Anomaly Detection — Methodology

AskBiz uses a seasonal decomposition approach to anomaly detection. Instead of a simple threshold ('alert if revenue drops 20%'), we model the expected value of each metric on each specific day, accounting for:

• Day-of-week patterns (Tuesday is typically different from Saturday)
• Week-of-month patterns (first week of month often differs from last)
• Seasonal patterns (December differs from June)
• Recent trend (if revenue has been growing 10% week-on-week, the expected value this week is 10% higher than last week)

An anomaly is flagged when a metric deviates more than 2 standard deviations from its expected value under this model.

Two standard deviations is a well-established statistical threshold. Under a normal distribution, approximately 95.4% of values fall within 2 standard deviations of the mean. Values outside this range occur by chance roughly 4.6% of the time.

In practice, this means: for a business tracking 10 daily metrics, you'd expect roughly one anomaly alert every 2–3 days by pure chance — without any genuine business problem. To reduce false positive rates, we apply a secondary confirmation filter: an anomaly must persist for at least 2 consecutive measurement periods before triggering an alert (except for critical metrics like cash position, which alert immediately).

The anomaly detection model requires a minimum of 14 days of data to build a reliable baseline. For the first 14 days after connecting a data source, anomaly detection is in 'learning mode' — it observes but does not alert.

With 30+ days of data, seasonal patterns begin to emerge. With 90+ days, the model becomes significantly more accurate — it can distinguish 'low Sunday sales' (expected) from 'abnormally low Sunday sales' (potential anomaly).

This is why connecting a data source immediately and letting it sync for several months — even before you actively use AskBiz — pays dividends later.

You can adjust sensitivity in Intelligence → Alerts → Settings:

High sensitivity (1.5 standard deviations): More alerts. Catches smaller deviations. Higher false positive rate. Recommended for: high-stakes periods (peak season, product launch, new market entry) where you want maximum visibility.

Medium sensitivity (2 standard deviations, default): Balanced. Catches genuine anomalies with manageable false positive rate. Recommended for: normal business operation.

Low sensitivity (2.5 standard deviations): Fewer alerts. Only significant anomalies surface. Lower false positive rate but may miss early warning signs. Recommended for: stable, predictable businesses or periods when you want to reduce notification noise.