Accelerometery is revolutionising the field of behavioural ecology through its capacity to detect the fine-scale movements of animals resulting from their behaviour. Because it is often difficult to infer the behaviour of wildlife on a continuous basis, particularly for cryptic species, accelerometers potentially provide powerful tools for remote monitoring of their behavioural responses to the environment. The goal of this study was to provide a detailed, calibrated methodology, including practical guidelines, to infer the behaviour of free-ranging animals from acceleration data. This approach can be employed to reliably infer the time budget of species that are difficult to observe in certain environments or at certain times of the day. To this end, we trained several behavioural classification algorithms with accelerometer data obtained on captive roe deer, then validated these algorithms with data obtained on free-ranging roe deer, and finally predicted the time-budgets of a substantial sample of unobserved free-ranging roe deer in a human-dominated landscape. The best classification algorithm was the Random Forest which predicted five behavioural classes with a high overall level of accuracy (approx. 90%). Except for grooming (34-38%), we were able to predict the behaviour of free-ranging roe deer over the course of a day with high accuracy, in particular, foraging head down, running, walking and standing (68-94%). Applied to free-ranging individuals, the classification allowed us to estimate, for example, that roe deer spent about twice as much time foraging head-down, walking or running during dawn and dusk than during daylight or nighttime. By integrating step by step calibration and validation of accelerometer data prior to application in the wild, our approach is transferable to other free-ranging animals for predicting key behaviours in cryptic species.