With recorded weather, a 28-day count scale reduced 2012 mean absolute error by 57% for a fixed 2011 Capital Bikeshare model

We tested whether recent count totals improve a rental model when it is used in a later year. We fitted a fixed calendar and weather model on 8,645 Capital Bikeshare hours in 2011. We scored 8,734 observed hours in 2012. Each target hour used recorded weather. A daily scale divided observed counts by fixed predictions over the previous 28 calendar days. No target-date count entered the scale. Mean absolute error fell from 90.18 to 39.12 rentals per hour, a 56.6% reduction. The paired reduction was 51.06 rentals per hour, with a nominal 95% interval of 43.11 to 58.53 from blocks of 14 dates. Bias changed from -85.12 to 0.20 rentals per hour. Error fell in all 12 months. A prior-week count reference had error of 60.60. A monthly refit on the previous 90 days, added after the main test, had error of 45.70. Checks with 7-day and 56-day scales, Poisson loss, and a calendar model without weather kept the reduction. The fixed model has no explicit growth trend. This is one historical system and assumes known target weather. It does not establish a new algorithm, a causal account of travel choices, or the best possible forecast. The package preserves the source, all predictions, code, checks, and run instructions.

bike sharing · demand prediction · time series · model calibration · recent counts · temporal validation · open data