Building sensor anomaly detection is crucial for operational efficiency and equipment reliability, yet standardized evaluation frameworks are lacking, leading to inconsistent method selection. This paper presents a comprehensive comparison of anomaly detection methods using a novel unsupervised evaluation framework that incorporates statistical consistency, temporal coherence, cross-method agreement, and domain logic compliance. We evaluate three paradigms—statistical (STL), machine learning (Isolation Forest), and deep learning (LSTM Autoencoder)—on real-world building sensor data spanning two years with eight sensor types. Our results show that Isolation Forest achieves the best overall performance with a total score of 0.464 and superior computational efficiency of 0.339 seconds, detecting four collective anomalies with excellent temporal coherence. Surprisingly, the LSTM Autoencoder achieves only a 0.263 total score despite 206.444 seconds training time, challenging assumptions about deep learning superiority. These findings demonstrate that well-designed traditional methods can outperform complex deep learning approaches when considering computational efficiency and practical deployment requirements, providing essential guidance for building monitoring system design.