It performs FIR modeling for an appropriately selected number of lags: y(t) = \sum_{i=0}{N-1} b_i*u(t-i). For example, you may do arx(data, [na nb nk]) where na = 0, nb = 20, nk = 0 for a SISO system.
The "data" here in an iddata object containing the measured output signal and a "pre-whitened" input signal (u_white). The prewhitening of the input is performed by filtering the actual input signal using the model: u_white(t) = 1/a(q) u(t). For example, do the following to obtain the prewhitened input:
M = arx(u, 4); A = M.a; u_white = filter(u,A,[1, zeros(1,3)]);
The time horizon selected for estimation includes a portion of negative real axis so that feedback effects in the data may be detected. Please see the code in iddata/impulse to see how this is done. In many situations, this detail is irrelevant.
