The use of genome-scale data to infer phylogenetic relationships has gained in popularity in recent years due to the progress made in target-gene capture and sequencing techniques. Data filtering, the approach of excluding data inconsistent with the model from analyses, presumably could alleviate problems caused by systematic errors in phylogenetic inference. Different data filtering criteria, such as those based on evolutionary rate and molecular clocklikeness as well as others have been proposed for selecting useful phylogenetic markers, yet few studies have tested these criteria using phylogenomic data. We developed a novel set of single-copy nuclear coding markers to capture thousands of target genes in gobioid fishes, a species-rich lineages of vertebrates, and tested the effects of data-filtering methods based on substitution rate and molecular clocklikeness while attempting to control for the compounding effects of missing data and variation in locus length. We found that molecular clocklikeness was a better predictor than overall substitution rate for phylogenetic usefulness of molecular markers in our study. In addition, when the 100 best ranked loci for our predictors were concatenated and analyzed using maximum likelihood, or combined in a coalescent-based species-tree analysis, the resulting trees showed a well-resolved topology of Gobioidei that mostly agrees with previous studies. However, trees generated from the 100 least clocklike frequently recovered conflicting, and in some cases clearly erroneous topologies with strong support, thus indicating strong systematic biases in those datasets. Collectively these results suggest that data filtering has the potential improve the performance of phylogenetic inference when using both a concatenation approach as well as methods that rely on input from individual gene trees (i.e. coalescent species-tree approaches), which may be preferred in scenarios where incomplete lineage sorting is likely to be an issue.