Time series forecasting has been extensively studied for decades and remains fundamental across diverse real-world domains. Recent advances in deep learning have
enabled models to capture complex and non-stationary temporal patterns more effectively. ...
Time series forecasting has been extensively studied for decades and remains fundamental across diverse real-world domains. Recent advances in deep learning have
enabled models to capture complex and non-stationary temporal patterns more effectively. However, deep learning rely on normalization for training stability and on large-scale data, which amplify multimodality in time series and introduce a trade-off between generalization and precise pattern modeling. In this paper, we propose Latent Aligned Dependent Model (AlignDep), which models inter series dependency by aligning entangled information in input time series within a latent space before incorporating relational information. Moreover, we design a Subspace-Aware Feed-Forward Block to mitigate structural limitations of window embedding models in capturing local temporal patterns. Across widely used benchmark datasets, our approach consistently reduces forecasting error compared to existing window embedding baselines. Comprehensive analyses further validate the structural strengths and model behavior of the proposed architecture.