Hierarchical Forecasting of District-Level Petrol and Diesel Demand in India: Public Covariates, Benchmark Discipline and a Refiner Supply Layer

Refiners and fuel marketers plan crude runs, product placement and depot stocks on cumulative demand over the coming one to twelve months, at district, state and national level, and by marketing company. This paper describes a demand forecasting system built to support that planning, for petrol (motor spirit) and diesel (high-speed diesel) across Indian states and union territories at district, state and national level, extended to company-wise demand and to a refiner’s sales at its supply locations. The system is built on four ideas. First, every forecast is made only with information actually released by the run month: the publication lag of each source is replayed in back-testing, and a truncation test, shown to detect a planted leak, verifies that the feature store contains no look-ahead. Second, a causal chain separates market demand from company shares and from the refiner’s effective share of its customers’ requirement, so supply-side decisions never contaminate demand estimates. Third, candidate models (seasonal naive, seasonal naive with drift, exponential smoothing and a global gradient-boosted model) are reconciled with MinT-shrink using each model’s own error covariance and selected under pre-declared kill criteria. Fourth, the covariate set is built from public sources: satellite rainfall aggregated to states with GIS, national toll-plaza traffic, vehicle registrations, goods and services tax collections, industrial production, a curated event and agricultural calendar, and state fuel levies. The sample deliberately starts in April 2022, after the COVID-19 disruption to fuel demand. On a fourteen-month hold-out the reconciled seasonal naive with drift forecasts three-month cumulative demand at the aggregate level with a weighted absolute percentage error of 1.8% (diesel) and 2.0% (petrol), and at state level 3.3% and 2.8%. The gradient-boosted model and most macroeconomic driver groups failed to earn their place out of sample. These negative results and the evaluation protocol are reported in full.