Demand forecasting
EcoMoveUS
Built a regression workflow to forecast bike-rental demand from weather, calendar, and temporal features, supporting fleet and staffing decisions.
Random Forest: R² 0.941 · RMSE 50.24
Executive summary
EcoMoveUS needed to forecast bike-rental demand ahead of time to plan fleet rebalancing and staffing, rather than reacting after demand had already been missed. I cleaned and feature-engineered weather, calendar, and temporal data, then benchmarked a linear regression baseline against a Random Forest ensemble. Random Forest explained 94.1% of demand variance versus 41.5% for the linear baseline, with hour of day standing out as the single dominant driver.
The problem
EcoMoveUS needed to anticipate bike-rental demand ahead of time so it could plan fleet distribution and staffing, instead of reacting to demand after it happened and losing rides to empty stations.
The outcome
Trained and compared regression models on weather, calendar, and temporal features; the Random Forest model explained 94.1% of demand variance (R² 0.941, RMSE 50.24), giving a reliable forecast to plan fleet and staffing decisions around.
1. Data cleaning
Range checks on temperature, humidity, and wind speed; removal of invalid records and outliers.
2. Feature engineering
Expanded hour, weekday, season, and business-day fields; added interaction terms (Hour × Temperature, Temp × Humidity); one-hot encoded categorical variables.
3. Modelling
80/20 train-test split (random_state=42); Linear Regression baseline versus Random Forest (100 estimators).
4. Evaluation
Compared RMSE and R² on the held-out test set, then reviewed feature importance for interpretability.
Model comparison
R²variance explained · higher is better
RMSEprediction error · lower is better
Feature importance
Architecture
Average rentals by weekday and hour, from the exploratory analysis — the morning and evening commute peaks are the main driver the model picks up on.