This article introduces a data-driven and model-free framework called Feature-and-Reconstructed Manifold Mapping (FRMM) for predicting multiple observations in complex systems. FRMM combines feature embedding and delay embedding to identify low-dimensional manifolds that represent the system's essential dynamics. These manifolds are used as generalized predictors to forecast all components of the system. The framework is validated on both model systems (e.g., Lorenz and Rössler systems) and real-world data (e.g., Indian monsoon, EEG signals, foreign exchange market, and traffic speed in Los Angeles). FRMM effectively addresses the curse of dimensionality and provides reliable predictions for high-dimensional systems. It outperforms traditional methods in terms of accuracy and robustness, particularly in handling complex, high-dimensional, and time-varying systems. The framework is also shown to be effective in identifying tipping points and predicting future states of complex systems. FRMM uses Gaussian process regression to train mappings between low-dimensional representations, making it a versatile and interpretable method for forecasting in various domains. The study highlights the potential of FRMM for applications in diverse real-world systems, demonstrating its ability to predict multiple observations accurately and efficiently.This article introduces a data-driven and model-free framework called Feature-and-Reconstructed Manifold Mapping (FRMM) for predicting multiple observations in complex systems. FRMM combines feature embedding and delay embedding to identify low-dimensional manifolds that represent the system's essential dynamics. These manifolds are used as generalized predictors to forecast all components of the system. The framework is validated on both model systems (e.g., Lorenz and Rössler systems) and real-world data (e.g., Indian monsoon, EEG signals, foreign exchange market, and traffic speed in Los Angeles). FRMM effectively addresses the curse of dimensionality and provides reliable predictions for high-dimensional systems. It outperforms traditional methods in terms of accuracy and robustness, particularly in handling complex, high-dimensional, and time-varying systems. The framework is also shown to be effective in identifying tipping points and predicting future states of complex systems. FRMM uses Gaussian process regression to train mappings between low-dimensional representations, making it a versatile and interpretable method for forecasting in various domains. The study highlights the potential of FRMM for applications in diverse real-world systems, demonstrating its ability to predict multiple observations accurately and efficiently.