Classification & Contribution to total

I'm writing a program and I am trying to come up with a function that takes as inputs a number of criteria (they are by and large numeric with a few "categoricals", but the categoricals can be made into dummy variables [or I've made a mistake]).

My question is this: My response variable is a vector of 15 unique "items", in this case representing prices. I have for each of those 15 "items" (which I treat as observations) 15 sets of corresponding criteria/grouping/regressor variables. However, there are only 5 criteria for each of the 15 grouping variables/regressors. I am trying to:

1) Determine the relative importance of each grouping var/regressor 2) Use that information (or something else)to try to disentangle the grouping vars/regressors' contribution to the price (response variable).

Problem: My regressors/grouping vars cannot be represented as a matrix in their current form (from what I can figure).

I tried using [ranked,weights] = relieff(X,Prices,K); but 1) My results are very sensitive to the choice of K; and 2) I have a lot more price information (the forward curve, and not just the average of prices [although it's not critical in this particular instance]), as well as historical price data so I feel I'm potentially throwing away some information. However, the categoricals/regressors are constants so I'm thinking there will be some problems with rank deficiency if I just do a repmat of those.

Any ideas how to approach this? I know this is rather general, but I'm not necessarily looking for code, just ideas/thoughts on ways this might be handled/approached intelligently.

Your ideas would be much appreciated - thanks!

Answer 1

[Fang]

I have had a similar problem but could not find a satisfactory solution. Would love to get some input too.