Recent advances in artificial intelligence and financial analytics have underscored the pivotal role of large language models in financial practice. However, their susceptibility to hallucinations and limited numerical reasoning capabilities raise significant concerns, as numerical accuracy is paramount in financial applications. In this paper, we introduce Self-Correction via Semantic Information (SCSI), a novel approach that enables LLMs to autonomously refine their numerical understanding without relying on external data or prior knowledge. Our framework quantifies semantic uncertainty in generated outputs along with semantic embedding produced during inference to assess numerical comprehension and triggers a self-correction mechanism when low confidence is detected. Experimental evaluations on two open financial datasets and three open-source LLMs demonstrate that SCSI framework enhances numerical reasoning performance across benchmark LLMs, surpassing state-of-the-art techniques such as Chain-of-Thought prompting. Overall, our method contributes to both artificial intelligence and financial analytics by offering a robust and interpretable framework that effectively mitigates concerns regarding numerical veracity in financial analytics and foresters a synergistic cross-fertilization between fintech and AI.