Hello all, suppose we have ensemble output that consists of N members. Each member is real valued (single floating point number). In addition, assume we have "truth" values that specify what the "Correct" value is that the ensemble output is trying to predict. The ensemble and truth values are generated every T hours, and we have several months of data available.
We want to quantify how we'll each member performs by comparing it to the truth value. In addition, we like to know if certain members are biased on a consistent basis. I'm looking for information on a technique that takes ensemble members and performs linear regression on the values in order to quantify how these members do over time. In addition, is it possible to determine whether certain members are off by a certain amount and apply weights to them to adjust their output? The idea is to train the algorithm on a historical data set, then use it to predict actual values in real time. What is this approach called that employs regression to generate/update the member weights? If anyone has any info on this I'd be very grateful. Or, if anyone knows which keywords i could use to dig deeper, that would be great. Everything I've found so far are PDFs of research papers that say they're using a technique like I've described, but give no details. Perhaps it's trivial and I'm thinking too hard about this, but I can't find anything that describes how to perform these kinds of calculations. Thanks!
[link][4 comments]