Closed-Boundary Social Complex System Ensemble Probabilistic Forecasting Engine

This paper proposes an ensemble probabilistic forecasting engine for closed-boundary social complex systems. We treat social system prediction as a probabilistic distribution optimization task rather than searching for a single deterministic future. The framework combines three layers: a large-population pretraining stage on synthetic complex-system dynamics, a transfer learning stage adapted to a small closed-island simulation, and an ensemble forecasting pipeline with ETKF state assimilation, stochastic perturbation schemes, and DBSCAN clustering to identify dominant, secondary, and low-probability high-impact branches. We benchmark three models on a synthetic island-domain dataset: the full pretrain+transfer+hybrid model, a pure Bayesian baseline, and a no-transfer baseline. Evaluation metrics include CRPS, Brier Score, reliability diagrams, main-path enrichment, and warning lead time on synthetic extreme event test sets. Results show that the transfer-enabled model achieves better calibration and sharper forecasts, and identifies high-impact branches earlier than baselines. All experiments are conducted on synthetic data under a unified statistical framework and do not refer to any real individuals or organizations.