I need a data-driven engine that can forecast winning numbers for a national lottery by marrying classical probability theory with modern machine-learning techniques. Historical draw data will serve as the foundation; from that, I want to extract frequency distributions, transition matrices, and other statistical features, then feed them into supervised models so both perspectives inform each prediction round. The workflow should cover data gathering (or cleaning if I supply the files), feature engineering, model training and back-testing, plus a simple interface—CLI notebook or lightweight dashboard—that lets me refresh data and generate new number sets on demand. Python with pandas, NumPy, scikit-learn, and either TensorFlow or PyTorch fits my stack preference, but I’m open if you can justify alternatives. Deliverables (all source included): • Clean, well-documented codebase and reusable pipeline • README or notebook explaining methods, assumptions, and how to rerun the model • Final trained model(s) with a short performance report (hit rate, lift vs. random) • Instructions for future retraining when new draws are published Target completion: within one month. I’ll be available for quick feedback throughout so iterations move fast and the hybrid approach reaches the best possible accuracy minimum 85%.