Skip to content
AT Ali Taqi
All Work

Boston University · M.A. Statistics · MA 575 · Linear Models · 2021 · Team project

Forecasting Bike-Share Demand

A linear-modeling study predicting daily Capital Bikeshare ridership from weather and season — and correcting for the year-over-year user growth that quietly breaks a naïve forecast.

Fitted versus actual daily registered ridership across 2011, with weekday and holiday observations marked.
Fitted versus actual daily registered ridership across 2011, with weekday and holiday observations marked.
Multiple linear regressionModel selectionCross-validation (LOOCV)Residual diagnosticsR · tidyverse

Given a day’s weather and season, how many people will rent a bike? Working from two years of Capital Bikeshare data, our team built multiple-regression models predicting daily ridership from temperature, humidity, windspeed, weather, and calendar features — with interaction terms for how a hot day plays differently on a workday than on a weekend.

The interesting failure showed up at validation: models fit on 2011 systematically under-predicted 2012. The cause wasn’t the weather — it was a growing user base. We estimated a year-over-year growth ratio (via an environmental-loss-function approach, plus a simpler windowed estimate) to rescale the forecasts, recovering roughly a 10% improvement in prediction error and turning a biased model into an honest one.

Built in R with leave-one-out cross-validation and residual diagnostics throughout. A team project, and a clean lesson in how a model can be locally right and globally wrong.