This was a great challenge and I’m grateful to have achieved 1st place on the private leaderboard, also my first 1st-place finish on a private LB, which is awesome! Huge thanks to my teammate @wuuthraad and to the competition organizers for setting up such a challenging problem.

Our solution used LightGBM with carefully constructed temporal features to predict congestion across all required future horizons. One of the biggest challenges was the weak correlation between local CV and the public leaderboard, which meant trusting cross-validation more heavily and tracking experiments and improvements thoroughly. Because of the strong class imbalance, we also relied on a two-stage balancing strategy, combining downsampling of fully free-flowing sequences with per-fold oversampling of minority -classes during training to optimize F1-macro without leaking information into validation along with an F1 eval function for early stopping.

The training data is constructed using strict, gap-free 15-minute windows, grouped by camera, and only samples with all future targets present are retained after building out lag features. This enforces temporal integrity and fully respects the embargo and real-time inference constraints. I also experimented with sequence-based LSTM models trained on full 15-minute multivariate inputs to predict all future minutes jointly, but these did not outperform the per-horizon LightGBM models in terms of generalization.

In addition, I explored video-derived features extracted using YOLO tracking (e.g. average, max, and standard deviation of vehicle counts across frames). These features consistently improved LSTM validation performance, but did not improve LightGBM results and did not generalize to the test set. In practice, the tree-based models appeared to generalize better by focusing on cleaner temporal signals, while higher-capacity sequence models were more sensitive to noise introduced by video-level features on testing.

Please find the model training notebook and repository for our final solution at the link below:

Notebook: https://github.com/daniel-bru/Barbados-Traffic-Analysis-Solution/blob/main/modelling_v2.ipynb

Repo: https://github.com/daniel-bru/Barbados-Traffic-Analysis-Solution/tree/main

Repo & notebook notes:End-to-end experiments can be run with `modelling_v2.ipynb`. Each experiment is saved under `lgbm_training_history/`, just update the notebook, and set a new name for the experiment in the notebook config.

Curious to hear others' approaches and whether extracting features from the videos helped improve your models