>On 19 Sep 2003 21:35:22 -0700, [EMAIL PROTECTED] (Anwar) wrote: >"Does anyone know of regression methods that satisfy the following three criteria: > 1) nonlinear,
Nonlinear regression? for beginners, one might try: http://www.curvefit.com/ since they seem to have a 'how-to' guide and offer a free demo. > 2) variable coefficients (i.e. updating of coefficient values), AND Given that you said "updating" I assume you are thinking of an algorithmic approach to obtaining a set of fixed coefficients by varying the coefficients initially? Given the formulation of your question, I doubt you are referring to threshold models. So we shall leave out that option entirely. Generally speaking -- Fitting coefficients is essentially a trial and error process within some large trial-and-error space. The question of HOW you want to do this, depends on the size of the parameter space you are searching. Big spaces (lots of possible solutions, possibly one global solution) to me often invokes ideas of simulated annealing. Smaller spaces (one minimum or maximum) might be traversed using simpler gradient based algorithms without restarts. Of course, you need to be more specific about what you are doing. > 3) minimize out-of-sample, one-step ahead, forecasting error >rather than in-sample fit? " Here's how I would interpret this -- Assuming I have a function (above nonlinear regression equation) that can do forecasts: 1. Split data in half. Use half for fitting coefficients to your function. 2. Use unused half for estimating "out of sample" forecast error. This sounds like a homework assignment. You need to think more carefully about what the question is asking, and be able to specify (precisely) exactly what your solution requires. Think about your inputs, your outputs, and your black-boxes. If you can't do this, you haven't gotten "half-way" to solving the problem yet. If you can, the solution should be clear. cheeriOs, pradyumna . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
