It is shown that there is a stochastic version of the fact that the standard IIR, AR, ARX and ARMAX model and filter parametrizations suffer from a worst-case exponential growth of the required relative coefficient accuracy to preserve model and filter stability. This means that, inevitably, such standard model and filter parametrizations become infeasible to use for a subclass of order n filters and models when n is not small.