Abstract: In order to gain insight into those training regimens that can minimise the risk of fracture in athletic populations, we conducted a large epidemiological study in racehorses. Thoroughbred racehorses provide a suitable model for studying fracture development and exercise-related risk factors in physically active populations. They represent a homogeneous population, undertaking intensive exercise programmes that are sufficiently heterogeneous to determine those factors that influence injury risk. Daily exercise information was recorded for a cohort of 1178 thoroughbreds that were monitored for up to 2 years. A total of 148 exercise-induced fractures occurred in the study population. Results from a nested case-control study showed a strong interactive effect of exercise distances at different speeds on fracture risk. Horses that exceeded 44 km at canter (< or =14 m/s) and 6 km at gallop (>14 m/s) in a 30-day period were at particularly increased risk of fracture. These distances equate to ca. 7700 bone loading cycles at canter and 880 loading cycles at gallop. Fifty-six fractures occurred in the subset of study horses that were followed since entering training as yearlings, when skeletally immature (n = 335). Cohort analysis of this data set showed that, in previously untrained bones, accumulation of canter exercise increased the risk of fracture (P < or = 0.01), whereas accumulation of high-speed gallop exercise had a protective effect (P < 0.01). However, increasing distances at canter and gallop in short time periods (up to one month) were associated with an increasing fracture risk. All training exercise involves a balance between the risk of fracture inherent in exposure to loading and the beneficial effect that loading has by stimulating bone cells to produce a more robust architecture. Results from our study provide important epidemiological evidence of the effects of physical exercise on bone adaptation and injury risk and can be used to inform the design of safer exercise regimens in physically active populations.