Multi-Target LPG Price Prediction & Market Intelligence System

Context

LPG prices experience severe volatility driven by complex macroeconomic factors, creating information asymmetry for consumers, distributors, and policymakers who lack real-time intelligence for planning and regulation.

Challenge

Simultaneously predicting retail prices at three granularity levels (per kg, 5kg, and 12.5kg cylinders) across 37 geographical regions while managing highly correlated macroeconomic variables.

Key Questions

• How do macroeconomic factors like Dollar exchange rates and PMS prices impact local cooking gas prices?
• What is the true effect of regional conflicts and festive seasons on retail energy pricing?
• Can geographical features be retained without overfitting or arbitrarily pruning less-populated areas?

Approach

• Integrated monthly state-level price records with macroeconomic indicators and binary flags for conflict and festive periods.
• Engineered standardized target variables and unified FX indicators to eliminate multicollinearity.
• Conducted a rigorous Geographical Feature Ablation Study to empirically determine the optimal spatial resolution.
• Trained and evaluated Multi-Target models (Random Forest, XGBoost) using 5-Fold Cross-Validation.

Outcomes

• Identified Dollar exchange rate (~67% importance) and PMS price as the primary drivers of LPG pricing.
• Confirmed that festive price surges are driven by local retail demand rather than international crude oil supply factors.
• Deployed an interactive Streamlit application allowing users to obtain live, localized price estimates.

Takeaways

• Non-linear relationships dominate energy pricing dynamics, requiring complex models over standard linear approaches.
• Geographical context is critical; retaining state and region features improved model accuracy significantly.
• A principled feature ablation study is safer than arbitrary feature importance pruning for preserving geographical identity.