Prediction of Shared Bicycle Demand with Wavelet Thresholding

Consumers are creatures of habit, often periodic, tied to work, shopping and other schedules. We analyzed one month of data from the world's largest bike-sharing company to elicit demand behavioral cycles, initially using models from animal tracking that showed large customers fit an Ornstein-Uhlenbeck model with demand peaks at periodicities of 7, 12, 24 hour and 7-days. Lorenz curves of bicycle demand showed that the majority of customer usage was infrequent, and demand cycles from time-series models would strongly overfit the data yielding unreliable models. Analysis of thresholded wavelets for the space-time tensor of bike-sharing contracts was able to compress the data into a 56-coefficient model with little loss of information, suggesting that bike-sharing demand behavior is exceptionally strong and regular. Improvements to predicted demand could be made by adjusting for 'noise' filtered by our model from air quality and weather information and demand from infrequent riders.