Search

Combining Multiple Model Probabilities

Here's another area of work for modeling Prosper loans. How do I combine the probabilities of multiple models? I started with the brain-dead approach of using a convex combination of probabilities:

$p^* = \sum_{i=1}^{n}w_ip_i$

This can be interpretated as the coefficients directly proportional to the amount of information each model has. The optimal weights can be estimated from training data. However I am wondering if I can improve this approach and, of course, the Google is all powerful. Time for some more bed-time reading.

Risk Analysis, Vol. 19, No. 2, 1999. Combining Probability Distributions From Experts in Risk Analysis. R. T. Clemen and R. L. Winkler.

Technical Report no. 543, Department of Statistics, University of Washington, Oct 2008. Combining Probability Forecasts. R. Ranjan and T. Gneiting.