Anomaly classification involves predicting whether an anomaly has occurred and determining its type. In industrial settings, it is decisive not only to detect and classify anomalies but also to understand their impact on other system components. Identifying the causal relationships between an anomaly and its effects is vital for effective damage control and system improvement. In this paper, we explore the use of large language models (LLMs) for variable selection, with a focus on explainability, in the context of fan coil unit anomaly classification. We evaluate six LLMs alongside five time-series classifiers, finding that LLMs significantly enhance feature selection for anomaly classification. Among the models, MetaMath-Mistral-7B proved the most effective. Additionally, our results show that LLMs can identify for some problems likely causal pathways between anomalies and their effects, supporting explainable decisions in variable selection.