Supervised Machine Learning for Hourly Bike-Sharing Demand Prediction Using Multivariate Time-Series Features

This study investigates hourly demand prediction for the Divvy bike-sharing system in Chicago using supervised machine-learning models and a multivariate time-series feature set. The analytical workflow integrated data acquisition, relational storage, ETL processing, exploratory analysis, model training, and result visualization. The raw dataset covered July 2020 through July 2022 and contained 11,183,815 trip rows with 13 fields. Trip records were consolidated into an hourly analytical dataset with 18,254 observations and enriched with calendar, seasonal, and weather variables. Exploratory analysis identified strong seasonality, positive skewness, outliers, non-normality, autocorrelation, and evidence of non-stationary behavior. Three regression approaches were evaluated: ordinary linear regression, random forest regression, and XGBoost regression. Under the original 80/20 random train–validation split with a fixed seed of 1234, random forest achieved the strongest holdout performance, with MAE of 81.50, RMSE of 136.85, and out-of-sample R2 of 95.50%. XGBoost also performed strongly, with MAE of 102.86, RMSE of 164.85, and out-of-sample R2 of 93.74%, whereas linear regression reached 58.20% out-of-sample R2. The results indicate that nonlinear tree-based models capture the observed demand patterns substantially better than a linear baseline. Because the original evaluation used a random split on time-indexed observations, the reported scores should be interpreted as predictive holdout performance rather than a strict forward-in-time forecasting benchmark. Future work should reproduce the experiment with chronological validation and systematic hyperparameter tuning.