Metrics

Connhex AI supports both essential and advanced evaluation metrics: while the latter might be less intuitive, they are often more representative of prediction quality. This is especially important in anomaly detection tasks, since anomalies are almost always windows of time instead of discrete points¹.

Metric	Shorthand	Forecasting	Anomaly detection
Mean Absolute Error	MAE	✅
Mean Absolute Ranged Relative Error	MARRE	✅
Root Mean Squared Error	RMSE	✅
symmetric Mean Absolute Percentage Error	sMAPE	✅
Root Mean Square Percent Error	RMSPE	✅
Mean Absolute Scaled Error	MASE	✅
Mean Scaled Interval Score	MSIS	✅
Mean Time To Detect	MTTD		✅
F1			✅
Precision			✅
Recall			✅
Pointwise F1			✅
Pointwise Precision			✅
Pointwise Recall			✅
Point-adjusted F1			✅
Point-adjusted Precision			✅
Point-adjusted Recall			✅
NAB Score			✅
NAB Score Low FN			✅
NAB Score Low FP			✅
F2			✅
F5			✅

Footnotes

1. In other words, you're usually better-off with point-adjusted metrics instead of pointwise metrics. You still have the option to select the latter, for example if models are performing poorly on short anomalies. ↩