Machine Learning with context, engineering and results.

I built a LightGBM model to identify clients with a higher risk of late payment, handling the class imbalance and comparing the result with the previous model used as a baseline. AUC rose from 0.719 to 0.764 and, in a simulation with 30,752 predictions, the new model found 218 more clients who would pay late, representing over ¥21.88 million in potential losses identified.

I analyzed the behavior of 11,573 bank customers to understand what was related to credit card cancellation, then compared Decision Tree, Random Forest and XGBoost to anticipate this churn. XGBoost delivered the best result, with 97.4% accuracy, combining the cancellation forecast with an analysis of the most important variables for each model.

I built a pipeline with Python and MySQL to convert, store and analyze 158,366,934 IMDb and TMDB records, totaling about 12.73 GB of data. On top of this structure, I developed 13 analytical queries and Power BI visualizations on genres, ratings, representation, regions and financial performance, turning a large database into an overview of the film industry.