Free anomaly detector — paste numbers, get outliers.

How does the anomaly detector work?

The detector applies a rolling z-score: for each point, it computes the mean and standard deviation of the previous N points (the rolling window) and asks how many standard deviations the current point sits from the rolling mean. Points beyond a configurable threshold are flagged.

• Z-score formula: (value − rolling_mean) / rolling_stdev
• Default thresholds: |z| ≥ 2 = warning (statistically unusual, ~5% of points in a normal distribution); |z| ≥ 3 = critical (~0.3% — almost certainly worth investigating).
• Default window: 12 (one year of monthly data, or one quarter of weekly). Shorter windows are more reactive; longer windows are more stable.
• Burn-in: the first N points have no rolling stats yet, so they're labelled "n/a" rather than scored.

When to use this

• Variance review — finance / FP&A: anomalies in line items often map directly to variances vs budget.
• Audit prep — flagged transactions become candidates for the audit work programme.
• Reconciliation — differences between systems often surface as outliers in one of them.
• Fraud screening — a starting filter for human review, not a verdict.
• Data quality — outliers often reveal export bugs or data-entry errors.

Limits of the rolling z-score

• Strong seasonality can produce false positives — a Q4 spike isn't an anomaly if Q4 always spikes. For seasonal data, deseasonalise first or use a more sophisticated model (Holt-Winters residual scoring on the platform).
• Structural breaks (a pivot, a merger, a regulation change) shift the underlying mean, generating short-term false positives until the rolling window adapts.
• Skewed distributions (where data isn't roughly symmetric around the mean) inflate the right tail's z-scores. Consider log-transform first, or use an interquartile-range-based detector.
• Multivariate anomalies (revenue and cost moving in opposite directions when they should track) need a correlation-residual approach. Available on the full DataHub Pro platform.

Want this on every numeric column of your spreadsheet?

This calculator runs one column at a time. DataHub Pro automatically scans every numeric column on upload, ranks anomalies by severity and value impact, ties them back to source rows, and includes them in your branded Word/PowerPoint reports. from $14.99/mo — 7-day free trial.