The forecast object becomes the product

Industrial metals forecasting is quietly being rebuilt for a reader that does not read. When an AI agent retrieves a copper or aluminum price via an MCP (Model Context Protocol) server and routes it straight into a procurement workflow, the headline number is the least interesting part of the payload. What matters is whether the benchmark is unambiguously specified, whether the uncertainty is calibrated, and whether the metadata travels with the number. The forecast object, not the forecast, is becoming the product.

The argument from MetalMiner is bluntly stated: most forecasting errors do not originate in the model. They originate upstream, in sloppy benchmark definitions. A number labelled simply 'aluminum' can refer to an LME (London Metal Exchange) three-month contract, a US Midwest premium, a specific product form, or a regional transaction price — each with distinct volatility, trend persistence, and drawdown

That specification reads like plumbing, but it is the whole point. Each of those fields is a resolution criterion — the kind of unambiguous anchor a prediction market would require before writing a contract. Without them, a forecast is a sentence; with them, it is a structured object that a downstream system can verify, version, and challenge. MetalMiner's position is that machine-readable intelligence collapses the moment any of those fields goes missing, because the consuming agent has no way to detect the loss.

Why calibration is the harder half

Structured metadata handles identity. Calibration handles trust. Here the most interesting external evidence comes from a large-scale evaluation of prediction markets flagged by Marginal Revolution, which examined roughly one million resolved contracts across Polymarket, Kalshi, Manifold, and Metaculus. Events assigned a probability of p percent occurred approximately p percent of the time. Aggregated human judgment, in other words, produces well-calibrated probability distributions at scale — a standard against which proprietary metals forecasts can now be measured, rather than taken on faith.

“Large-scale aggregation of judgments contains meaningful information about real-world uncertainty; when examined in aggregate, the predictions generated by the platforms provided informative estimates of event likelihood.”

— Marginal Revolution

The implication for metals vendors is uncomfortable. If a forecast carries a 70 percent confidence interval, that interval should contain the realised price 70 percent of the time across a long run of calls. Most published industrial price forecasts have never been audited this way. In an MCP world where agents will query the same forecast endpoint thousands of times, calibration becomes empirically testable almost immediately — and reputations will adjust accordingly. US hot-rolled coil, LME copper, LME aluminum, 304 stainless, cobalt, and lithium carbonate each demand their own calibration regime; a single confidence model stapled across the complex will fail visibly.

The dossier offers no quantified forecasts on adoption timing or vendor concentration, so we decline to invent any. The structural claim stands on its own: benchmark fidelity, structured metadata, and calibrated uncertainty are becoming hard requirements for any price feed destined for an autonomous workflow. Vendors who treat forecasts as prose will be filtered out by the agents that cannot read them; vendors who treat forecasts as objects will become the default upstream dependency. The transition is already underway, and it rewards the unglamorous work of definitional discipline over narrative flair.