The honest answer about ML in commodity forecasting is unfashionable: it wins on some problems, loses on the ones people most want it to win, and the difference is knowable in advance.
Where ML genuinely creates edge
- Nowcasting and signal extraction. Learning from high-dimensional inputs - satellite storage estimates, freight rates, weather ensembles, order-flow microstructure - is where gradient boosting and modern ML outperform classical models. When the signal exists but is nonlinear and high-dimensional, ML finds it.
- Anomaly and regime detection. Unsupervised methods flagging abnormal basis behaviour, storage builds or curve kinks are production-ready and genuinely additive to a risk workflow.
- Volatility and correlation dynamics. Forecasting the covariance structure that drives portfolio risk - rather than point prices - is a problem ML handles well, and one where even small improvements compound through VaR and ES.
Where it fails, predictably
Multi-week directional price prediction is the graveyard. Commodity prices are close to martingale at most horizons; the features that would beat the market are the ones everyone is bidding on, and the backtest that looks spectacular is usually leakage - look-ahead bias, survivorship, or training on the regime you are testing in. The failure is not the algorithm; it is the question. A model that cannot beat a random walk out-of-sample should not be traded on, however good its training metrics - and no amount of architecture tuning changes a market's signal-to-noise ratio.
Governance is the difference between tool and liability
In a risk-sensitive book, an ML model is a model - subject to the same discipline as a VaR engine: documented features and lineage (training data reproducible via time-travel queries), champion-challenger evaluation against classical baselines, out-of-time validation protocols agreed before results are seen, drift monitoring in production, and a named owner who can explain a prediction to a regulator. The 2026 interagency model-risk guidance reads as if written for exactly this moment; firms deploying ML for forecasting should treat it as the checklist it is.
Our practice, distilled: deploy ML where it demonstrably wins - signals, anomalies, volatility and correlation dynamics - and resist the directional-forecast temptation. The edge is real, but it lives in the unglamorous corners.