Energy / Current / Jun 2026 - Current

Natural Gas Forecasting & Risk Research

Walk-forward study of weekly Henry Hub returns using EIA prices and storage, NOAA weather demand, seasonality, momentum, and volatility to test whether forecasting or risk control adds more value.

Role
Independent Quant Researcher / End-to-end owner
Scope
Data pipeline, feature design, walk-forward models, risk evaluation
Coverage
1997-2026 / daily public data modeled weekly
1,260
walk-forward forecasts
50.3%
final direction accuracy
0.42
final Sharpe (0% RF)
22 pts
lower max drawdown

Research decision

Prediction was noisy. Risk control carried the result.

1997-2026 / weekly
Research inputsPublic data pipeline
Market
EIA Henry HubDaily cash prices
Weather
NOAA CONUSHeating and cooling demand
Fundamentals
EIA storageLevel, change, and seasonal surplus
15 lagged inputsWeather, returns, volatility, seasonality, and storage
Model evidenceWhat survived validation
ARIMA
48.9% directionBaseline, outlier-sensitive
Weather ARIMAX
48.5% directionRaw lagged weather was weak
Linear + storage
50.3% directionFinal risk-adjusted leader
Storage ARIMAX
50.0% directionPositive after adding market tightness
  1. 01
    Walk forward260-week initial window
  2. 02
    Control outliers1st / 99th percentile caps
  3. 03
    Normalize riskCommon 25% volatility target
Final risk-adjusted leader

Linear Regression + 25% volatility target

Annual return
7.7%
Sharpe (0% RF)
0.42
Max drawdown
-53.5%

The question

Natural gas prices react to weather expectations, storage tightness, volatility, and extreme cash-market dislocations. The project asks whether those public signals can forecast weekly direction reliably, and whether disciplined exposure sizing is more useful than a more complex model.

Pythonpandasscikit-learnstatsmodelsARIMAARIMAXGARCHEIANOAA