Build a Click-Through-Rate Prediction Model with Logistic Regression and Feature Crosses Free
Predict click probability for 480 million daily ad requests with hashed feature crosses and FTRL logistic regression, cutting normalized entropy from 0.92 to 0.795 while holding calibration at 1.00 and scoring 600 candidates in 7ms.
Build a Credit Risk Scoring Model with Logistic Regression and Calibration Free
Score loan applications for 12-month default probability using weight-of-evidence binning and regularized logistic regression, lifting Gini from 0.36 to 0.58 while cutting expected calibration error from 0.071 to 0.006.
Build a Customer Churn Classifier with Gradient Boosted Trees Premium
Predict which subscribers will cancel in the next 30 days from behavioral event snapshots using LightGBM, lifting PR-AUC from 0.19 to 0.42 over a logistic baseline while scoring each user in under 20ms.
Build a Customer Lifetime Value Regression Model Free
Predict 24-month net revenue per customer on a zero-inflated, heavy-tailed target using a LightGBM Tweedie objective, lifting top-decile revenue capture from 31% to 52% and Spearman correlation from 0.41 to 0.68.
Build a House Price Prediction Model with Regularized Linear Regression Free
Predict home sale prices from tabular listing data using Ridge and Lasso regression, cutting test RMSE by 18% over plain OLS while keeping the model fully interpretable for underwriting review.
Build a Loan Default Predictor with XGBoost and SHAP Explainability Free
Predict 12-month default on unsecured personal loans with XGBoost and per-decision SHAP attributions, lifting Gini from 0.41 to 0.58 while producing adverse action reasons a lender can legally send to every declined applicant.
Build a Multi-Class Product Categorization Classifier with Naive Bayes and TF-IDF Free
Assign marketplace listings to one of 1,200 leaf categories from seller-written titles using TF-IDF and Complement Naive Bayes, lifting macro-F1 from 0.38 to 0.67 while scoring each listing in under 3ms.
Build an Imbalanced Fraud Classifier with SMOTE and Cost-Sensitive Learning Free
Catch card-not-present fraud at a 0.17% base rate by combining cost-sensitive gradient boosting with SMOTE, lifting PR-AUC from 0.124 to 0.612 and cutting net fraud loss 42% inside a 40ms scoring budget.