Objective: To determine the ability of clinical measures of balance to distinguish fallers from non-fallers and to determine their predictive validity in identifying those at risk of falls. Data sources: AMED, CINAHL, Medline, Scopus, PubMed Central and Google Scholar. First search: July 2015. Final search: October 2017. Review methods: Inclusion criteria were studies of adults with a definite multiple sclerosis diagnosis, a clinical balance assessment and method of falls recording. Data were extracted independently by two reviewers. Study quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 scale and the modified Newcastle–Ottawa Quality Assessment Scale. Statistical analysis was conducted for the cross-sectional studies using Review Manager 5. The mean difference with 95% confidence interval in balance outcomes between fallers and non-fallers was used as the mode of analysis. Results: We included 33 studies (19 cross-sectional, 5 randomised controlled trials, 9 prospective) with a total of 3901 participants, of which 1917 (49%) were classified as fallers. The balance measures most commonly reported were the Berg Balance Scale, Timed Up and Go and Falls Efficacy Scale International. Meta-analysis demonstrated fallers perform significantly worse than non-fallers on all measures analysed except the Timed Up and Go Cognitive ( p < 0.05), but discriminative ability of the measures is commonly not reported. Of those reported, the Activities-specific Balance Confidence Scale had the highest area under the receiver operating characteristic curve value (0.92), but without reporting corresponding measures of clinical utility. Conclusion: Clinical measures of balance differ significantly between fallers and non-fallers but have poor predictive ability for falls risk in people with multiple sclerosis.