Numerical Binary Classifier Needed

I have a CSV files that contain Timeseries purely numerical features and a binary target column. I want a fully reproducible workflow that starts with exploratory analysis, applies the right pre-processing steps—specifically normalization or scaling, plus intelligent handling of low-variance features and highly correlated predictors to implement rolling window feature selection—and finishes with a robust and well-tuned model that I can drop straight into production.

You are free to compare algorithms (logistic regression, random forest, gradient boosting, XGBoost, stack ensemble or regime based stack model, rolling threshold optimization model) but please justify your final pick with solid cross-validated metrics. Use Python and scikit-learn (other open-source libraries are welcome as long as they are clearly listed in a requirements.txt).

Deliverables:
• Jupyter notebook or .py script covering EDA, preprocessing, model training, evaluation, and inference pipeline
• Saved, versioned model artifact ready for deployment
• requirements.txt with all library versions
• Brief report (ROC-AUC, precision-recall, confusion matrix, feature importance) explaining why the chosen model is the best performer and how it improves on a naïve baseline

Acceptance criteria: the model must outperform the baseline by a meaningful margin on unseen data, and the notebook must run end-to-end with no manual intervention other than placing the raw CSV files in the specified folder.