Configurable Variables

Module-level configuration

Path: includes/custom/modules/anomaly_detection/config.json (or .yaml / .js)

Field	Type	Required	Notes
`enabled`	boolean	Yes	Enables the module when `true`
`version`	number	Yes	Config version marker
`cases`	string[]	Yes	Case filenames (without extension) to execute

Default values are defined in includes/core/modules/anomaly_detection/default_config.json.

Path: includes/custom/modules/anomaly_detection/cases/{case}.yaml

Field	Type	Required	Description
`case_data_project`	string	Yes	Source project
`case_data_dataset`	string	Yes	Source dataset
`case_data_table`	string	Yes	Source table
`time_series_timestamp_col`	string	Yes	Timestamp/date column used by the model
`time_series_data_col`	string	Yes	Metric expression or column to aggregate
`time_series_data_agg`	string	No	Aggregation method: `sum`, `count`, or `count_distinct`. Defaults to `sum` when omitted
`time_series_id_cols`	string[]	No	Dimension columns that define a unique time series. Omit or set to `[]` to model a single global series
`top_n_time_series`	number	No	Limits training to the top N series by total training-window volume. Omit or set to `null` to include all series
`metric_cap`	number	No	Caps the metric value at this number before model training using `LEAST(_y, metric_cap)`. Prevents extreme outliers from skewing the ARIMA model. Omit or set to `null` to disable capping

Use one complete mode:

Mode	Required fields
Explicit dates	`training_start_date`, `training_end_date`
Rolling offsets	`training_end_days_ago`, `training_window_days`

Explicit dates take precedence when both modes are present. If neither mode is fully defined, the module throws a compile-time error.

Field	Type	Description
`model_type`	string	BQML model type; `ARIMA_PLUS` is the recommended default
`data_frequency`	string	Time grain of the series; use `DAILY` for GA4 data
`decompose_time_series`	boolean	Separates trend and seasonality components before fitting
`clean_spikes_and_dips`	boolean	Removes transient spikes from training to improve baseline quality
`adjust_step_changes`	boolean	Handles permanent level shifts in the series
`auto_arima`	boolean	Lets BQML select the best ARIMA order automatically
`model_version`	number or string	Appended to the model name to version the BQML artifact
`model_cron`	string	Cron expression controlling when the time-series preparation and model training actions run. On days that match the expression, the pipeline rebuilds the training table and retrains the model. On non-matching days, those two actions are disabled entirely and the existing model is left in place. Anomaly scoring always depends on the most recently trained model.

Field	Used in	Description
`training_min_series_days`	model training	Minimum number of rows per series required to include it in training
`training_min_series_avg`	model training	Minimum average metric value per series required to include it in training
`detection_min_series_days`	scoring flags	Minimum rows in the detection window for a series to be marked as strong
`detection_min_series_avg`	scoring flags	Minimum average metric in the detection window for a series to be marked as strong

Field	Description
`anomaly_detection_end_days_ago`	How many days ago detection ends (e.g., `1` targets yesterday)
`anomaly_detection_window_days`	Number of days to score (e.g., `1` scores a single day)
`anomaly_prob_threshold`	Probability threshold passed to `ML.DETECT_ANOMALIES`; higher values reduce false positives

time_series_data_agg must be one of: sum, count, count_distinct.
time_series_id_cols must be an array when provided; it may be empty.
top_n_time_series must be a positive integer when set.
model_cron must be a valid cron expression when set.
Training dates must resolve to exactly one valid mode (explicit or rolling). If neither mode is fully defined, the module raises a compile-time error.