Evolving with Open Worlds: Towards Recursive Self-Improvement for Future Forecasting