Forecasting demand for bus journeys 15 days in advance using historical booking data, search trends, and enriched holiday features.

This repository contains my top-performing solutions for the RedBus DataDecode Hackathon 2025, hosted by Analytics Vidhya in collaboration with redBus. The challenge was to forecast route-level seat demand for a specific date of journey (DOJ), 15 days in advance, using historical booking and search data.

📈 Final Rank: 21
🏆 Recognized for analytical rigor and creative feature engineering.

Predict total seats booked for each route on a specific date of journey, using data available 15 days in advance.

👉 Click to visit the detailed problem statement: RedBus DataDecode Hackathon 2025

• Impact of holidays, weekends, school calendars, and wedding seasons.
• Regional variability in holiday effects.
• Day-of-week and temporal search trends.
• Noise in user search behavior and booking delays.

redBus-DataDecode_rank21_Sol/
├── BestSol/
│   ├── 15.06.25_v4.ipynb          <- Final best performing solution
│   ├── submission.csv             <- Corresponding submission file
│   ├── LB Standings .png          <- Proof of leaderboard rank
├── AnotherSol/
│   ├── 22.06.25_v1.ipynb          <- Alternate solution with holiday enrichment
│   ├── submission_with_holidays.csv
│   ├── holiday_dates.csv          <- Holiday data scraped manually (2023–2025)
│
│── Data/
│   ├── train.zip            <- train data 
│   ├── test.csv             <- test data
│
├── Model Architecture.png         <- Model Block Diagram
└── README.md                      <- This file

Visual pipeline of the solution:

🖼️ View Block Diagram

• 🧠 ML Models: LightGBM, XGBoost (with fine-tuned hyperparameters) and Ridge (meta-learner)
• 📅 Temporal Features: DOJ & DOI-based time deltas, week/day flags
• 📊 Feature Engineering:
  • Route frequency statistics
  • Search-to-booking ratios
  • Holiday influence (via custom calendar CSV)
  • Long weekend detection
• 🔍 External Data: Manually curated holiday calendar (India-wide)
• 📦 Tools: Pandas, NumPy, Scikit-learn, Seaborn, Matplotlib

Both models were designed with strict 15-day-ahead forecasting logic, avoiding future leakage and optimizing for RMSE.

RMSE (Root Mean Squared Error)
Used to evaluate model predictions against actual seats booked.

• Top 25 Rank (Rank 21) among 694 teams.
• Uploaded on GitHub to share reproducible research.

• redBus and Analytics Vidhya for hosting the challenge.
• Public sources for Indian holiday data (e.g., digitalpathsala.com, official state calendars).
• Community on GitHub and AV for discussions and guidance.

If you find this work helpful or have any questions, feel free to reach out:

• 📧 [email protected]
• 🌐 LinkedIn