Academic team research on online passenger-load forecasting for ten bus stops in Salvador, Brazil.
Python · CapyMOA · pandas · NumPy · scikit-learn · Jupyter
Paper · Experiments · Setup · Data · Team
Developed for COMPX523-25A · Machine Learning for Data Streams, University of Waikato, 2025. The study predicts the average five-minute passenger load over the next 30 minutes, using observations from ten stops between March and May 2024.
The experiments compare autoregressive approaches, differencing, station/time encodings and contextual features with KNN, FIMT-DD and Adaptive Random Forest regressors. A prototype uses predicted crowding to suggest alternative stops.
The paper presents the study and its findings. Reproduction checks cover the full model comparison, feature importances and recommendation examples.
All 25 experiment/model combinations in Table 2 have been reproduced. Every MAE and RMSE matches the precision published in the paper. The complete comparison records the values for all 50 metrics.
| Approach | MAE | RMSE |
|---|---|---|
| Naive baseline | 7.431 | 17.863 |
| Per-stop autoregression, variant 1 · KNN | 6.336 | 14.702 |
| Integrated model · FIMT-DD | 6.855 | 14.892 |
| Seasonal integrated model · FIMT-DD | 6.240 | 14.104 |
Adding time and station information improved the best reported model; adding the complete set of external variables did not improve it further. The full comparison and experimental assumptions are in the paper.
![]() |
![]() |
| Error over time · original paper figure | Station crowding · retained simulation output |
The recommendation prototype assumes compatible bus routes and feasible walking between stops. It does not model actual walking routes or transfers.
- The consolidated experiment notebook, with paths adapted to this folder structure.
- Historical transport-data exploration and earlier classification/recommendation prototypes, clearly separated from the main experiments.
- Original Python model and preprocessing scripts under
src/legacy/. - Selected aggregate transport/context data and stop coordinates, with sources and attribution.
- The final academic paper, original figures and retained animation.
| Team member | Contributions |
|---|---|
| Aitana Cantó Rico | Conceptualization, analysis, investigation, methodology, software and original draft; development and evaluation of the station/hour time-aware baseline with FIMT-DD, KNN and Adaptive Random Forest. |
| Christopher Mayo | Conceptualization, data curation, analysis, investigation, methodology, software, validation, visualization and writing. |
| Ryan Manchester | Conceptualization, resources, software and presentation. |
| Bohui Zhong | Conceptualization, methodology, software, visualization and original draft. |
| Matthew Rolleston | Conceptualization, analysis, investigation, methodology, software and visualization. |
| Jayden Litolff | Conceptualization, data curation, analysis, investigation, methodology, project administration, software and visualization. |
The project was developed collaboratively by the team.
notebooks/
experiments.ipynb Main experiment and application notebook
archive/ Earlier exploration and prototypes
src/legacy/ Retained Python models and preprocessing
data/
raw/5min/ Aggregate transport observations
raw/30min/ Aggregate inputs used by earlier prototypes
processed/ Derived targets and contextual features
reference/ Bus-stop coordinates
docs/
paper.pdf Submitted academic paper
images/ Paper figures and retained animation
SETUP.md Environment and notebook instructions
REPRODUCIBILITY.md Provenance, known issues and validation limits
scripts/check_project.py Offline structural and data checks
outputs/ Ignored local results
To inspect the prepared files without installing ML dependencies:
python scripts/check_project.pyFor the notebooks and full experiment runner, follow local setup. The tested environment uses CapyMOA 0.8.2, Python 3.12 and Java 17. Reproduction notes summarise the checks and evaluation limitations.
The source study is Autoregressive Modelling for Streaming Time Series Data, COMPX523-25A Assignment 3, June 2025. Credit all six authors when describing its findings.
Transport observations derive from SUNT, DOI: 10.17632/85fdtx3kr5.1. Online models use CapyMOA. See data attribution for dataset sources and reuse conditions.

