This repository contains all relevant notebooks and code that helped our team 𝐋𝐨𝐰 𝐁𝐢𝐚𝐬 𝐇𝐢𝐠𝐡 𝐕𝐢𝐛𝐞𝐬 secure #5 on the private leaderboard (out of 8000+ participants) in the redBus Data Decode Hackathon hosted by Analytics Vidhya.
We made a strong leap from #11 on the public leaderboard to #5 on the private, with CV and test scores staying closely aligned — always a good sign of a generalizable solution.
Our final score was 533.58587541, just 0.1 RMSE shy of the 2nd Runner-Up spot — a better seed or refined hyperparameters could have made the difference.
Forecast the number of bus seats booked for a given route on a specific Date of Journey (DOJ) using only data available 15 days before the DOJ — including seat search and booking history, route metadata, and other features.
We opted for a simple yet effective strategy centered around a single LightGBM model, powered by exploratory data analysis, thoughtful feature engineering, and targeted post-processing.
- Lag features: 3-day lag of cumulative
seat_searchesfor the current route to mitigate cumulative seat_searches as seen during EDA. - Time-based features: Extracted weekday, day-of-month, long weekends, etc., which provided more signal than raw dates.
- External calendar features: Added regional and national holidays/festivals as a single feature ordinally encoded with respect to its relevance.
- Train/Test Distribution Analysis: Boosted prediction mean and variance to offset LGBM's lack of extrapolation capabilities (unlike neural nets).
- A trusty CV split to hold it all together: Stratified based on route and time, enabling the model to generalize better across unseen patterns.
The test data corresponded to January and February 2025, which saw an unusual spike in bookings due to special events like Mahakumbh and Coldplay concert. These events were not represented in the training data, making CV-to-LB correlation weaker than expected. However, our distribution-aware boosting technique — originally designed to handle LGBM’s extrapolation limitations — turned out to be highly effective in compensating for this shift, giving us a significant edge on the private leaderboard.
Several experiments did not yield improved performance on validation/test splits:
- Target normalization: To battle mean shift trends, as they are not catered by tree-based models (highly train distribution dependent) , but failed probably due to unrelaiable batch-statisitcs.
- TabNet, RuleNet, and MLPs: Unable to generalise seasonality effectively on the CV set.
- Other Route's features: Including bookings/searches on correlated routes and 3-day lag of return journey routes only slightly improved validation scores.
- Classical time-series models: SARIMA and LSTMs failed to outperform CV baseline and suffered on unseen trends.
Thanks to redBus and Analytics Vidhya for organizing a well-structured and practical competition based on real-world data.
A huge shoutout to my teammates Mukil M, MOHD ASHAZ KHAN, and Dilshad Raza — your collaboration, breadth of exploration, and ability to converge on the right path made this solution possible.