Predictive Alerts

Set up proactive alerts — warn the user before a metric crosses a critical threshold, not after.

When to Use

• "Alert me if conversion is about to drop below 15%"
• "Warn me if errors are trending toward our SLO limit"
• "Tell me when rage clicks are on track to double this month"
• "Set up a watch on the new checkout flow — alert if anything changes"
• "What should I be worried about tomorrow?"

Mental Model

Reactive monitoring tells you what already happened. Predictive alerting tells you what's about to happen based on the trend.

If error rate is 0.8% today, 1.2% yesterday, and 1.6% the day before — you're on track to hit your 3% SLO limit in 3 days. Alert now, not when you breach it.

Workflow

Step 1: Define the alert

What metric, what threshold, what direction?

Alert: Checkout conversion drops below 15%
Current: 21%
Trend: Declining -0.5pp per week
Predicted threshold breach: ~12 weeks (if trend holds)
→ No immediate alert needed, but set a watch for <18% as early warning

Step 2: Analyze the trend

Build a trend metric and assess the trajectory:

fullstory:build_metric(query="checkout conversion rate", output_type="trend")
fullstory:compute_metric(metric_id, time_range="last_90_days")

Analyze the trend:

• Direction: Up, down, or flat?
• Velocity: How fast is it changing? (per day, per week)
• Acceleration: Is the change speeding up or slowing down?
• Volatility: How much does it bounce around? (noisy trends are harder to predict)

Step 3: Predict threshold breach

Simple linear projection:

Current value: 21%
Weekly change: -0.5pp
Threshold: 15%
Weeks to breach: (21 - 15) / 0.5 = 12 weeks

Caveat heavily: "This is a simple linear projection based on the last 4 weeks. Trends change. This is an early warning, not a forecast."

Step 4: Check for acceleration

Is the decline speeding up?

Last week: -0.3pp
This week: -0.5pp
Week before: -0.2pp
→ The decline is accelerating (from -0.2 to -0.5). If it continues accelerating, the breach could happen in 6-8 weeks, not 12.

Step 5: Assess confidence

Step 6: Set the alert

Based on the analysis, define:

• Watch threshold: Early warning (e.g., <18% — triggers investigation)
• Alert threshold: Action required (e.g., <15% — triggers incident response)
• Check frequency: How often to re-evaluate? (daily, weekly)

"If checkout conversion hits 18%, run funnel-doctor on the checkout flow. If it hits 15%, alert the on-call engineer."

Step 7: Report

## Predictive Alerts — Current Watches

### 🔴 ACTIVE: Error Rate Approaching SLO
- Current: 2.1%
- SLO limit: 3.0%
- Trend: +0.15pp per week (accelerating)
- Predicted breach: ~6 weeks (was 12 weeks last check — it's getting worse faster)
- Confidence: High (steady trend, low volatility)
- Action: run `error-forensics` now — don't wait for the breach

### 🟡 WATCH: Checkout Conversion Declining
- Current: 21%
- Alert threshold: 15%
- Trend: -0.5pp per week (steady)
- Predicted reach: 12 weeks at current rate
- Confidence: Medium (moderate volatility)
- Action: monitor weekly. At 18%, run `funnel-doctor`.

### 🟢 HEALTHY: Rage Click Rate Under Control
- Current: 3.2% of clicks
- Alert threshold: 5%
- Trend: Declining (-0.1pp per week)
- No breach predicted

### New This Week
- Mobile conversion gap widening (desktop 22%, mobile 11% — from 12% last month)
- Not at alert threshold yet, but worth investigating

Alert Configuration Template

Guidelines

• This is trend projection, not machine learning. Be explicit: "I'm projecting from the recent trend. This is directionally useful, not mathematically precise."
• Acceleration matters more than velocity. A metric declining at a steady rate is predictable. A metric declining faster each week needs immediate attention.
• A single data point breaking the threshold is not an alert — it might be noise. Require 2-3 consecutive points outside the watch threshold before triggering.
• Alerts are only useful if they have a defined action. Don't create alerts for "FYI" metrics — every alert should trigger a specific investigation or response.
• Re-baseline after major changes. If you deploy a fix for the declining conversion, the old trend is invalid. Reset the projection.
• For critical business metrics (revenue, checkout), set both watch AND alert thresholds. For secondary metrics, a single watch threshold is enough.