Back to library

📉Detect Anomalies in Time-Series Data

Stop alerting on every weekend dip and missing the real incidents — learn to separate point, contextual, and collective anomalies, match each to the right detector (Z-score, isolation forest, LSTM autoencoder), and design an SLI alerting rule that survives weekly seasonality plus a slow trend.

Applied14 drops~2-week path · 5–8 min/daytechnology

Phase 1The Three Anomaly Types You're Conflating

Separate the three anomaly types your alerts confuse daily

4 drops
  1. Your weekend dip is not an incident

    6 min

    Your weekend dip is not an incident

  2. One value that doesn't belong

    6 min

    One value that doesn't belong

  3. The value is fine; the timing is wrong

    7 min

    The value is fine; the timing is wrong

  4. Each value is normal; the sequence is the problem

    7 min

    Each value is normal; the sequence is the problem

Phase 2Detectors in Practice: Z-Score to LSTM

Run isolation forest and LSTM autoencoder on real telemetry

5 drops
  1. The cheapest detector works perfectly — until your data has a trend

    7 min

    The cheapest detector works perfectly — until your data has a trend

  2. Build trees that isolate outliers in three splits

    7 min

    Build trees that isolate outliers in three splits

  3. Teach a network what normal looks like, then alert on surprise

    8 min

    Teach a network what normal looks like, then alert on surprise

  4. Match the detector to the failure mode

    6 min

    Match the detector to the failure mode

  5. Run all three on the same telemetry and watch them disagree

    7 min

    Run all three on the same telemetry and watch them disagree

Phase 3Why Static Thresholds Fail on Seasonal Data

See why static thresholds fail on seasonal SLIs

4 drops
  1. Break the signal into trend + season + residual

    7 min

    Break the signal into trend + season + residual

  2. The threshold that fired 400 times this month

    7 min

    The threshold that fired 400 times this month

  3. Alert on what's left over, not what's expected

    7 min

    Alert on what's left over, not what's expected

  4. Slow drift hidden inside weekly waves

    7 min

    Slow drift hidden inside weekly waves

Phase 4Design an Alerting Rule That Survives Seasonality

Design an alerting rule that survives weekly seasonality

1 drop
  1. Design an alerting rule for an SLI with weekly seasonality and a slow trend

    8 min

    Design an alerting rule for an SLI with weekly seasonality and a slow trend

Frequently asked questions

What is anomaly detection in time-series data?
This is covered in the “Detect Anomalies in Time-Series Data” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
What's the difference between point, contextual, and collective anomalies?
This is covered in the “Detect Anomalies in Time-Series Data” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
When should I use isolation forest vs LSTM autoencoder for anomaly detection?
This is covered in the “Detect Anomalies in Time-Series Data” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
Why do static thresholds fail on seasonal metrics?
This is covered in the “Detect Anomalies in Time-Series Data” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
How do I write an alerting rule that doesn't fire on every weekend dip?
This is covered in the “Detect Anomalies in Time-Series Data” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.