Home Forums Analytics / Predictive Analytics Dec 2020 PA exam

Viewing 7 posts - 1 through 7 (of 7 total)
• Author
Posts
• #993
Narwal
Participant

Starting this thread since Reddit is not super helpful in figuring out game plan with December sitting considering states are going into lockdown

#994
Narwal
Participant

In WA, there is 25% capacity limit for business to hold training and testing if it cannot be done remotely. Tried calling SOA and Prometric today and no helpful responses. Anyone know if December exam will be postponed into January given that whole country is in worse situation with COVID-19 compared to March this year.

#1076
alex theodorakakos
Participant

How’s studying going for this? I basically started studying yesterday and the exam is on December 9th lol

#1077
Narwal
Participant

Meh! The exams are long. Little concerned about finishing it on time. Question on June 16, 2020 model solution: In the executive summary the solutions states “When this approach is compared to the recommended model on unseen data reserved for performance measurement, the predictions from the recommended model lead to 25% reduction in error”. Where is the 25% coming from?

#1108
GhostOfMFE
Participant

Guessing here, but take some metric of error using both models, and I imagine the percent change for said metric from one model to the other is around -25%.

#1116
Narwal
Participant

I get 31% on training set and 37% for the test set…still not close to 25%. I feel like I am still missing something here

GLM (from Task 7, excluding PC1)

Statistic on Training set: 1.5569789
Statistic on Test set: 1.445401

GLM, intercept model only
(Task 12 code in the rmd file solution)
Statistic on Training set:2.053229
Statistic on Test set:2.011089

#1204
Mel Phant
Participant

Hi, I hope someone in here can clear this up for me. I’m using a GLM, Gamma family with a Log link and want to interpret my findings. If I exponentiate the betas, is it true to say that figure (e^beta) for the intercept is the expected mean of the target variable using the base levels for each factor feature and zero for any numerical feature?

Viewing 7 posts - 1 through 7 (of 7 total)
• You must be logged in to reply to this topic.