Anomaly detection¶

FinSentry flags a month when a metric moves far from its own recent history. It looks at each metric separately, and separately for each cost center.

How a month gets flagged¶

For one metric in one cost center:

1. Take the six months right before the month being checked.
2. Work out the average and the normal amount of month-to-month swing over those six months.
3. Compare the current month against that. If it is more than two "normal swings" away from the average, the month is flagged.

A metric needs at least seven months of history before anything can be flagged (six months to learn what is normal, plus the month being checked). Shorter histories produce no anomalies.

Direction¶

Every flagged month is marked up (above its recent average) or down (below it).

Grouping months into one anomaly¶

Flagged months are combined into a single anomaly when they share all of:

• the same cost center,
• the same affected metric,
• the same direction,
• and fall in consecutive months.

The combined anomaly shows the full date range and how many months it covers. A gap in the months, a change of direction, or a different metric starts a new anomaly.

When it runs¶

Anomaly detection runs automatically once a dataset finishes importing, and again whenever metric formulas are recalculated. Each run replaces the previous results for that dataset.

Where to see results¶

The anomalies page lists every anomaly for the active dataset. Opening one shows the affected metric, the period, the largest drivers, and a short "what happened / why / what next" summary based on the surrounding months.