归因分析是根据“结果”发生的条件而推导出因果关系结论的分析过程，可以用来识别变量间的因果关系。其中，利用结构因果模型可以用来解决反事实层问题，即想象中的没有在现实中发生的事件，如“假如地震没有发生，现在会什么样？”解决这一层问题后，我们可以得到关于因果关系的洞见。本文利用循环神经网络来构建结构因果模型。然后利用企业财务指标的数据训练神经网络，来预测介入操作后非现实中发生的“结果”的值，从而得到介入分布上的平均因果效应。得到反事实的结果后，即可推导因果关系。实验结果表明，保持销售额对公司稳定发展至关重要，流动债务对维持公司资金流动和合理的资产结构至关重要，而公司的突然缩水可能是由不合理的收购造成的。

Attribution methods are used for deducing causality according to the conditions of the "outcome", which can identify causal relationship among variables. One of the methods is to address counterfactual problems with structural causal models, namely imaginary events that never happened in reality. We can gain knowledge about causality after answering questions like "what would happen if the earthquake had not happened?". Our method is to view feedforward neural networks, especially recurrent nueral networks, as structural causal models. Then we train these networks with enterprises' financial data to obtain counterfactual "outcome" after intervention, and calculate the average causal effect on the interventional distribution. Accordingly, we can figure out features that have most significant causal effects and deduce the causal relationship. As our result suggests, it is crucial for enterprises' steady development to maintain sales. Current liability is also important to guarantee fund flow and asset structure. Besides, sudden shrink of enterprises' assets is likely caused by an unreasonable acquisition.