Seismic time series forecasting remains challenging due to the nonlinearity, non-stationarity, and noise of earthquake data, and because deep learning models are sensitive to preprocessing and hyperparameter settings. Although recent studies have improved neural architectures and optimization techniques, preprocessing is often treated as a fixed or manually designed stage, with limited integration into model optimization. To address this, this paper proposes an integrated, data-driven modelling framework that combines guided preprocessing with systematic hyperparameter optimization for seismic prediction, specifically forecasting earthquake magnitude from seismic catalog time-series data, with experiments conducted on Canadian seismic records. The method uses a Large Language Model to guide data preparation and feature engineering, rather than fully automate them, and applies deep learning-based forecasting with the N-HITS architecture, optimized via metaheuristic-assisted feature selection and hyperparameter tuning. The Football Optimization Algorithm (FbOA), employed as a metaheuristic optimization strategy in this study, is evaluated and compared with several well-known optimizers under identical conditions. The results show significant performance gains, with FbOA achieving superior accuracy, robustness, and convergence compared to baseline and competing methods. Notably, error metrics are reduced (MSE $3.10\times {10}^{-7}$, RMSE $5.57\times {10}^{-3}$), with high performance indicators ( $r=0.982$, $R^2=0.979$, NSE $=0.981$, WI $=0.985$). These results highlight the value of integrating guided preprocessing with optimization and demonstrate a scalable framework for high-precision time-series prediction in geophysical and related domains.