A new approach is proposed for the prediction of bit error rate (BER) in binary FSK receivers subjected to impulsive noise. This is accomplished by calculating the conditional error probability of the receiver and then integrating over the probability distribution of the incoming noise. The analysis shows that the BER prediction is crucially dependent on accurately modelling the envelope of the receiver filter impulse response. The closer the model is to the actual receiver impulse response, the better the agreement between prediction and measurement. The new approach is verified experimentally, first for a uniform train of constant-strength impulses and secondly for different noise strength distributions (such as those determined from the noise amplitude distribution) for both steady and fading signal cases.