Skip to main content

Anomaly Detection

What Is Snowflake ML Anomaly Detection?​

SNOWFLAKE.ML.ANOMALY_DETECTION identifies data points that deviate significantly from expected patterns. It supports both supervised mode (with labeled anomalies) and unsupervised mode (learning normal behavior from unlabeled data). The function works on time series and tabular data.

Supervised vs Unsupervised Modes​

ModeLabel ColumnHow It Works
SupervisedRequired (BOOLEAN)Learns from historical examples of known anomalies
UnsupervisedNot providedLearns normal patterns and flags deviations

Use supervised mode when you have reliable historical labels. Use unsupervised mode when anomalies are rare or unlabeled.

Training Data Format​

Time Series Anomaly Detection​

ColumnTypeRequired
TimestampTIMESTAMP_NTZ or DATEYes
Target valueNUMBER or FLOATYes
Series identifierVARCHAROptional (for multi-series)
LabelBOOLEANOptional (supervised only)

Tabular Anomaly Detection​

For non-time-series data, omit the timestamp column. The model detects outliers based on feature distributions.

Training an Unsupervised Model​

-- Prepare clean training data (assumed to be mostly normal)
CREATE OR REPLACE VIEW normal_transactions AS
SELECT
transaction_date,
account_id,
daily_transaction_count,
daily_total_amount,
unique_merchants
FROM daily_account_metrics
WHERE transaction_date BETWEEN '2023-01-01' AND '2024-06-30';

-- Train unsupervised anomaly detection
CREATE OR REPLACE SNOWFLAKE.ML.ANOMALY_DETECTION txn_anomaly_model(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'NORMAL_TRANSACTIONS'),
TIMESTAMP_COLNAME => 'TRANSACTION_DATE',
TARGET_COLNAME => 'DAILY_TOTAL_AMOUNT',
SERIES_COLNAME => 'ACCOUNT_ID'
);

Training a Supervised Model​

-- Training data includes a label column
CREATE OR REPLACE VIEW labeled_metrics AS
SELECT
reading_timestamp,
sensor_id,
temperature,
pressure,
vibration,
is_anomaly -- TRUE for known anomalies, FALSE for normal
FROM sensor_readings_labeled;

-- Train supervised model
CREATE OR REPLACE SNOWFLAKE.ML.ANOMALY_DETECTION sensor_model(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'LABELED_METRICS'),
TIMESTAMP_COLNAME => 'READING_TIMESTAMP',
TARGET_COLNAME => 'TEMPERATURE',
SERIES_COLNAME => 'SENSOR_ID',
LABEL_COLNAME => 'IS_ANOMALY'
);

Detecting Anomalies​

-- Run detection on new data
CALL txn_anomaly_model!DETECT_ANOMALIES(
INPUT_DATA => SYSTEM$REFERENCE('TABLE', 'RECENT_ACCOUNT_METRICS'),
TIMESTAMP_COLNAME => 'TRANSACTION_DATE',
TARGET_COLNAME => 'DAILY_TOTAL_AMOUNT',
SERIES_COLNAME => 'ACCOUNT_ID'
);

The output includes:

ColumnDescription
SERIESSeries identifier
TSTimestamp of the data point
YActual observed value
FORECASTExpected value based on learned patterns
IS_ANOMALYBOOLEAN indicating whether the point is anomalous
PERCENTILEHow extreme the value is (0.0 to 1.0)
DISTANCEDeviation from expected in standard deviations
-- Store and filter results
CREATE OR REPLACE TABLE detected_anomalies AS
SELECT * FROM TABLE(RESULT_SCAN(LAST_QUERY_ID()))
WHERE is_anomaly = TRUE;

Practical Example: Transaction Monitoring​

-- Train on 6 months of normal transaction patterns
CREATE OR REPLACE SNOWFLAKE.ML.ANOMALY_DETECTION fraud_detector(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'CLEAN_TRANSACTION_HISTORY'),
TIMESTAMP_COLNAME => 'TXN_DATE',
TARGET_COLNAME => 'TXN_AMOUNT',
SERIES_COLNAME => 'CUSTOMER_ID'
);

-- Detect anomalies in yesterday's transactions
CALL fraud_detector!DETECT_ANOMALIES(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'YESTERDAY_TRANSACTIONS'),
TIMESTAMP_COLNAME => 'TXN_DATE',
TARGET_COLNAME => 'TXN_AMOUNT',
SERIES_COLNAME => 'CUSTOMER_ID'
);

-- Alert on high-confidence anomalies
SELECT
series AS customer_id,
ts AS transaction_date,
y AS actual_amount,
forecast AS expected_amount,
distance AS std_deviations_away
FROM TABLE(RESULT_SCAN(LAST_QUERY_ID()))
WHERE is_anomaly = TRUE
AND distance > 3.0
ORDER BY distance DESC;

Practical Example: Sensor Data Monitoring​

-- Detect temperature anomalies across factory sensors
CREATE OR REPLACE SNOWFLAKE.ML.ANOMALY_DETECTION temp_monitor(
INPUT_DATA => SYSTEM$REFERENCE('TABLE', 'SENSOR_BASELINE'),
TIMESTAMP_COLNAME => 'READING_TS',
TARGET_COLNAME => 'TEMPERATURE_C',
SERIES_COLNAME => 'SENSOR_ID'
);

CALL temp_monitor!DETECT_ANOMALIES(
INPUT_DATA => SYSTEM$REFERENCE('TABLE', 'LATEST_SENSOR_READINGS'),
TIMESTAMP_COLNAME => 'READING_TS',
TARGET_COLNAME => 'TEMPERATURE_C',
SERIES_COLNAME => 'SENSOR_ID'
);

-- Join anomalies with sensor metadata for investigation
SELECT
a.series AS sensor_id,
s.location,
s.equipment_type,
a.ts AS reading_time,
a.y AS actual_temp,
a.forecast AS expected_temp,
a.distance
FROM TABLE(RESULT_SCAN(LAST_QUERY_ID())) a
JOIN sensor_metadata s ON a.series = s.sensor_id
WHERE a.is_anomaly = TRUE
ORDER BY a.distance DESC;

Best Practices​

  1. Train on clean data — In unsupervised mode, the training set should represent normal behavior. Remove known outages or incidents before training.
  2. Use the DISTANCE column — Not all anomalies are equal. Higher distance values indicate more extreme deviations. Set thresholds based on business tolerance.
  3. Combine with alerting — Schedule anomaly detection as a Snowflake Task and use notification integrations to alert on-call teams.
  4. Multi-series at scale — Group related entities (accounts, sensors, stores) into a single model for efficiency.
  5. Retrain after regime changes — If the business undergoes a structural change (new product launch, acquisition), retrain the model to establish a new baseline.