Weather changed 2012 Capital Bikeshare daily casual share error by -0.20 percentage points, with a 95% block interval of -0.75 to +0.44, for ridge models fitted on 2011

We tested whether recorded daily weather improves estimates of the casual share of Capital Bikeshare rentals in a later year. We fitted fixed ridge models on all 365 dates in 2011 and tested all 366 dates in 2012. The calendar model used month, weekday, holiday, and day group terms. Other models added temperature, humidity, wind, and weather class. Each date had equal weight. Adding all weather terms changed mean absolute error from 3.574 to 3.371 percentage points. The paired change was -0.203 points. Its 95% block interval was -0.746 to +0.443 points when both years were resampled and models were fitted again. The all-weather model did not clearly beat the temperature model either: +0.005 points, interval -0.367 to +0.485. Its error fell on weekends and holidays but rose slightly on working days. Its mean positive bias grew from 0.604 to 1.587 points. All planned model checks kept small negative error changes relative to calendar, with fixed-fit intervals that included zero. A model with other weather terms but no temperature had a smaller point error than the all-weather model. These results do not show that weather has no useful information. They describe this daily outcome, model family, and year split. The test uses recorded weather and does not identify a causal effect or a forecast before weather is known. The package preserves source data, provenance, the local plan, figures, all daily predictions, bootstrap draws, checks, and repeat instructions.

bike sharing · human behavior · weather · rental share · temporal transfer · model evaluation · uncertainty