Independent project · July 2026 – Present
AI Commodities Research & Forecasting System
A multi-agent research system that turns scheduled EIA releases into scenario forecasts and point-in-time backtests of hypothetical positions in energy futures, ETFs and equities, after transaction costs and risk controls.
- Data
Collect the release and related data, preserving when each item became available.
- Analysis
Compare the release with market expectations and explain the potential price impact.
- Forecasts
Develop scenario forecasts with explicit assumptions.
- Backtesting
Build hypothetical positions and test them against historical prices after transaction costs.
- 35
- Historical EIA releases tested
- 15
- Instrument baskets tested
- US$100,000
- Simulated capital per portfolio
All results are simulated. The project does not manage capital or execute trades.
Research question
Scheduled energy-data releases can move prices within hours. This project asks whether a structured research process can translate those releases into useful forecasts and hypothetical positions after transaction costs. I use the petroleum inventory reports published by the US Energy Information Administration (EIA) as the test case.
My role
I designed the research questions, selected the data and set the workflow, controls and evaluation rules. I used AI coding tools to build the system and reviewed the research outputs myself.
Inputs
The system combines market, macroeconomic and inventory data to test event-driven scenarios across energy futures, exchange-traded funds and equities.
Approach
Four controls govern the backtest:
- Point-in-time data. Each decision uses only information that would have been available when it was made.
- Calculations in code. A deterministic Python engine performs the calculations. The AI agents interpret the information and draft the research.
- Costs and risk. The backtest applies transaction costs, position sizing and risk controls to each US$100,000 paper portfolio.
- Human review. I review the scope, uncertain sources and key assumptions before they enter the analysis.
Output
The system produces point-in-time backtests and research reports, with the inputs, assumptions and calculations recorded alongside each result. That record makes the analysis traceable and reproducible.
Limitations
This remains an experimental research project. All results come from historical backtests in simulated portfolios. The system is not connected to a broker, does not execute trades and has not managed capital.