# Predicted By Best Practices

Source: https://userguide.qualytics.io/data-quality-checks/predicted-by/deep-dive/best-practices/

Guidelines for getting reliable signal from Predicted By checks while keeping the noise (and the maintenance) low.

## Write the prediction as the relationship you can defend

The expression is the definition of correct for this column. Base it on something that genuinely determines the value, a documented formula, a rate, another column it is derived from, rather than on a curve that happens to fit today's data.

## Size the tolerance from observed drift, not from a round number

Look at how far the column actually strays from the prediction over a few weeks, then set the band just outside that. A tolerance guessed at zero reports rounding, and one guessed too wide never reports anything.

## Remember the band does not scale

The tolerance is an absolute number. On a column whose values span orders of magnitude, one band cannot serve the whole range. Scope the check with a filter, or move to [Satisfies Expression](https://userguide.qualytics.io/data-quality-checks/satisfies-expression/introduction/) where you can write a relative comparison.

## Pair with Not Null when inputs are required

A row whose field or expression result is `NULL` passes. When those values are mandatory, add a [Not Null](https://userguide.qualytics.io/data-quality-checks/not-null/introduction/) check on the field and on each column the expression depends on.

## Prefer a simpler rule when the relationship is simple

If the prediction is just another column, [Equal To Field](https://userguide.qualytics.io/data-quality-checks/equal-to-field/introduction/) says so more clearly and offers a relative tolerance. If it is a constant, [Equal To](https://userguide.qualytics.io/data-quality-checks/equal-to/introduction/) does. Keep Predicted By for relationships that need an expression.

## Keep coverage at 100% unless a known backlog exists

At 100% coverage every failing row is reported as a Record Anomaly, which tells you exactly which rows break the rule. Below 100% the check reports a single Shape Anomaly only when the failing fraction crosses the tolerance, and no per-row detail is produced. Lower coverage only while a known set of legacy rows is being cleaned up.

## Scope with a filter instead of loosening coverage

When the rule only applies to part of the table (one segment, one channel, one period), express that with a filter clause rather than by lowering coverage. The filter removes the out-of-scope rows from evaluation entirely, and the expression is echoed in every anomaly message.

## Route the anomalies to the right people

A failure usually points at the system that produced the value, not at the warehouse. Set an **Anomaly Assignee** from the team that owns that producer, and tag the check so related checks are easy to find.

## See Also

- **Permissions**

  The team permission each action needs: view, create, edit, archive, restore, and delete.

  [Permissions](https://userguide.qualytics.io/data-quality-checks/predicted-by/deep-dive/permissions/)
- **How It Works**

  The complete reference: definition, field scope, the prediction expression, how the tolerance band works, date and time targets, NULL handling, filter behavior, and coverage.

  [How It Works](https://userguide.qualytics.io/data-quality-checks/predicted-by/deep-dive/how-it-works/)
- **Anomaly Reporting**

  The anomaly messages the check produces, what the numbers mean, Source Records behavior, and Custom Anomaly Description.

  [Anomaly Reporting](https://userguide.qualytics.io/data-quality-checks/predicted-by/deep-dive/anomaly-reporting/)
- **Examples**

  Three production scenarios with sample data, anomaly messages, and the SQL equivalent of what the check evaluates.

  [Examples](https://userguide.qualytics.io/data-quality-checks/predicted-by/deep-dive/examples/)

*Last updated: October 19, 2026.*
