The present study examined the predictive validity of a dynamic test of decoding in which participants are taught three novel letters and how to synthesize the corresponding letter sounds into new words. One version of this dynamic test was administered to 158 kindergarten children before the onset of formal reading instruction along with traditional predictors of reading. Similarly, a parallel version of the dynamic test was administered to the same children after a few months of formal reading instruction. At the end of grade 2, the children were assessed on outcome measures of reading and categorized as having disabilities with either accuracy or fluency measures. Administered before as well as after the onset of formal reading instruction, the dynamic test of decoding contributed uniquely to the prediction of difficulties with reading "accuracy" at the end of grade 2 after control for traditional predictors of reading. Difficulties with reading "fluency" were also predicted by the dynamic decoding test, but the unique prediction value was more limited. This study showed that a dynamic assessment of decoding can be a useful addition to traditional test batteries for early identification of children at risk for reading disabilities. Even when taken before formal reading instruction, a combination of the dynamic assessment and two traditional measures (letter knowledge and rapid automatized naming) yielded a very high prediction accuracy of reading difficulties at the end of grade 2.