Back to Portfolio
2026
LightGBMPythonClickHouseForecasting

Festival Demand Prediction Model

"A machine-learning model that forecasts seat demand day by day for festival peaks like Diwali and Dussehra — an ensemble of booking-curve and calendar models trained on three years of bookings and a hand-built festival calendar. Within ~7% on Dussehra 2026."

Executive Summary

Festival weeks decide a bus company's year, so I wanted demand known weeks in advance instead of guessed. I built a forecasting pipeline that pulls daily booking curves from the data warehouse, adds a festival calendar for every major Indian festival since 2023, and blends two gradient-boosted models — one learning how bookings fill up as the date approaches, one learning seasonal and festival effects. Backtested across three Diwalis and Dussehra, it now gives the team seat and traveller forecasts per journey date to plan inventory, campaigns and budgets.

Key Outcomes

  • ~7% error on Dussehra 2026; ~14% across three years of festival backtests.
  • Day-by-day seat and traveller forecasts for every festival window.
  • Stable ratios discovered (seats → travellers → bookers) that planning now uses.
  • Proved batch prediction works on our data — the base for propensity models next.

Next Case Study

Retention CRM Platform

Our own CRM server on Kubernetes — WhatsApp agents, a ticketing desk, a cross-brand DND/opt-out register synced to every send, and live admin dashboards — so retention runs on infrastructure we control instead of vendor limits.