Estimating daily bicycle counts with Strava data in rural and urban locations

Reliable information on daily bicycle traffic provides a fundamental basis for city planners and scientists. To estimate daily bicycle counts for various German locations with different degrees of urbanisation, we applied Generalised Boosted Regression Models. Altogether 44,136 daily datapoints from 46 counter locations covering a time period of four years were considered. Crowdsourced fitness tracker data from Strava, socio-demographics, land use and weather data were used as independent variables. Our results indicate that weather has the strongest influence on estimated bicycle counts, exceeding the relevance of fitness tracker data. In an overall model daily bicycle counts were estimated with a mean absolute percentage error (MAPE) of 27.9 %. In terms of location-specific estimations, a MAPE of 11.2 % was reached. With our approach, high-quality out-of-sample predictions are also feasible. Based on our estimations, we assume the volatility of fitness tracker user share to have a major impact on model accuracy.