In the financial world, a transaction graph is commonly used for modeling the ever-changing payee-payor relationships. Every online transaction corresponds to a directed edge from the paying party to the receiving party in this graph. Even though the superior learning capability of Graph Neural Networks (GNNs) has led to several successful financial applications like fraud detection and anti-money laundering, most of these existing works do not have fairness considerations. Apparently, the lack of fairness guarantees during the GNN-based decision-making process would cause increasingly serious societal concerns from both buyers and sellers. Furthermore, the time-varying property of the financial networks makes the fairness requirements more challenging, since current fairness measures on graph learning tasks and fairness-aware GNN models are all designed for static graphs only. In this work, we present a new generic definition of individual fairness for dynamic graphs and propose a regularization-based method to debias the GNN model in the temporal setting. We perform some preliminary experimental evaluations on two real-world datasets and demonstrate the potential efficacy of the proposed methods.