Traffic demand forecasting is often treated as a "deep learning first" problem. However, for the Flipkart GridLock 2.0 Hackathon, I took a different approach: a hierarchical basis-day routing matrix combined with ridge calibration.

The result? A 95.007% R² score that is both computationally efficient and fully interpretable.

The Challenge

The competition (organized by Flipkart and the Bengaluru Traffic Police) required predicting demand on a specific hidden horizon (Day 49). The data included spatial geohashes, temporal timestamps, and road-level metadata.

The Architecture

Instead of a black-box neural network, the pipeline uses four logical stages:

1. 
   Proxy Construction: Using non-overlapping Day 49 labels to learn spatial structures.


2. 
   Hierarchical Routing: Fitting simplex-constrained weights across global, geo_4, and geo_5 granularities.


3. 
   Ridge Calibration: A linear model trained on the observed early-morning Day 49 slice to adjust for real-time trend shifts.


4. 
   Residual Correction: A bounded additive geohash-level fix based on historical weekly patterns.

Why This Matters

By avoiding deep learning, the model runs in seconds on standard hardware and provides clear insights into why a specific demand prediction was made—crucial for urban mobility decision support.

I've open-sourced the full pipeline, including the optimized .xz data handling and EDA companion notebooks.

View the full project on GitHub: https://github.com/SNiPERxDD/GridLock2.0

pip install -r requirements.txt

python gridlock_release_pipeline.py

GridLock 2.0/

├── README.md

├── requirements.txt

├── gridlock_release_pipeline.py

├── notebooks/

│   ├── gridlock_release_pipeline.ipynb

│   └── gridlock_eda_companion.ipynb

├── data/

│   ├── competition_train.csv

│   ├── competition_test.csv

│   ├── evaluation_ground_truth.csv

Tags: #bengaluru #urbanmobility #flipkart #gridlock