Petro4Cast

The Challenge

U.S. fuel demand is shaped by economic, market and operational factors. Although traditional models predict demand, none explain why it changes.

Correlation is Not Causation.
Crude price, industrial output and demand move together; regression weights do not say which lever actually moves demand.

Structural Breaks Break Models.
A mobility collapse, a refinery closure or a strategic reserve release changes the data-generating process, not just the noise.

Conditional on Sampled History.
A forecast conditioned on history cannot evaluate an intervention that has never happened in the sample.

Our Solution

Petro4Cast is a framework that forecasts demand, identifies its real drivers, explains the what, when, how, & why a change happened, and shows how the market would have evolved under different conditions.

Each layer is one of the challenge questions that Petro4Cast answers.

Where traditional forecasting stops

✕Predicts demand, but cannot explain the move

✕Treats every correlated series as a cause

✕One accuracy number, no driver-level attribution

✕Breaks silently across structural shifts

✕No answer for what would have happened otherwise

What Petro4Cast adds

✓Splits every forecast into seasonality, trend and driver contributions

✓Keeps a driver only if it survives the causal graph and refutation tests

✓Scores every forecast against univariate benchmarks across windows

✓Dates structural breaks with weekly event flags and avoids edge fitting

✓Answers what-if questions with do(·) interventions, both forward-looking and historical

How Petro4Cast Works

Select a stage to know more.

Public data inEIA weekly dataFRED macro seriesCrude and retail pricesDegree days and event flags

Decisions out3-6M forecastsDriver attribution per fuelScenario branchesReady-to-share reports

Our What-If Engine

The counterfactual engine fixes a driver with a do(·) intervention and propagates the change through the fitted model.

Scenario branches

Illustrative shape only. The live scenario outputs for each fuel are on the Counterfactual page.

Open the counterfactual engine →

Built to be trusted and deployed

Our governance pillars shape every output, and our architecture keeps the core model safe and sound inside the private cloud while internet search remains isolated.

Open & Free Data

Every source is public: EIA, FRED and market data. Every transformation is logged, so the panel behind the model can be audited end to end.

Explainability & Interpretability

Every forecast decomposes into seasonality, driver contributions and trend. Every causal estimate carries its backdoor adjustment set and refutation results.

Understandability

Outputs speak the language of the decision they feed: volumes, days of cover and margin exposure, with a conversational assistant to interrogate them.