Affective events classification, which seeks to identify the sentiment associated with short texts that describe events, has garnered significant attention in recent years. Previous studies have highlighted the primary challenges of this task: the implicit nature of event sentiment and the scarcity of labeled data and contexts. While considerable efforts have been made to address these issues, we propose an additional perspective. We argue that the sentiment of an event can vary significantly among different people with diverse conditions and opinions. Recognizing that relevant contexts are frequently not provided alongside the event itself, in this work, we introduce Chain-of-Thought Context-Aware Network (CoT-CAN), which analyzes sentiment by integrating generated context information with the event description. We employ a two-stage chain-of-thought prompt to guide pretrained language models in generating relevant contexts for each event, and introduce an interaction module composed of cross-attention, which effectively integrates contextual information into the event embeddings. We also apply a self-training strategy and propose a new method to update the weights and pseudo-labels of pseudo-labeled data, based on the consistency with historical pseudo-labels. Experimental results on two standard datasets show effectiveness of the proposed approach.