recursiveARX: How to improve simulation error?

I'm studying online parameter estimation in Matlab with recursiveARX function. It's MISO model with 2 inputs and 1 output, there is 1 unmeasured disturbance of low magnitude so the 2 inputs should describe the output fairly well. As we can see in this picture this is not the case.

Blue line is y measured, red one is yHat which is the third argument from the adaptation step function, i.e. [A,B,yHat]=step() and the yellow line is y simulated calculated as follows: ySimulated=f(uMeasured(t,t-1,..),ySimulated(t-1,t-2,..)). The simulation function was tested and it works well. The yHat practically coincide/overlaps with yMeasured, but the ySimulated which is true indicator of model goodness is rather bad.

ForgettingFactor (FF=0.996) method turned out to be "best" solution in this case. I've noticed there is no additional options available to specify 'Focus' to 'Simulation' or any other parameters except the already tested adaptation algorithms.

What could I do to further improve the simulation error or model?