The software company that created FT Mercati Go to ftmercati.com

Case study

Machine learning in support of scrap quotations

FT Rottami: seven ML.NET algorithms read the merchants' contributions and determine each week the changes in the prices recorded at the scrap yards.

In production FT Mercati's weekly scrap price assessment Updated on

  • 7regression algorithms compared on the same data, tested in batches with different configurations and cross-validation
  • 5years of history for training, and sixteen months out of sample for testing
  • 3ways of combining the models: average, error-weighted average, median
  • 3blocks of categories that by rule move together, whatever the model says
  • 1scheduled entry per week, which writes only where a value is missing; the pricing team can always overwrite

In brief

FT Rottami is FT Mercati's machine learning system that reads all the contributions received from the merchants and determines the price changes for each category, as recorded at the yards: seven algorithms trained on five years of history, tested on sixteen months never seen before, with the group constraints enforced by rule. The weekly entry is automatic; the pricing team can review and correct. It is not a market forecast.

The starting point

Every week FT Mercati publishes its quotations for ferrous and non-ferrous scrap, by category. The assessment starts from the merchants’ contributions, collected between Monday and Tuesday, and is compared with linked sources that come out at different intervals: steel industry price lists, Chamber of Commerce price surveys, trade association data. On Wednesdays the pricing team had to decide, category by category, how far to move the quotation, while also respecting the constraints between categories in the same block. It is work for experts, repetitive and with a lot of numbers to look at.

What we built

A system based on machine learning, the branch of artificial intelligence (AI) that learns from data, written in .NET on the history of the assessments:

  • The variables. From the contributors: average, median and spread of the proposed changes, share of rises and falls. From the linked categories: the net change in the external sources. From the calendar: week, month, holiday periods. From the history: trend over the last four weeks and volatility.
  • The constraints. The categories in the same block (sheet scrap, demolition scrap, turnings) must move by the same amount: this is imposed by a rule written in the code, not by the model.
  • The algorithms. Seven regression methods compared on the same data, tested in batches with different configurations and cross-validation, and three ways of combining them into an ensemble: average, average weighted by the inverse of the error, median.
  • The test. Training from October 2019 to July 2024; out-of-sample test from September 2024 to December 2025; error metrics and percentage of correct direction by category; an archive of the experiments with chart comparison and export.
  • Production. A tab with the chosen model, the Wednesday date and the schedule: in production, entry is scheduled and writes the changes by itself where there is no value yet. With “Confirm and Insert” the pricing team can review the proposals, change them and overwrite a value already written.

Where generative AI comes in

In the lab, not in the quotation. The person running the experiments writes in plain language, in Italian in our case, what they want to try, “five conservative FastTrees”, “ten configurations with shallow trees”, and a language model turns it into training configurations ready to run, with self-explanatory names. Only the request goes to the model: the market data does not leave the company’s servers.

How we keep the model under control

  1. Automatic entry, correction always possible. The scheduled entry writes only where there is no value yet and never touches the ones already there; the pricing team can review and overwrite from the Production tab. There is no mandatory sign-off before entry, and we say so.
  2. Out-of-sample test. We do not put into production a model that does not hold up over sixteen months it has never seen, whatever its error in training.
  3. Constraints by rule. The blocks of categories move together because of a written rule, whatever the model says.
  4. Everything on-premise. The models are trained and run on FT Mercati’s servers; no market data leaves the company.
  5. Tracked experiments. Every training run stays in the archive with its parameters and its metrics: we always know why a model was chosen.

Declared limits

  • the model is only as good as the contributions it receives: a week with few contributions produces less reliable values, and the pricing team can correct them;
  • the linked sources come out at different intervals, weekly or fortnightly: a late source shows in the proposals;
  • the lab’s configuration generator depends on a language model named in the code: it has to be updated when the provider retires it, as happened in 2026;
  • it is an aid to FT Mercati’s internal assessment, not a service for forecasting prices, which FT Support does not offer.

What it means for your company

Many companies have an assessment or an estimate that is repeated every week: an internal price list to update, a premium to set, a consumption figure to estimate. Machine learning can determine the number from the history and from the current data; the rule is that the model derives it from the data collected and the person can correct it, with an out-of-sample test beforehand and a log afterwards: an internal assessment, never a market forecast. We discuss it in the guide on how to connect LME prices to your ERP (the company’s management software) and in the one on AI in the cloud or on-premise. If you have a weekly assessment that today runs on experience and spreadsheets, let’s talk.

Frequently asked questions

ML for scrap quotations: questions and answers

Does the model forecast scrap prices?

No. The model reads all the contributions received from the merchants, the linked categories, the recent trend and seasonality, and determines the weekly change for each category as recorded at the yards. It is an internal FT Mercati assessment, not a market indication and not a forecasting service.

Who decides the final quotation?

In production, entry is automatic and scheduled: on Wednesdays the model reads all the contributions received and writes the changes where there is no value yet. The pricing team can always step in from the Production tab: review the proposals, change them and overwrite a value already written, with "Confirm and Insert". There is no sign-off before entry: the pricing team's check comes afterwards, when it is needed.

How can you tell whether a model is good?

With the out-of-sample test: sixteen months never seen during training, mean error and squared error, percentage of correct direction, category by category. The experiments stay in the archive and are compared with one another: this is the yardstick for deciding which model deserves to go into production.

Where does generative AI come in?

In the lab: a request in plain language such as "five conservative FastTrees" becomes, through a language model, a set of training configurations ready to run. Only the request goes to the model: the market data stays on FT Mercati's servers.

Other case studies

Do you have a process you would like to see working better?

Tell us about it in a first conversation, with no obligation: together we work out whether and where AI brings a measurable benefit, and which step to start with.