# Predicted By Troubleshooting

Source: https://userguide.qualytics.io/data-quality-checks/predicted-by/troubleshooting/

Common problems when creating and running a Predicted By check, their causes, and how to resolve them.

## Validation Fails When Creating the Check

Clicking **Validate** returns an error instead of a success message.

**Cause:** The rule could not run against the selected data. The most common reasons are an invalid **Expression**, a **Tolerance** that is missing or not a number, an invalid **Filter Clause**, or a container that failed to load.

**Resolution:**

1. Review the error message: it explains what needs attention.
2. Check the **Filter Clause** for typos: it must be a valid SQL `WHERE` expression for the selected container.
3. If the message says the container is marked as **Unloadable**, the container was skipped after repeated operation failures. Follow the steps in [Unloadable Container Error](https://userguide.qualytics.io/operations/scan/troubleshooting/#unloadable-container-error).

## Almost Every Row Is Reported

The check fires on nearly the whole container.

**Cause:** Usually the tolerance is far tighter than the column's real drift, or the expression predicts on a different scale than the field (a fraction against an amount, for example).

**Resolution:** Compare a few predicted values against the actual ones to confirm the two are on the same scale, then set the tolerance just outside the drift you observe.

## Rows With NULL Values Are Not Flagged

Records where the field or the expression's result is `NULL` never appear as anomalies.

**Cause:** When the expression returns `NULL` or the field is `NULL`, the comparison has nothing to measure and the row passes.

**Resolution:** 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.

## The Tolerance Fits Small Values but Not Large Ones

The check is quiet on part of the data and noisy on the rest.

**Cause:** The tolerance is an absolute number and does not scale with the size of the prediction.

**Resolution:** Scope the check to a comparable slice with a filter, or express a relative comparison with [Satisfies Expression](https://userguide.qualytics.io/data-quality-checks/satisfies-expression/introduction/).

## Validation Fails on a Date or Timestamp Field

Clicking **Validate** on a check whose field is a date or timestamp returns a data type error.

**Cause:** The **Expression** can transform only the predicted value. The selected field stays a date or timestamp, so the difference between the two is not a number the tolerance can measure.

**Resolution:** Write the complete comparison in a [Satisfies Expression](https://userguide.qualytics.io/data-quality-checks/satisfies-expression/introduction/) check instead, converting both sides to the same numeric unit (for example, `abs(unix_timestamp(l_receiptdate) - unix_timestamp(date_add(l_shipdate, 15))) <= 86400`).

## Validation Fails on the Expression

Clicking **Validate** returns an error instead of a success message.

**Cause:** The expression is not valid SQL for this container: a misspelled column, an unbalanced parenthesis, or a function the source does not provide.

**Resolution:** Read the error message, then check the column names against the container and simplify the expression until it validates.

## An Edited Check Keeps Behaving the Old Way

You changed the configuration but the anomaly list did not change.

**Cause:** Edits take effect on the next Scan. Saving the check does not re-evaluate the data, and anomalies raised under the previous configuration are not modified.

**Resolution:** Run a Scan on the container (or wait for the scheduled one). Old anomalies stay open until you triage them or a Full scan with [Auto Resolve](https://userguide.qualytics.io/operations/scan/deep-dive/read-strategies/#auto-resolve-on-full-scans) clears them. See [What Happens to Existing Anomalies](https://userguide.qualytics.io/data-quality-checks/predicted-by/how-tos/edit-a-check/#what-happens-to-existing-anomalies).

## No Anomalies Although the Data Looks Wrong

A Scan ran, the data clearly breaks the rule, but nothing was reported.

**Cause:** One of these configurations is excluding the violations:

- The check is in **Draft** status: Draft checks are not evaluated by Scans.
- The **Filter Clause** excludes those rows before the evaluation runs.
- **Coverage** is below 100% and the failing fraction stayed within the tolerance, so the check passed.

**Resolution:** Confirm the check is **Active** and test the filter expression against the offending rows. Then review the coverage setting; see [Coverage and Tolerance](https://userguide.qualytics.io/data-quality-checks/predicted-by/deep-dive/how-it-works/#coverage-and-tolerance).

## Expected One Anomaly per Row, Got a Single Rolled-Up One

Many rows failed, but the scan reported one Shape Anomaly instead of per-row Record Anomalies.

**Cause:** When the number of failing rows exceeds the scan's rollup threshold, the per-row findings are grouped into one rolled-up Shape Anomaly.

**Resolution:** This is expected behavior; the rolled-up anomaly carries sampled source records. To change the threshold, see [Maximum Record Anomalies per Check](https://userguide.qualytics.io/operations/scan/deep-dive/scan-settings/#maximum-record-anomalies-per-check).

*Last updated: October 19, 2026.*
