Multi-Scale Forecasting of Citi Bike Demand in New York City: Comparing Time Series, Machine Learning, and Deep Learning Models

Forecasting bike-share demand is central to efficient operations and planning in urban micromobility systems because service quality depends on anticipating when and where bicycles are needed. This dissertation addresses the research question: How accurately can bike-share demand be forecast, and what factors most influence model performance and interpretability? Using complete Citi Bike trip records from 2017–2024 in New York City, augmented with weather, holiday, air-quality, mobility-count, and built-environment inputs, the study evaluates forecasting accuracy across spatial (system, cluster, station) and temporal (daily, hourly) resolutions. A leakage-controlled workflow compares classical time series baselines (ETS and SARIMAX with and without exogenous regressors, and Prophet), tree-based supervised learning (Random Forest and XGBoost), and a station-level CNN-LSTM model. Models are assessed in one-step-ahead forecasting with fixed chronological splits (training 2017–2021, validation 2022, testing 2023–2024) and standard error metrics. Results show that forecast quality varies by temporal resolution and aggregation level, and no single model family dominates across tasks. For system-level daily demand, ETS(A,N,A) with weekly seasonality and exogenous variables is the strongest baseline in the 2023–2024 test period, closely followed by SARIMAX with the same covariates, while Prophet is weaker. Weather accounts for most incremental gains from exogenous inputs. For system-level hourly demand, Random Forest with engineered autoregressive and calendar–weather features outperforms XGBoost. Disaggregating demand to functional station clusters improves the relative performance of supervised learning at both hourly and daily horizons. At the station level, the CNN-LSTM reduces MAE relative to a tabular baseline but increases RMSE, indicating higher sensitivity to large errors in sparse settings. Interpretability analyses align with these patterns, highlighting precipitation, holidays, and temperature as key drivers once weekly seasonality is captured.