Neural network question - Instability of the forecasts in a for loop

Dear all,

We have the following puzzle when using the neural network toolbox in a for loop. We try to implement a recursive experiment to assess the forecasts produced with narxnet. At each iteration of the loop, the sample is increased with one observation.

When we change the order of the recursions, the forecasts (and estimated weights) change (although the recursions are independent).

We guess that we should reinitialize the network. We tried to clear all the variables involved in the net between two loops but it does not solve the problem.

Thank you very much for your help.

Marie