How to use a trained recurrent neural network on new input data (without having output data!)

I have trained a RNN in matlab and I want to use it to predict the outputs for an input vector. but in order to use the net, always the output vector is also required!

even for one-step ahead net without loop! I have tried to feed the one-step ahead net by creating a for loop and build an input vector in each time-step by combining X(t) and the output of RNN from previous time-step as follow:

Xi=0;                      % Assumed output for the first timestep
for j=1:size(X,2)  
    input={X(:,j);Xi};     % It is like using function      
    ys = nets(input);      % get the output of RNN     
    y(end+1)=cell2mat(ys); % convert it to mat and save it  
    Xi=cell2mat(ys);       % save the last output for next input
end

But the performance is very bad so it seems that something is wrong.

what is the solution?