Missing data in time series can severely compromise the reliability of forecasting, anomaly detection, and system monitoring tasks, particularly in domains where data integrity is critical. This work proposes a hybrid imputation framework that integrates Least Squares Support Vector Machines for modeling long-term trends with a lightweight ensemble of residual learners to capture short-term variations. Unlike traditional single-step imputation methods that often struggle to generalize under high missingness, the proposed approach adopts a decoupled structure that isolates smooth global patterns from localized fluctuations, enhancing both accuracy and interpretability. The residual ensemble dynamically adjusts its weights based on cross-validated performance, allowing the most reliable models to contribute more heavily to the final prediction. Additionally, the use of temporal, harmonic, and lag-based features in residual modeling increases robustness to irregular and non-random data gaps. Grid-based optimization further enables automated tuning of hyperparameters for the trend component, ensuring adaptability across diverse missingness scenarios. The result is a modular and scalable solution that consistently outperforms conventional techniques by preserving structural fidelity and capturing fine-grained dynamics, even under challenging real-world conditions.