A scrollytelling case study on forecasting bike availability across Barcelona’s Bicing network using temporal filtering, feature engineering, and tree-based ensemble models.
The prototype learned the broad rhythm of the system, but it also exposed the harder truth: bike-share prediction is a spatial rebalancing problem, not just a tabular regression task.
This project set out to predict the proportional availability of bikes at any Bicing station at any hour in Q1 2024. What begins as a machine learning exercise quickly becomes an urban systems problem: demand shifts uphill and downhill, anomalous years distort the training distribution, and not every feature that seems intuitive actually carries predictive value.
The source article frames the feature set as a tension between intuition and actual model utility. Time was the strongest signal, spatial context was under-specified, and temperature alone was too weak to stand in for weather-driven suppression.
| Category | Features | Expected role | Observed utility |
|---|---|---|---|
| Spatial | Latitude, longitude, altitude | Capture slope & neighborhood context | Useful, but incomplete without spatial lag |
| Temporal | Hour, day type, month | Encode commuting & leisure cycles | Strongest predictive signal |
| Meteorological | Hourly temperature | Approximate weather deterrence | Weak proxy without rain, wind, humidity |
| Demand context | Capacity, population density | Represent supply pressure & local demand | Moderate, but still blunt |
Forecasting bike availability is not just about one station at one hour. It is about how flows move across a network shaped by commute cycles, topography, and neighborhood asymmetry.
Barcelona’s bike-share network spans flat central districts and steeper uphill neighborhoods. That means bikes do not simply appear and disappear randomly; they drain downhill during commute cycles and empty faster uphill, making rebalancing a spatial systems issue.
Normalizing available bikes by station capacity lets the prediction target stay comparable across the network. That choice makes the output operationally meaningful, but it also raises the bar for the input features: simple tabular signals are not enough to explain the dynamics by themselves.
The original problem statement is strategic: a rider needs more than a short forecast window when station availability depends on wider urban flows.
Before model selection, the project had to answer a more important question: which years still describe the world the model will be deployed into?
The longitudinal availability plot showed two structural breaks: the Bicing 2.0 platform rollout in 2019 and the COVID demand shock in 2020. Including those periods would have polluted the training distribution with behavior that no longer matched the deployment environment.
2019 reflects transition noise from the Bicing 2.0 launch. 2020 reflects a lockdown collapse and anomalous rebound. Training on both would teach the model that those distorted availability baselines are normal.
The project sourced hourly temperature, but Barcelona’s cycling suppression is more tightly coupled to intense rain events than to winter cold. The weather charts help explain why the meteorological channel stayed weaker than expected.
That choice preserved seasonality and public-holiday variation while avoiding regime-change periods that would bias the model away from current operations.
Once the training window is stable, the story shifts to signal design: which features genuinely explain bike availability, and which model family handles those relationships best?
Larger districts dominate raw population totals simply because they cover more area. That makes absolute counts a misleading proxy for bike-share demand. The original article correctly reframes this as a density problem.
Density improves the demand proxy, but the article’s own diagnosis is sharper: a POI density score or neighborhood attractor index would explain cycling demand better than residential density alone.
A single decision tree fits the local structure quickly, which makes it useful for inspection. It also makes it easy to overfit discontinuities in training data when the feature set still contains noise.
Averaging many decorrelated trees reduces variance inside the training distribution. But the article’s key lesson is that lower training MSE is not proof of better learning when the deployment period is structurally different from the memorized patterns.
It was not the most optimistic model on the training set. It was the one that stayed closer to the shape of the real signal once the project crossed into an unseen quarter.
The PDF deck included code-level screenshots for the prototype models. They are useful as supporting evidence of how the classifiers were trained, but not strong enough to carry a full scrollytelling act on their own.
Useful as appendix material because it grounds the narrative in an actual training workflow without interrupting the higher-level comparison of signal design and model behavior.
Kept as supporting reference from the slide deck: informative for process context, secondary to the stronger visual evidence in the audit, feature, and diagnosis acts.
The reveal in the results is not just which model won. It is why the best training score was misleading, why the predictions were systematically smoothed, and what a production version would need to do differently.
The Q1 2024 plot shows the model following broad weekly and seasonal movement while dampening the sharper morning and evening peaks. That smoothing is the visual signature of a system that generalizes some structure but still misses the abrupt transitions that matter most operationally.
The retrospective diagnosis is precise: there was no time-series cross-validation, no explicit regularization target, and too many features that sounded plausible but carried limited generalizable signal.
Spatial lag variables, graph encodings of the station network, and richer meteorological features would have made the model more structurally aware. The project’s own critique is that raw coordinates cannot substitute for relationships across the network.
The case study matters not because it solved Bicing forecasting completely, but because it made the failure modes legible. That makes the prototype strategically useful: it defines what a stronger mobility model must encode next.
Regime-change years can corrupt a training distribution before any algorithm touches the data. Auditing the label over time is as important as model tuning.
Spatial lag, slope differentials, POI density, rain, and rebalancing events describe the actual system better than raw coordinates or temperature alone.
A stronger version would move toward graph-based spatiotemporal learning or a richer boosted-tree pipeline with cyclical encoding and continuous retraining.
Bicing Prediction Models with Machine Learning was developed at the Institute for Advanced Architecture of Catalonia in the Master in City & Technology by Sharon Megan Wong, Diederik de Vries, Nupur Sircar, and Siddharth Nambiar during Digital Tools & Big Data III with Andre Resende.
This version keeps the original Bicing narrative in semantic HTML but restructures it into a Nocturne-style story: system problem, data audit, signal design, model comparison, and forward path. Local assets replace the source page’s remote figures wherever equivalents were available, while the slide-deck code exports are kept as appendix material rather than promoted into the main scroll.
