1

Basically, I came across an article where the authors first ran a logistic regression on a data set to predict the probability (q=demand) of buying their product, as a function of price p and various other customer characteristics. Then they created a histogram, presented the so called "predicted demand buckets". That is, the observed nbr of clients within each range of predicted demand.

1.) How does these prediction buckets represent the segmentation power of the test?

Then they plotted the actual demand against predicted demand within each prediction bucket, and showed that the actual demand is within the confidence interval.

2.) How does this even make sense? To me, it feels like they use the data they regressed the model on, and then show that the model fits!

fritzenbauer
  • 31
  • 5
  • How big was the data set they used? And how many betas in the model? The answer to your question depends on this. :) – probabilityislogic Apr 15 '16 at 14:51
  • 10,000 customers, 6 betas ( 5 excluding intercept)! :) – fritzenbauer Apr 16 '16 at 10:05
  • My response to this question - http://stats.stackexchange.com/questions/20010/how-can-i-help-ensure-testing-data-does-not-leak-into-training-data/20044#20044 should hopefully give an answer for question 2. This is definitely a "big n, small p" model. – probabilityislogic Apr 17 '16 at 00:18

0 Answers0