PROJECT / ANOMALY-MONITOR
Detecting the unexpected.
When does a temperature change deserve attention? Explore a signal, review its alerts and introduce changes to compare two ways of detecting anomalies.
Reproducible projectWhat you can explore
Browse a historical temperature series and see which readings each detector flags as unusual. One uses Isolation Forest; the other measures how far the temperature departs from a statistical reference. In the second section, introduce a spike or a sustained shift, adjust its strength and choose where it starts. You can see how the alerts change without altering the original data. The app uses a public series from Numenta Anomaly Benchmark and lets you download what you are exploring.
What the project tells us
Detecting changes is only part of the problem: unnecessary alerts and missed anomalies matter too. In this test, both methods detected the two anomaly windows in the evaluation period. Isolation Forest produced fewer false alerts, but the statistical reference achieved a slightly better F1 score. Neither wins on every measure. This comparison uses one series and scores individual readings; it is not the official NAB score or a guarantee for other sensors.
EXECUTED EVALUATION / 2026.09.30
| Method | F1 | False positives | Alerts |
|---|---|---|---|
| Isolation Forest | 0.307 | 234 | 390 |
| Statistical baseline | 0.318 | 295 | 469 |
The baseline achieves better F1; Isolation Forest reduces false positives. A more complex model does not automatically improve every metric.