For the dynamic transaction networks, financial fraud detection has been an ongoing major challenge due to the constantly increasing interactions, temporal dependencies and hidden casual relationships. Conventional graph-based approaches process structural and temporal patterns well, but false correlations sometimes are falsely identified as causal relationships, leading to unreliable predictions during the temporal evolution. To tackle this, in this paper we introduce CaT-GNN, a Causal-Temporal Graph Neural Network that integrates causal reasoning with temporal graph representationlearning. The proposed model includes four essential modules, i.e., a Temporal Graph Encoder (TGE) for time-evolving embeddings capture, a Causal Inspector (CI) for causal attention weights learning, a Causal Intervener (CE) for intervention-based debiasing and a Classifier (CF) to estimate fraud probability from causally refined representations. Extensive experiments on benchmark datasets, including Elliptic and PaySim, demonstrate that CaT-GNN outperforms state-of-the-art temporal and causal graph models in terms of AUC and F1-score, while exhibiting superior robustness and interpretability across temporal and out-of-distribution shifts.