Financial markets operate under constantly shifting conditions, making it difficult to distinguish between normal fluctuations and fraudulent or manipulative activities. This paper proposes the Adaptive Temporal-Residual Network (ATR-Net), a hybrid framework designed to strengthen fraud detection while keeping the system interpretable for practitioners. The model combines Transformer-based time series forecasting with a Variational Autoencoder (VAE) that analyzes deviations between expected and observed outcomes. To improve accuracy, we incorporated both historical stock data and sentiment extracted from financial news through the Financial News and Stock Price Integration Dataset (FNSPID). In addition, a Market-Adaptive Swarm Intelligence (MASI) algorithm was applied to fine-tune model parameters in response to evolving conditions. The interpretability of results was ensured through SHAP values, which linked anomalies to practical indicators such as trading volume, volatility, and shifts in investor sentiment. Experimental results demonstrate that ATR-Net not only improves detection performance but also provides clear reasoning for its outputs, making it a practical tool for regulators, analysts, and risk managers.