CycleCast: Bike Demand Forecasting

Bike-share operators must decide how many bicycles to make available at each time and place. Too few bikes leave customers without a ride; too many in low-demand areas tie up equipment and require unnecessary redistribution. CycleCast is a data-science prototype that predicts hourly rental demand from historical bike-share records and contextual features such as date, hour, season, holiday, temperature, humidity, and weather. The project uses the UCI Bike Sharing dataset covering 2011–2012. It excludes casual and registered user counts from the predictor set because they are components of the total-rentals target and would not be known when forecasting. Candidate regression models include Linear Regression, Decision Tree, Random Forest, XGBoost, and LightGBM. Records are sorted chronologically; the earliest 80% are used for training and time-series cross-validation, while the latest 20% are reserved for final evaluation. The notebook is configured to tune XGBoost with RandomizedSearchCV and TimeSeriesSplit and evaluate using RMSE, RMSLE, MAE, and MAPE. The intended users are bike-share operators and city mobility planners. In the current prototype, an analyst loads prepared data into a Python notebook, runs the modeling workflow, and reviews hourly predictions, evaluation tables, and plots to help plan bicycle distribution. The time-aware validation is designed to better represent forecasting future demand. The current deliverable is a notebook prototype, not a deployed city-wide service; final performance metrics will be added after the corrected notebook is rerun.