The early prediction of math difficulties (MD) is important as it facilitates timely support. MD are multifaceted, and several factors are involved in their manifestation. This makes the accurate early prediction of MD particularly challenging. In the present study, we aim to predict MD in Grade 6 with kindergarten-age (age 6) measures by applying a neural networks model. We use a set of 49 variables assessed during kindergarten from the domains of early arithmetic skills, cognitive skills, the home learning environment, parental measures, motivation, behavioral problems, and gender, which have been shown to have associations with mathematical development and/or MD. A two-step approach was used: First, we examined whether the neural networks approach can provide a solution for the effective early identification of MD based on all 49 variables and, then, by using the most important predictors as identified by the initial model. The initial model achieved an area under the curve (AUC) of .818, demonstrating excellent performance. The most important predictors of Grade 6 MD came from the domains of arithmetic and cognitive skills (arithmetic skills, rapid automatized naming, number concepts, spatial skills, counting) and behavioral problems (attention-orientation). The model with only the most important predictors achieved an AUC of .776, indicating good performance. Our results provided proof of concept for using neural networks in MD prediction in Grade 6 using information already available in kindergarten. In schools, these results could be used to identify children at potential risk of developing MD and to provide access to early support. Educational Impact and Implications Statement Approximately 4%-15% of children suffer from math difficulties (MD), and many more struggle with them without a formal diagnosis. MD have been shown to put children at increased risk of lower academic achievement, lower motivation, anxiety, depression, and even higher unemployment. Predicting MD accurately and early facilitates timely support. The current study demonstrates the potential of neural networks models to facilitate the early identification of those at risk of developing MD. The performance of our model provided proof of concept for using neural networks for the prediction of MD in Grade 6 using information already available in kindergarten. In a school setting, such prediction knowledge could be used to identify children at potential risk of developing MD and to provide access to early support.