Agricultural biosecurity interventions are aimed at minimizing introductions of harmful non‐native organisms to new areas via agricultural trade. To prioritize such interventions, historical data on interceptions have been used to elucidate which factors determine the likelihood that a particular import is carrying a harmful organism. Here we use an interception data set of arthropod contaminants recorded on plant imports arriving in South Africa from 2005 to 2019, comprising 13,566 samples inspected for arthropod contaminants, of which 4902 were positive for the presence of at least one arthropod. We tested 29 predictor variables that have previously been used to explain variation in rates of detection and three variables describing possible sources of additional variation and grouped these into six mutually exclusive “factor classes.” We used boosted regression trees as a non‐parametric stochastic machine‐learning method to build models for each factor class and interactions between them. We explored the influence of these variables with data split either randomly or chronologically. While we identified some specific patterns that could be explained post‐hoc by historical events, only inspected volumes were reliably correlated with detection of arthropod contaminants across the whole data set. However, inspected volumes could not predict future interceptions of arthropods, which instead relied on contextual factors such as country, crop or year of import. This suggests that, although certain factors may be important in certain circumstances or for particular crops or commodities, there is little general predictive power in the current data. Instead, an idiographic approach would be most beneficial in biosecurity to ascertain the details of why a particular pest arrived on a particular pathway and how it might move (and be stopped) in future.