With the development and proliferation of social networking services, the widespread dissemination of false and misleading information—so-called fake news—has become a serious societal issue. Fake news can negatively affect various domains, including politics and economics, making early detection critical. Recent studies on automated fake news detection have primarily leveraged language models, with growing evidence that incorporating social context improves classification performance. However, as social context accumulates over time, a trade-off emerges between detection accuracy and timeliness. To address this challenge, we propose an early fake news detection method based on sequential analysis of social context and progressive threshold adjustment. Our method aims to enhance classification performance in the early stages of diffusion by removing news articles with high-confidence predictions from further analysis. By dynamically lowering the decision threshold as diffusion progresses, the method balances precision and recall—favoring confident early predictions while maintaining broader coverage later. Experimental results on two real-world datasets demonstrate that our method outperforms the baseline method in terms of F1 score during early detection stages, confirming its effectiveness in early-stage fake news identification.