Output Data Dictionary

Main output table

Table: {OUTPUTS_DATASET}.anomaly_detection_report

Partitioned by date; clustered by case_name and time_series_id.

Column	Type	Description
`date`	date	Date of the scored or training point
`case_name`	string	Case identifier (e.g., `ga4_events`)
`time_series_id`	string	Concatenated dimension value identifying the series
`metric_value`	float	Actual metric value for the row
`lower_bound`	float	Predicted lower confidence bound; `null` for training rows and series that did not meet detection thresholds
`upper_bound`	float	Predicted upper confidence bound; `null` under the same conditions
`anomalies_point`	float	Metric value when the point is anomalous; `null` otherwise
`is_anomaly`	boolean	Series-level anomaly flag: `true` when any point in the series was flagged during the detection window
`source`	string	`training` for historical context rows; `anomaly_detection` for scored rows
`is_strong_series`	boolean	`true` when the series meets the configured detection quality thresholds (`detection_min_series_days` and `detection_min_series_avg`)
`in_training_not_in_detection`	boolean	`true` when a series appeared in the training window but has no data in the detection window. The pipeline synthesizes a placeholder row for these series (metric and bounds are `null`, date is set to `CURRENT_DATE`) so they remain visible in the output rather than silently disappearing

Table: {TRANSFORMATIONS_DATASET}.int_anomaly_detection_{case}_anomalies

Partitioned by date; clustered by the case's time_series_id_cols.

Column	Type	Description
`date`	date	Date for scored point or synthetic missing-series marker
`{id columns}`	string or mixed	Case-specific time-series dimension columns
`_series_id`	string	Concatenated ID built from the configured dimensions
`{metric column}`	float	Case metric used for scoring
`lower_bound`	float	Model lower confidence bound
`upper_bound`	float	Model upper confidence bound
`anomalies_point`	float	Metric value when the point is flagged; `null` otherwise
`is_anomaly`	boolean	Point-level anomaly flag from `ML.DETECT_ANOMALIES`
`config_name`	string	Case name used at execution
`inserted_at`	timestamp	Processing timestamp
`is_strong_series`	boolean	Detection-window strength flag
`in_training_not_in_detection`	boolean	`true` for training-only missing series markers

Table: {TRANSFORMATIONS_DATASET}.int_anomaly_detection_{case}_time_series

Column	Type	Description
`{timestamp col}`	date or timestamp	Case timestamp field
`{id columns}`	string or mixed	Case dimensions (full series and single-dimension stacked variants)
`{metric column}`	numeric	Aggregated metric. If `metric_cap` is set, values are capped using `LEAST(_y, metric_cap)` before the column is written. If `metric_cap` is `null` or omitted, the raw aggregated value is used

source = 'training' rows provide historical context and do not contain anomaly bounds.
source = 'anomaly_detection' rows contain scored bounds and anomaly flags.
anomalies_point is null for non-anomaly rows; it only carries a value when the point is flagged.
Both lower_bound and upper_bound are clamped to 0. If the model predicts a negative lower bound, it will appear as 0 in the output.
When metric_cap is configured, training data is capped at that value. The metric_value column in the report reflects the original uncapped value, but the model's bounds were learned from capped data.
For alerting, pair is_anomaly with is_strong_series and a persistence check across consecutive days.