A hurdle model combined with Bernoulli part and truncated Poisson part can be used to predict zero-inflated geographic count response. To get the prediction with a hurdle model, the estimation of fixed effects can be easily solved as generalized linear model (GLM) does. An ad-hoc method, which re-fits the hurdle model to compute the predicted random effect for geographic IDs with missing response, is applied. However, no study has examined the performance of this prediction method for hurdle model, especially for the spatially correlated count responses with excessive zeros. This paper aims to check how well the hurdle predictors perform in ideal and real situations, by means of cross validation. The performance of the hurdle model based prediction is compared with the performance of the predictors from the universal kriging which is most widely used on spatial predictions. The simulation result shows that hurdle performs better than universal kriging based on mean absolute errors. The ideal situation is generated by using Monte-Carlo simulation. In order to examine the comparative performance with real data situations, two real data examples are presented. The results show that, in prediction using single observation per location (e.g. one year’s spatial observation) with excessive zeros, hurdle model does not perform well, while universal kriging also failed in the same situations especially for those non-zero points.