Consider candidates that are tied according to distance with the current
worst predictor in order to consistently use the lowest indexed candidate
and to prefer neighbours in coarser LoD to neighbours in current LoD.