Anomaly Detection

Find genuinely unusual data points in any metric — not just big percentage changes, but statistically significant deviations from the expected seasonal pattern.

Ask Orion to detect anomalies

Orion detects anomaly intent automatically. Try questions like:

• "Are there any anomalies in orders this month?"
• "What's unusual about revenue in the last 90 days?"
• "Did anything unexpected happen with conversions?"
• "Find anomalies in daily active users"

Orion responds with:

• A list of anomaly cards (date, actual vs expected, % difference, z-score)
• Severity label (high / medium)
• Plain-English explanation of why each point is unusual
• Or a green "no anomalies" confirmation if everything looks normal

How it works

Celiq uses Prophet residual z-score detection:

1. Fit a Prophet model to the full history (removes weekly + seasonal patterns)
2. Calculate residuals = actual − expected
3. Flag any point where |z-score| > sensitivity threshold

This means a normally-low Monday won't get flagged as an anomaly just because it's lower than Tuesday — the model already knows Mondays are lower.

Orion Pulse integration

The Orion Pulse (SpotIQ) scanner that generates proactive alerts uses the same ML detection when the ML service is available:

• ML available: Uses Prophet residuals, only surfaces anomalies in the last 7 days
• ML unavailable: Falls back to rule-based detection (≥15% week-on-week change)

This eliminates false positives from weekend patterns and normal seasonal variation.

Sensitivity settings

Orion uses 2.5 by default. Orion Pulse also uses 2.5.

Anomaly severity

Limitations

• Needs at least 14 data points; returns empty result with fewer
• Best with daily-grain data over 30–90 day windows
• Very sparse or highly irregular metrics may produce noisy results