The interplanetary magnetic field sector polarity at the Earth's location can be inferred from ground measurements of the cosmic ray north-south anisotropy. We present here the results of a systematic comparison between inferred and directly measured polarities in the period from January 1971 through December 1997. The overall success rate of the prediction method is 72%, in rather good agreement with previous findings. On the other hand, our analysis shows for the first time the limits of the method and calls for warning in its use. In fact, it turns out that the inferred polarities are affected by a bias, which becomes more relevant around the sunspot minimum. Even though the bias cannot be removed, it is possible to select conditions, occurring for 30% of the time, under which the inferred polarities are very reliable (the success rate being 95%).