Such difficult analyses as disease association studies, which aim at mapping genetic variants underlying complex human diseases, rely on high-throughput genotyping techniques. However, a shortcoming of these techniques is the generation of missing calls. Computational inference of missing data represents a challenging alternative to genotyping again the missing regions. In this paper, we present SNPShuttle, an algorithm designed to improve inference accuracy with respect to a former software designed by Roberts and co-authors, NPUTE (Roberts et al., 2007). Given an SNP panel, NPUTE algorithm infers missing data through a single scan of the panel, exploiting local similarity within sliding windows. Instead, SNPShuttle scans an SNP panel in an iterative bi-directional way, to resolve missing data with more confidence. We evaluate the accuracy gain, systematically comparing a variant of NPUTE and SNPShuttle for controlled missing data percentages (pmiss) ranging from 5 to 30%. The corresponding benchmarks are built from the high resolution map of mouse strains made available by the Perlegen Project. In all cases (all missing data percentages, all 20 chromosomes studied), SNPShuttle is shown to bring a gain of accuracy. For pmiss percentages comprised between 5 and 15%, the average gain ranges between 1.56 and 1.81%. All missing data percentages confounded, accuracies obtained with the original method range in interval 92%-96.5%, for all 20 chromosomes analysed. In all cases, we show that SNPShuttle always increases the accuracy, by 0.88 up to 1.81%. We show that scanning an SNP panel in an iterative bi-directional way, only fixing SNPs inferred identically through two successive scans, brings invaluable accuracy gain with regard to the original method.