How the model performed during China's 2023 gallium export controls
A retrospective analysis of how our ML models detected early signals of the July 2023 gallium export restrictions and the cascade effect on related minerals.
Background
This analysis draws on the same news-event pipeline and machine-learning models that generate the daily predictions. It is educational context on how the models behave — the specific figures customers act on come from the live feed and the backtests, not from this article.
Understanding the mechanics behind the predictions helps subscribers decide how to weight and integrate the signals into their existing workflows. This is especially relevant for Nickel, where signal density and model behaviour differ from more widely covered minerals like copper or gold.
Methodology
The models process hundreds of features per mineral per day across rolling windows of 3, 7, 14, and 30 days. These features fall into several broad categories: news intelligence, severity analytics, supply-chain risk, geographic analysis, positioning and inventory, and policy/geopolitical context. Each contributes differently depending on the mineral and the current information environment. The models use no price data as an input.
For precious and battery metals, supply-chain features tend to dominate when the model shifts to ELEVATED or above. This contrasts with base metals like copper and aluminium, where news-volume features are typically the primary driver — reflecting the more concentrated supply chains of precious and battery metals.
Key patterns
1. Signal lead time matters. The models are built to flag elevated risk before a volatility spike, not after — forward-looking context is more useful for risk management than reacting once a move is already under way.
2. Confidence correlates with reliability. Predictions where multiple signal categories are densely populated are more reliable than low-confidence ones. The confidence score is published alongside every prediction so subscribers can calibrate how much weight to give it.
3. Cross-mineral contagion is measurable. When one mineral in a related group (for example LME base metals: nickel, copper, zinc) triggers HIGH, adjacent minerals become more likely to trigger within a few trading days. The contagion feature captures this and can give earlier warning for the second and third minerals in a chain.
Practical applications
For subscribers integrating this data into automated workflows, the structured format (CSV + JSON) is designed for direct ingestion. The alert_flag boolean can drive automated notifications, the risk_level string maps to threshold-based logic, and the three risk_factor fields provide human-readable explanations for audit trails.
Subscribers commonly review the 14-day horizon for nearer-term context and the 30-day horizon for longer-term risk monitoring. The two horizons together often surface a developing trend earlier than either alone. How you weight any horizon is entirely your own decision.