From morning departures to evening returns: a probabilistic model for forecasting bike-sharing demand
Abstract
BSS face a persistent operational challenge, spatial and temporal imbalances between stations that force operators into costly manual redistribution. Existing demand-forecasting models, from count data regressions to deep learning architectures, treat each time slot independently and overlook a structural feature of BSS usage the dependency between morning and evening flows. This paper formalizes that dependency through a Hurdle Negative Binomial-Binomial model applied to the VélôToulouse network. For each origin station, the model estimates p, the probability that a morning trip generates a same day evening return, using aggregated flows without individual level data. The model reveals marked spatial heterogeneity : periphery to center pairs show the highest return propensity (median around 0.42, reaching 0.80 on some axes), while periphery-periphery pairs show the lowest (median around 0.20), consistent with commuting behavior. Using p to predict evening demand from morning flows alone yields a substantial accuracy gain over a historical average baseline and a perfect symmetry baseline (33% and 37% reduction in MAE). The resulting net imbalance signal, available as soon as the morning peak ends, shows central stations accumulating bikes while peripheral stations run short,a pattern directly usable for proactive rebalancing. Limitations and extensions toward covariate dependent, dynamically updated versions are discussed.
Related
- Can Geopolitical Risk Improve the Forecasting of Thai Stock-Market Returns? Evidence from Econometrics and Machine-Learning Models
- SDECast: Probabilistic Weather Forecasting in Continuous Time with Neural SDEs
- Probabilistic Forecasting of Electric Vehicle Charging in a Time-varying Hierarchical Infrastructure
- Combining Weather Forecast Aggregation and State-Space Models for Adaptive Probabilistic Electricity Load Forecasting
- Rollcast: Proper-Score Gated Rolling Anchors for Adaptive Probabilistic Time-Series Forecasting
BibTeX
@misc{salifou2026from,
title = {{From morning departures to evening returns: a probabilistic model for forecasting bike-sharing demand}},
author = {Mohamadou Salifou},
howpublished = {HAL},
year = {2026},
month = aug,
url = {https://hal.science/hal-05737851v1},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - From morning departures to evening returns: a probabilistic model for forecasting bike-sharing demand
AU - Mohamadou Salifou
PY - 2026
DA - 2026/08/24
PB - HAL
UR - https://hal.science/hal-05737851v1
AB - BSS face a persistent operational challenge, spatial and temporal imbalances between stations that force operators into costly manual redistribution. Existing demand-forecasting models, from count data regressions to deep learning architectures, treat each time slot independently and overlook a structural feature of BSS usage the dependency between morning and evening flows. This paper formalizes that dependency through a Hurdle Negative Binomial-Binomial model applied to the VélôToulouse network. For each origin station, the model estimates p, the probability that a morning trip generates a same day evening return, using aggregated flows without individual level data. The model reveals marked spatial heterogeneity : periphery to center pairs show the highest return propensity (median around 0.42, reaching 0.80 on some axes), while periphery-periphery pairs show the lowest (median around 0.20), consistent with commuting behavior. Using p to predict evening demand from morning flows alone yields a substantial accuracy gain over a historical average baseline and a perfect symmetry baseline (33% and 37% reduction in MAE). The resulting net imbalance signal, available as soon as the morning peak ends, shows central stations accumulating bikes while peripheral stations run short,a pattern directly usable for proactive rebalancing. Limitations and extensions toward covariate dependent, dynamically updated versions are discussed.
ER -