The Value of a Steel Wholesaler's Predictive Model

The problem wasn't intuition, it was confidence

The company didn't need a model to understand its market. It is a stable business with years of experience and strong instincts. It already tracked historical prices and had dashboards that showed how orders and prices changed over time.

The gap was elsewhere. When a decision felt risky, the team often held back or acted too late. So the question became: what could give them the confidence to act on decisions they already suspected were right?

The answer was a predictive model built as a decision-support tool, not a decision-maker. If the model agrees with a call that feels risky, the team can act with more confidence. Get just a few of those decisions right each year and the model is profitable.

Asking a better question

Forecasting exact steel prices is hard. The data doesn't suit classic time-series models. Test data is scarce, inflation pushes prices outside the range the model was trained on, and the pandemic years distorted history.

Instead of excluding awkward data to make the models look good, the team changed the question. They asked the business which price moves actually matter. The answer was significant rises and drops over the coming months, not the exact price. So the model predicts one of four outcomes:

• Heavy rise
• Rise
• Drop
• Heavy drop

This made the target easier to validate and the output easier for decision-makers to understand. Trend-based features, careful feature selection and interpretable models kept it from becoming a black box.

Proving it works

The project began with a six-week proof of concept on a small set of products. Regular check-ins with stakeholders kept the work tied to what the business needed.

Stage	Result
First modelling round	3 of 8 products predicted well
After adding two more data sources	6 of 8 products showed promising results
About 4 months on live data	Models still performed, with no sign of overfitting
3 more months	Results remained surprisingly strong

A good score on historical test data proves little on its own. The real test was whether the model kept working as new months came in, and it did.

The first real decision

The first real decision paid back 13% of the total project cost. Management faced a purchase they considered too risky. The model said the timing was right, so they went ahead, and it paid off.

That is where the value comes from. The model doesn't have to be right every time. It only needs to help the business make a few well-timed decisions it would otherwise have avoided.

What it takes to create value

The wholesaler's model shows that predictive analytics pays off when it supports decisions instead of replacing them. Three things made the difference:

• Start from the business decision, not the algorithm. The goal was confidence to act, not a perfect price forecast.
• Frame the problem so people can trust it. Four clear outcomes are easier to validate and explain than a single price point.
• Be honest about uncertainty. Only time on live data proves a forecasting model, so the rollout was built around watching it perform.

Next, the model will scale to around 30 products, with predictions in an interactive dashboard and regular retraining so it keeps up with a changing market.

What this means for a metals desk

The lesson generalises well beyond steel. MetalAlert applies the same idea to metal markets: for each metal it tracks, it estimates the probability that the average price over a future month runs at least 2%, 5% or 10% above or below the current month's average, over one- and three-month horizons. It doesn't replace judgement, it adds confidence to the calls a buyer already suspects are right, backed by a visible, walk-forward track record rather than a slide on trust.