Altered Merit Function in Levenberg Marquardt

jbeerhalter · August 21st, 2006, 1:27 pm

The standard Levenber Marquardt algorithm uses a basic SSE as its merit function and I would like to change that to a weighted sum of squared errors. Is there any mathmatical reason why I can't do that? Specifically, I'm trying to get a rough estimate of the current vol curve in the market, and I would like to weight the higher vega options more than the lower vega options.

zeta · August 21st, 2006, 7:40 pm

LM can only be used with SSE; better to use conjugate gradient descent. I'm by no means an expert on these matters; you may find statsoft more helpful

Athletico · August 21st, 2006, 8:06 pm

LM must be used with SSE, but AFAIK you're fine using a weighted sum of squares as long as your Jacobian and merit function are mutually consistent.