; once the failing fraction goes past that, the check reports a Shape Anomaly. Use lower coverage with care: a real regression that happens to fall just under the threshold will not raise an alert at all.

Can I change the expression or the tolerance later?

Yes. In the UI, open the check, edit either property, and click Update; see Edit a Check for the full steps. Through the API, a PUT to /api/quality-checks/{id} updates the properties object. The rule type, the target container, and the associated Check Template stay immutable.

What permission do I need to create or edit a Predicted By check?

The Drafter team permission on the datastore covers Draft work (creating a check as Draft or editing it while it stays Draft). Anything that puts the check into evaluation, such as creating or editing an Active check, archiving, or deleting, requires the Author team permission (or above). Viewing only requires Reporter. See Permissions for the full matrix.

Should I use Predicted By or a simpler rule?

Use Predicted By when the expected value has to be computed per row. When the reference is simply another column, Equal To Field says so more clearly; when it is a constant, Equal To does.