SOLVED. Cannot replicate GLM predictions
SOLVED. It was the offset - needed to take the natural log of it before calibrating.
Hello,
I built a model using the GLM Classification plugin. The AUC is ~0.8 so it's fitting my data well but when I implement the GLM formula manually into Tableau the predictions are far too low despite having the correct shape. The model coefficients are also far from what I expected - this model has been in production for years, and this is just the first year that we've updated with new data using Dataiku.
I used Dataiku's What If? analysis to check a single record and confirmed that my prediction in Tableau matches the prediction of the What If? analysis. But when I score the data in Dataiku the prediction for that same record is significantly higher!
Why are the scored predictions different from the What If? predictions, and why is the model fitting my data so horribly when it has an AUC of 0.8? Algorithm and Train/Test inputs are attached. I am only doing a single fold, and I am using an offset but I made sure not to rescale it in the features handling.
