Bike Demand Forecasting (Hourly)

Forecast the total number of bikes rented in an hour from calendar and weather information. This is a new project in the portfolio and uses a chronological holdout so the test set comes strictly after training and validation in time.

The UCI hourly file contains 17,379 observations from 2011-01-01 through 2012-12-31. The model starts from 12 usable source signals and adds 8 calendar features, for 20 input columns

Seven regression baselines and model families are compared: mean baseline, Ridge, Decision Tree, Random Forest, Extra Trees, Gradient Boosting, and tuned HistGradientBoosting. The histogram gradient model is tuned with seven randomized configurations and three expanding-window TimeSeriesSplit folds using MAE. The winning model is selected by validation MAE, then refit on train plus validation before test evaluation.