Why nobody can forecast the exact steel price, and what to track instead
Every few months a new model claims better accuracy on steel or copper prices. None of them fix the actual problem, because a single predicted price was never the right target. Here's the research behind why, and what to track instead.
The error-profit paradox
A 2026 study on AI-based copper price prediction, published in Machine Learning and Knowledge Extraction, tested statistical and AI models against each other under different market regimes and found what the authors call an error-profit paradox: models with the highest statistical forecasting accuracy did not necessarily produce the best trading outcomes. In several cases, a model with larger prediction errors performed better in practice, because it captured the direction of the move more reliably than the model that scored better on paper. (Bareith et al., 2026)
That's a direct challenge to how forecasting is usually judged. A model can post a lower average error and still be the wrong model to act on, because the metric being optimized (how close the number is on average) isn't the same thing a buyer actually needs (whether a specific move happens or not).
Why point forecasts break down for real buyers
A separate, more practitioner-focused critique from MetalMiner makes a related point from the buying side. Generic exchange benchmarks like LME or CME often diverge from what a specific buyer is actually exposed to: your material might be coated coil, a specific extrusion, or a particular stainless grade, not generic “steel” or “aluminum,” and a global index can move differently than your regional physical market.
The piece also points at a subtler failure: a model can be directionally right most of the time and still be useless if it can't answer the question a buyer actually has, something closer to whether to extend coverage six months or hold 50% spot for now, not a single number for next quarter.
What to track instead of a single number
Both critiques point the same way: away from “what will the price be” and toward “how likely is a move large enough to matter.” That reframing turns an unfalsifiable point estimate into a claim you can check. If a model says a 30% chance of a move, that should happen close to 30% of the time across many such calls, a property called calibration. And because commodity markets pass through structural breaks that alter which model works best, a probability's track record needs to be validated walk-forward, tested only against data the model never trained on, not just backtested on the same history it learned from.
Where MetalAlert fits
MetalAlert is built around that reframing. For each metal it tracks, it estimates the probability that the average price over the coming month or quarter runs at least 2%, 5%, or 10% above or below where it sits today, rather than producing a single predicted price. Each model's track record is shown per metal, scored for calibration and walk-forward performance, so the number behind an alert is something you can check against what actually happened, not something you have to take purely on trust.
FAQ
Why can't steel and other metal prices be forecast precisely?
Structural breaks, geopolitical shocks, and mismatches between generic exchange indices and a buyer's actual product and region all limit precision. Research on copper price prediction has found that models with the best statistical accuracy don't necessarily produce the best real decisions, and the reverse also holds: a model with larger errors can still be more useful if it captures direction and timing well.
Are AI or machine learning models better at forecasting metal prices?
They can reduce error on historical data, but studies comparing forecasting methods on copper have found model rankings shift across time periods, horizons, and volatility regimes. No single model type has held a consistent statistical edge, which is itself evidence that the exact number is the wrong target.
What is the error-profit paradox?
A finding from research on AI-based copper price prediction: models with the highest statistical forecasting accuracy do not necessarily generate the best trading or decision outcomes. In some cases a model with larger prediction errors performed better in practice because it captured the direction of the move more reliably.
If not a price forecast, what should a metal buyer actually track?
The probability that the price moves past a threshold that matters to your business, over a horizon that matches your buying cycle, rather than a single predicted number. That reframes the question from "what will the price be" to "how likely is a move big enough to change what I should do," which is the question a purchasing decision actually depends on.
How does MetalAlert apply this?
For each metal it tracks, MetalAlert estimates the probability that the average price over the coming month or quarter moves at least 2%, 5%, or 10% in either direction, calibrated so a "30%" call happens close to 30% of the time, and validated walk-forward against outcomes the models never trained on.
See the numbers behind this for the metals most buyers ask about: the copper price forecast and 2027 outlook and the aluminium (aluminum) price forecast, each with the LME monthly average and the odds of a 2%, 5% or 10% move.
See the probability, not a point guess
The same classification-over-regression idea, proven out by a real buying decision, is covered in how a steel wholesaler's predictive model paid back 13% of its cost.