Abstract: Auestion answering for table-text hybrid data in the financial domain is a challenging task. This paper proposes a positive-negative chain-of-thoughts method. Inspired by the influence of positive and negative examples on modeling, chain of thought (COT) samples are extracted from the reasoning results of large language models to construct positive and negative sample pools. A combination of static and dynamic strategies is employed to select different types of demonstration samples: choosing the best positive examples ensures the accuracy of the large language model's output, while selecting similar negative examples guides the model in avoiding incorrect reasoning. The experimental results show that the method improves the accuracy on the FinQA dataset by 3.6%, and on the FinQA-fix dataset by 12.73%.