Logistic Regression hardly ever predicts neutral conditions
I recreated your lr
model. I always get nice probability distributions for above normal
and below normal
ranging between 0.1 and 0.8 with a mean of .33, which is what I would expect. However, the distribution for normal
is very sharp, i.e. mean around .33 but only with a very small spread when looking at many forecast_times. Also it looks like .33 is the upper bound for normal
, and therefore normal
isnt predicted (never highest p). The training data (0,1,2 conditions) are quite evenly distributed. And therefore 1/3 of the cases, the neural ones, cannot be predicted.
Did you get the same in your model? Or is the expected for lr
?