anomaly-detector

Anomaly Detector

Scan your product data for things that don't look right — sudden spikes, unexpected drops, metrics that broke from their normal pattern.

When to Use

• "Did anything unusual happen this week?"
• "Alert me if conversion drops below its normal range"
• "Scan all our key metrics for anomalies"
• "Why did page views spike on Tuesday?"
• "Is the error rate abnormally high right now?"
• "Check if any funnel step changed unexpectedly"

Mental Model

An anomaly is a data point that deviates significantly from the expected pattern. Three types:

1. Spike: Sudden increase (errors, traffic, rage clicks)
2. Drop: Sudden decrease (conversion, engagement, revenue)
3. Break: The pattern itself changed — e.g., a metric that was cyclical suddenly went flat, or a metric that grew steadily now oscillates

Workflow

Step 1: Choose what to scan

If the user has specific metrics in mind, scan those. If they say "scan everything," pick a default set:

• Page views (traffic health)
• Errors (console errors, network failures)
• Conversion (key funnel completion)
• Frustrations (rage clicks, dead clicks from get_opportunities)

Ask: "I'll scan page views, errors, conversion, and frustrations over the last 14 days. That OK?"

Step 2: Build trend metrics

For each metric, build a trend over a wide time window (14-30 days):

fullstory:build_metric(query="page views", output_type="trend")
fullstory:compute_metric(metric_id, time_range="last_30_days")

Step 3: Detect anomalies

For each trend, look for:

Spikes/drops (>50% change from previous period's average):

• "Page views: 12,340 daily average, but Tuesday hit 31,200 (+153%) — possible anomaly unless there was a campaign or launch."

Breaks (pattern change):

• "Checkout conversion was stable at 20-22% for 3 weeks, then dropped to 14% on July 15 and hasn't recovered."

Zero events (metric that had data, now doesn't):

• "Rage clicks on /settings dropped to zero on July 20. Was a fix deployed? Or did tracking break?"