Handling anomalies is crucial in multivariate time series forecasting. However, existing methods treat anomaly preprocessing and model training separately, relying solely on data distribution and ignoring spatiotemporal cues. This often leads to incorrect anomaly handling and the loss of valuable training data. To address this, we propose STTS-EAD—an end-to-end framework that integrates Spatio-Temporal learning-based Time Series prediction with Embedded Anomaly Detection. STTS-EAD jointly optimizes forecasting and anomaly detection by alternating between the two tasks during training, leveraging spatiotemporal information to identify and utilize latent anomalies effectively. To the best of our knowledge, STTS-EAD is the first to unify anomaly detection and forecasting within the training phase for multivariate time series. Experiments on a public stock dataset and two real-world sales datasets demonstrate that STTS-EAD significantly improves forecasting accuracy and outperforms existing baselines by effectively addressing anomalies during training.