Hello!I know I'm not the first one to ask about parameter estimation (betas) in Svensson zero-coupon curve model. The thing is that I want to use rather an unusual error criteria and don't know if there's a convenient optimization technique for it.I'd like to define the error as 0 if the estimated (via curve) bond price falls between bid and ask prices and as an absolute difference between estimated price and bid (ask) price otherwise. (following Bliss, 1996)The error functional is then defined as a sum of individual errors described above. This functional is not quadratic and not even convex. So, I wonder if there is any optimization method which converges to a global minima in my case? I could implement genetic-annealing-voodoo magic- kind of optimization but I'd want to be sure that there is no other way than heuristics.Kind regards

it would surely be hard to find a gradient based minimzation method for this kind of error function.but you can go with search on param space, that i guess doesn't require differentiability of error function, but again non-convexity.. i would suggest tinkering with error function to keep it convex. you can try being innovative with some penalty function which takes care of your requirement(error 0 if price is within bid and ask and so on..) and also has good properties