The recent application of accelerometers for remotely classifying animal behaviours has improved our understanding of the ecology and physiology of many taxa. Statistical models are employed by researchers to predict specific behaviours using incremental changes in an animal’s acceleration. However, model construction is not standardised with covariates often selected and modified arbitrarily. The aims of this study were twofold: 1) To produce a rigorously tested and repeatable statistical method for accurate classification of ecologically important behaviours from low-frequency (1Hz) acceleration recordings, 2) Apply this method to determine the behaviour of a wild population of Australia’s largest terrestrial predator, the dingo (Canis dingo). We categorised six behaviours using video footage of captive dingoes fitted with accelerometers and manually annotated them to the raw acceleration data. The predictive ability of five widely employed classification models was compared by systematically testing different combinations of covariates followed by cross-validation. We then applied our best model to accelerometer data recorded from wild dingoes at Kalamurina, South Australia. Our best model predicted walking, running, trotting, foraging, standing, and lying down with >85% accuracy. Our study is the first to quantitatively measure behaviours in free-ranging dingoes, revealing a proclivity for lying down and no significant relationship between dingo activity and time of day. Interestingly, dingoes foraged in a temporally cyclical pattern, specific to individuals. The ability to classify ecologically important behaviours using such low-frequency data enables months of per-second behavioural information, which when integrated with GPS data will catapult our understanding of how animals interact with living systems.