Model

U.S. Gasoline Inventory Signal

Predicts whether gas prices will be higher next week by tracking U.S. weekly gasoline inventory levels from the EIA.

Accuracy75%

AlgorithmGradientBoostingClassifier · 100 trees · depth 3 · lr 0.08

DataEIA — Weekly Petroleum Status Report (WGTSTUS1: U.S. Ending Stocks of Total Gasoline)

Decision surface

The model's buy probability across its two most influential features. Every point was computed by calling model.predict_proba on the actual .joblib. Drag to rotate.

What the model pays attention to

stocks_4wk_change48%

gasoline_stocks_mb29%

stocks_vs_5yr_pct23%

Approach

I pulled 4 years of weekly U.S. gasoline stock data (EIA series WGTSTUS1) and engineered three features: the current inventory level in million barrels, the 4-week directional change in stocks, and the percent deviation from the 5-year seasonal average. The key insight is that when inventories fall below their seasonal norm and are actively drawing down, refineries are producing less than the market is consuming — a classic supply squeeze that pushes retail prices up within 1-2 weeks. I trained a GradientBoostingClassifier with standard scaling and validated on a 25% holdout set. The model is most confident when inventory signals are extreme (well above or below seasonal averages) and appropriately uncertain when stocks are near average, which matches real-world behavior.

Features

gasoline_stocks_mbfloat

U.S. weekly ending stocks of total gasoline (million barrels, converted from thousand barrels by dividing by 1000)

stocks_4wk_changederived

Change in gasoline stocks over the past 4 weeks (million barrels). Negative = inventory is drawing down (bullish for prices).

stocks_vs_5yr_pctderived

Percent deviation of current stocks from the 5-year seasonal average for that week of the year. Negative = below average = tighter supply = upward price pressure.

Built by

Hrijoo Bosu MullickLinkedIn

Graduate student in Business Analytics at ASU W. P. Carey interested in energy economics and commodity markets. Focused on building interpretable ML models that translate macroeconomic supply signals into actionable consumer decisions.

Trained on

2022-01-01 to 2025-12-31

All models