The increasing popularity of connected devices and systems in the building sector has led to the democratization of smart buildings. These buildings can be considered as a complex combination of sensors, actuators, software and systems that generate and collect a large and diverse amount of data. This data can be analyzed and used to provide a wide range of services and applications for occupants, third-part users, or the building itself (e.g., control strategies). As a result, its accessibility and availability have become major concerns. However, due to rare and abnormal events, the quantity and quality of the data can become a limitation, hindering the effective deployment of such services. To mitigate this issue, it is essential to identify such anomalous events in order to correct them, if possible. This paper presents a comparison of machine learning models, including Isolation Forest (IF), K-Nearest Neighbors (KNN), and several variants of Autoencoders (AE), for the detection of four types of anomalies commonly encountered in building measurement data: spikes, noise, offsets, and stuck-at faults. A qualitative investigation based on the permutation importance method is conducted to assess the respective influence of data-, model-, and anomaly-related properties on detection performance, using a set of 280 artificially generated test datasets. Results demonstrate that with appropriate feature engineering (or field-expertise), simpler machine learning models (e.g., KNN) can achieve competitive or even superior performance compared to complex deep learning approaches. Therefore, this study highlights that optimizing data representation is more critical for detection performances than increasing model complexity.