Closed antoinecarme closed 1 year ago
Also add AUC
PyAF models can be optimized/selected based on these measures ("KendallTau" , "KS", "MWU" and "AUC"), but this is very experimental and , of course, can generate strange models.
Higher values of these performance indicators are usually better (distribution values are compared).
Some tests :
tests/Kendall_KS_MWU_AUC_Perfs/test_ozone_Optimized_By_AUC.py
tests/Kendall_KS_MWU_AUC_Perfs/test_ozone_Optimized_By_Kendall_Tau.py
tests/Kendall_KS_MWU_AUC_Perfs/test_ozone_Optimized_By_KS.py
tests/Kendall_KS_MWU_AUC_Perfs/test_ozone_Optimized_By_Mann_Whittney_U.py
FIXED
Add some outlier-resistant forecasting performance measures for robustness.
Some based on order statistics / ranks : Mann–Whitney U (equivalent to AUC), or Kendall rank correlation coefficient (Tau).
Some based on signal distribution : Kolmogorov–Smirnov (KS).
At least to have their value as a model debriefing indicator and to test their usage as a model selection criterion.
All these indicators have already an implementation in scipy. Easy.