This longitudinal study aimed to determine changes in students' predictive reasoning across one year. Forty-four Australian students predicted future temperatures from a table of maximum monthly temperatures, explained their predictive strategies, and represented the data at two time points: Grade 3 and 4. Responses were analysed using a hierarchical framework of structural statistical features. Students were more likely in Grade 4 than Grade 3 to make reasonable predictions (87% vs 54%), to demonstrate data transnumeration in their representations (71% vs 19%), and to describe data prediction strategies based on extraction, clustering, aggregation of data, and observations of measures of central tendency.