Climate Commodity Alpha Lab

A quantitative research project for commodity return forecasting using price dynamics, weather anomalies, and extreme climate-risk signals.

Research Question

Can weather and climate-risk variables improve commodity return forecasts under a robust walk-forward validation framework?

Motivation

Commodity markets are physically constrained: energy supply disrupts when hurricanes reach the Gulf of Mexico, grain yields fall under drought, and demand for natural gas spikes with temperature anomalies. This project transforms that physical domain knowledge into systematic, testable predictive signals using rigorous time-series validation.

Results (summary)

Walk-forward validation (500-day train / 21-day test / 21-day step) over 7 commodity ETFs, 2007–2026, net of 5 bps one-way transaction costs.

• Climate + price features beat price-only baselines on 5 of 7 assets by Information Coefficient (IC).
• Best assets: SOYB (IC 0.121) and GLD (IC 0.113, net Sharpe 1.38).
• Climate features add +0.05 to +0.07 IC on agricultural and gold tickers; the Gulf hurricane-disruption proxy ranks top-5 by XGBoost gain importance for energy.
• Equal-weight portfolio: net Sharpe ≈ 0.82 after costs.

Scope & honesty: this is a research study on whether climate / alternative data adds predictive signal (measured by IC), not a deployable trading strategy. Drawdowns are large and some assets (e.g., UNG) are not forecastable with this feature set and are excluded from portfolio views. All results use strictly temporal validation (no look-ahead) and are reported net of costs. Full details in reports/quant_research_note.md and reports/backtest_report.md.

Architecture

flowchart LR
    A[Commodity Prices] --> D[Aligned Research Panel]
    B[Weather Variables] --> D
    C[Hurricane / Extreme Event Data] --> D

    D --> E[Feature Engineering]
    E --> F[Return Targets]
    F --> G[Walk-Forward Modeling]

    G --> H[Forecast Scores]
    H --> I[Signal Ranking]
    I --> J[Portfolio Construction]

    J --> K[Backtesting]
    K --> L[Risk Metrics]
    K --> M[Research Report]

    L --> N[Sharpe / Drawdown / Turnover]
    M --> O[Quant Research Note]

Loading

Assets

Proxy	Commodity	Climate Relationship
UNG	Natural Gas	Temperature anomalies, hurricane disruptions
USO	Crude Oil	Gulf hurricane risk, supply shocks
DBA	Agriculture broad	Temperature, precipitation
CORN	Corn	Precipitation, heat stress
WEAT	Wheat	Drought, temperature extremes
SOYB	Soybeans	Precipitation, growing season climate
GLD	Gold	Macro/risk-off control

Data Sources

Dataset	Source	Purpose
ETF prices	Yahoo Finance (yfinance)	Commodity price proxies
Weather data	NOAA GHCND	Temperature, precipitation
Hurricane tracks	IBTrACS v4	Tropical cyclone risk features
Macro variables	FRED / public sources	Optional regime context

Methodology

1. Price feature engineering — returns, momentum, volatility, drawdown, skewness
2. Weather anomaly construction — day-of-year climatology, rolling anomalies, drought proxies
3. Hurricane risk scoring — distance-weighted wind intensity for key infrastructure regions
4. Forward return targets — 1d, 5d, 10d, 20d horizons; directional and regression targets
5. Walk-forward model training — strictly temporal, no shuffling, no random splits
6. Signal generation — cross-sectional ranking and time-series forecasts
7. Portfolio backtesting — with transaction costs and turnover constraints
8. Risk and robustness analysis — regime breakdown, stress tests, drawdown analysis

Models

Model	Role
Ridge Regression	Baseline, regularized
Logistic Regression	Directional baseline
Random Forest	Non-linear reference
XGBoost / LightGBM	Primary forecasting model
Bayesian Regression (PyMC)	Uncertainty quantification
Volatility / Climate Regime Detection	Conditional performance analysis

Evaluation Framework

Metric	Purpose
Information Coefficient (IC)	Predictive quality of scores
Hit Rate	Directional accuracy
Sharpe Ratio (net)	Risk-adjusted performance
Max Drawdown	Downside risk
Turnover	Cost efficiency
Performance by regime	Robustness under different conditions

Repository Structure

climate-commodity-alpha-lab/
├── src/
│   ├── data/           # Data loading and panel construction
│   ├── features/       # Price, weather, hurricane and regime features
│   ├── models/         # Baseline, tree, Bayesian models + walk-forward
│   ├── backtesting/    # Signal-to-position, portfolio, performance
│   ├── risk/           # Drawdown, VaR/CVaR, stress tests
│   └── utils/          # Metrics, date utilities, leakage validation
├── notebooks/          # Ordered research notebooks (01–08)
├── reports/            # Quant research note, backtest report, figures
├── tests/              # Unit tests for all modules
└── data/               # Raw, interim and processed data (not tracked)

Notebooks

Notebook	Content
01_price_data_eda.ipynb	Price data, returns, feature distributions
02_weather_climate_features.ipynb	Weather anomalies, drought proxies
03_hurricane_risk_features.ipynb	Storm tracks, risk scoring, energy exposure
04_signal_engineering.ipynb	Feature selection, cross-asset correlations
05_walk_forward_modeling.ipynb	Model comparison under temporal validation
06_backtesting_strategy.ipynb	Signal-to-portfolio, equity curve, metrics
07_regime_detection.ipynb	Volatility and climate regime analysis
08_research_report.ipynb	Full findings as a quant research note

Notebooks 01 and 03–08 run end to end on the committed processed data (Python 3.10). Notebook 02 (ERA5 weather features) is optional: it requires downloading ERA5 data via the Copernicus CDS API (and the cfgrib/eccodes stack). The pipeline runs without it — ERA5 features are skipped automatically when the raw files are not present.

Why this project matters

Physical climate-risk modeling and systematic commodity research share the same core challenge: transforming noisy, non-stationary time-series data into robust predictive signals. This project applies hydrometeorological domain knowledge — tropical cyclone tracks, precipitation, temperature — directly to commodity return forecasting, with the same rigor expected in production quant systems: temporal validation, realistic transaction costs, and structured reporting.

Setup

git clone https://github.com/GaiskaSalomon/climate-commodity-alpha-lab.git
cd climate-commodity-alpha-lab
pip install -r requirements.txt
pip install -e .
cp .env.example .env  # fill in API keys
pytest tests/

CV Summary

Built a systematic commodity forecasting pipeline using price dynamics, weather anomalies and tropical cyclone risk features. Implemented walk-forward validation, XGBoost/LightGBM models, signal generation, transaction-cost-aware backtesting and risk reporting. Designed to test whether climate-risk alternative data improves commodity return forecasts.