FuelCast

• 

• 

⛽ FuelCast: The AI Energy Price Oracle

Project Name: FuelCast: AI Energy Price Oracle

Summary: FuelCast is a full-stack, predictive analytics dashboard that forecasts U.S. retail gasoline prices using a non-linear machine learning model, designed to mitigate the economic shock of energy price volatility. Our model achieved a 45% reduction in prediction error compared to the industry standard, providing a viable tool for consumer and business budgeting.

🌍 UN Sustainable Development Goal Alignment

Primary Focus: SDG 7: Affordable and Clean Energy

Target Addressed: Target 7.1 (Ensure universal access to affordable, reliable, and modern energy services).

Impact: Volatile fuel prices are a major threat to economic stability and affordability for small businesses and low-income households. Our high-accuracy model (MAE of 8.5 cents) enables proactive budgeting and hedging against price spikes, making necessary transportation energy more economically reliable and accessible.

💡 Inspiration

Traditional time-series forecasting (SARIMAX) fails when applied to volatile commodity markets because it assumes linearity. Our inspiration was to solve this "Linear Trap". We knew that retail gas price is a downstream effect of the global crude oil price, but with a complex time lag and non-linear market reaction. We set out to build a system that could identify and exploit this hidden causal relationship.

🛠️ What it does

FuelCast is a real-time predictive dashboard that delivers the most accurate retail fuel price forecast possible using historical data.

Prediction: Forecasts U.S. Regular Gasoline prices for the near-term future.

Visualization: Displays predictions from two models (SARIMAX and XGBoost) against the actual price curve on a Next.js dashboard using Recharts.

Interpretation: Provides clear Feature Importance scores from the XGBoost model, explaining why the price is moving (e.g., confirming the influence of crude oil).

🔨 How we built it

Data Ingestion: We sourced weekly U.S. Retail Gas Prices (Target) and Daily WTI Crude Oil Futures (Exogenous Feature) from high-authority datasets (EIA-sourced data).

Feature Engineering: We transformed the time series by creating powerful, lagged features, most critically the 4-Week Lag for the Crude Oil price and features for Seasonality (month, day-of-year).

Modeling Pipeline: We trained two models for comparison:

Baseline: SARIMAX, a linear statistical model.

Challenger: XGBoost Regressor, a non-linear ensemble model, trained on our engineered features.

Full-Stack Deployment: We saved the trained models and designed the front-end (Next.js/TypeScript/Tailwind) to fetch data from a planned FastAPI Python service.

🚧 Challenges we ran into

The Negative $R^2$: Our primary challenge was proving the "Linear Trap." The SARIMAX baseline model's R² score was -1.4241, which meant it was performing worse than a random guess (the mean). This required an immediate and absolute pivot to the machine learning approach.

Data Alignment: Successfully merging the daily WTI data with the weekly retail gas price data required careful resampling and ensuring the time indices aligned correctly for prediction.

SARIMAX Instability: The auto_arima function struggled with model convergence, highlighting the fragility of statistical methods when facing noisy, volatile data.

✅ Accomplishments that we're proud of

Error Reduction: Achieving a 45% reduction in prediction error by reducing the RMSE from $0.2264$ (SARIMAX) to $0.1238$ (XGBoost).

MAE Score: Delivering an average absolute prediction error (MAE) of just 8.5 cents per gallon. This is a business-ready metric for hedging and planning.

Causal Validation: Using the XGBoost Feature Importance to empirically prove that the 4-Week Crude Oil Lag is the dominant causal factor in retail price movement.

🧠 What we learned

We learned that in finance and commodity prediction, the modeling choice is secondary to Feature Engineering. By converting the time-series problem into a supervised learning problem and feeding the XGBoost model specific, lagged causal features, we generated superior performance compared to relying on the model's inherent temporal capabilities alone.

🚀 What's next for FuelCast

EIA Integration: Replace historical Kaggle data loading with direct, live integration to the EIA Public API for weekly automated updates.

Cloud Deployment: Deploy the trained XGBoost model via the FastAPI service to a scalable cloud platform (e.g., Google Cloud Run).

Monetization & SDG Expansion: Expand the dashboard into an API service for regional logistics companies to optimize their supply chain costs based on the forecast, directly enhancing affordability for the transport sector.

Built With

• fastapi
• lucide
• next.js
• numpy
• pandas
• rechart
• tailwind
• typescript
• uvicorn