In this notebook we builds a regression model to preduct forest-fire burned area. It loads the classic forest-fires dataset, stabilizes the target using a logarithmic transform, and rebalances extreme burned-area cases using SMOGN. It then prepares features with cyclical encodings, trains several regression models, tunes hyperparameters, compares performance in a leaderboard. The final model is evaluated with cross-validation and can be saved or reloaded for deployment.