Need help with problems like this?

Let's talk about your next steps.

Published

November 5, 2026

How to judge a price forecast before you trust it

How to judge a price forecast: compare it with no change, check how it was tested, ask for its misses and a range, and question the inputs.

The quotes for the first quarter are on your desk, and someone already has a view on where prices go next. It might be a slide in a supplier review, a market report, a software demo, or an answer a colleague got from ChatGPT: resin down 8% by spring, so wait. Whether you lock now or hold off may come down to a line someone else drew.

A price forecast earns trust when it beats a simple “no change” forecast

, was tested only on information available at the time, shows its track record with its misses, comes as a range rather than one line, and rests on inputs that make economic sense. Forecast accuracy is a record over many periods, not one good call.

None of these checks needs a statistics background. They are questions you can put to any supplier, analyst or software vendor.

In short:

  • Compare it with “no change”. A forecast that can’t beat today’s price carried forward adds nothing.
  • Ask how it was tested. Results built with hindsight look better than they are.
  • Ask for every forecast and its outcome, not the best calls, and see whether the misses lean one way.
  • Ask for a range. One line hides how unsure the forecaster is.
  • Check the inputs. A forecast inherits the errors in its assumptions.

About to act on a price forecast? We’ll go through it with you against these checks.

What does forecast accuracy mean for a price forecast?

It is how far the forecasts land from what actually happened, over many forecasts, at the horizon you decide on, compared with a simple alternative. One correct call proves little, because every forecaster gets some right. What counts is how large the misses are over time, and in which direction they fall.

Even the best-resourced forecasters miss by a lot when shocks hit. Ben Bernanke’s April 2024 review of forecasting at the Bank of England found that “the accuracy of the Bank’s economic forecasts has deteriorated significantly in the past few years”, and that the shocks behind the errors “were generally not predicted by financial markets or external experts”. For a buyer, that doesn’t make forecasts useless. It means a forecast is only usable when you know how wrong it tends to be.

Providers report accuracy in different ways, often as MAPE (mean absolute percentage error) or RMSE. You don’t need the formulas. Ask for the errors in terms you buy in, euros per tonne or per cent, at the horizon you decide on, such as three or six months, and next to the benchmark in the next section.

Does it beat a forecast of “no change”?

The simplest benchmark is a naive forecast: assume the price in three or six months will be what it is today. It costs nothing, so a forecast you pay for, or act on, should miss by less over the same periods and horizon. Without that comparison, you can’t tell whether the forecast adds anything.

“No change” is the standard benchmark in price forecasting, so ask for that comparison. And check that the forecast is scored on the price your contract uses: a monthly average, a month-end quote, a published index or a delivered price.

A worked example, with hypothetical numbers: a forecast misses by 6% on average at three months, which sounds good. If assuming no change missed by 5% over the same months, the forecast did worse than doing nothing. If no change missed by 10%, the forecast cut the error almost in half. The 6% means little until you see the second number.

Was it tested only on what was known at the time?

A fair test, often called a backtest, runs the forecast as it would have been made on each past date, using only the data available then, and scores it against what followed. If later information leaked in, such as revised data or a model that had already seen the outcome, the test flatters the forecast.

Leakage is easy to miss. A model tuned on the whole history and then tested on part of it has already seen the answers. Input data revised after the fact is not what a forecaster had at the time. And picking the best of many model versions after seeing the results is a form of hindsight too. Ask when the model was fixed, and how long it has run live since. Forecasts made and recorded before the outcomes were known are the strongest evidence you can get.

Can you see the track record, misses included?

Ask for every forecast made over a long period, with the outcome next to each, not a few good examples. Look at how large the misses are, how they behave in volatile months, and whether they lean one way. Misses that keep pointing in the same direction are forecast bias, which you can spot and correct for.

The European Central Bank puts it simply in its July 2024 review of its own inflation projections: “A bias indicates whether there is a systematic and persistent over or underprediction of the target variable.” The review found the projections unbiased overall, but not in every period. A track record that covers only calm months tells you little about forecast accuracy in the months that matter.

The Bernanke review recommended that Bank staff be “charged with highlighting significant forecast errors and their sources”. A provider worth relying on does the same without being asked.

Ask forWhyRed flag
Every forecast over a year or more, with outcomesShows the typical miss, not the best oneOnly case studies, or "we called the last spike"
Errors at your decision horizonA one-month forecast can be good while six months is notOne accuracy figure with no horizon
The same errors for "no change"Shows what the forecast addsNo benchmark
Misses in volatile monthsShows how it behaves when it mattersA track record that starts after the last shock
The direction of the missesReveals biasErrors reported only as averages without a sign

Does it give a range or just one line?

A single line says nothing about how sure the forecaster is. A range, or prediction interval, shows where the price is likely to land and how wide the uncertainty is, and scenarios show what would move it. Both let you weigh locking against waiting, instead of betting on the middle of a line.

The width of the range is information in itself. If it runs from below your budget price to well above it, waiting is a bet, and a fixed price may be worth its premium; we work through that trade-off in is a fixed-price offer worth the premium? If the range is narrow and sits below today’s quote, you have grounds to counter.

A range should also be honest. A 90% range should miss about one time in ten. If outcomes never fall outside it, it is too wide to help you decide; if they often do, it is too narrow. Scenarios add the “what if”: the Bernanke review recommended that “the central forecast should be regularly augmented by alternative scenarios”, so readers can see what a different path for the main drivers would do.

Do the inputs make economic sense?

Every price forecast rests on assumptions about what drives the price: feedstock, energy, capacity, demand, duties. Ask what the inputs are, where their values come from and when they were last updated. A forecast built on wrong inputs will be wrong however good the model, and one with no visible inputs can’t be checked.

Inputs carry much of the error. The ECB’s 2024 review found that adjusting for errors in its oil price assumption alone “reduces the HICP inflation RMSE by more than 15%”. An earlier ECB analysis notes that in less exceptional times its technical assumptions “explain the vast majority of the errors seen when forecasting energy inflation”.

For a material you buy, a few checks go a long way:

If a forecast says a resin will fall while its feedstock and energy costs are rising, ask why. There may be a good answer, such as new capacity or weak demand, but it should be one the forecaster can give.

Couldn’t we just ask ChatGPT?

You can ask, and you will get a plausible answer. It won’t be a forecast you can check: there is no track record, no stated inputs, no range you can test, and you may get a different answer if you ask again. Put the five checks to it and most of them have no answer.

General chatbots are good at explaining what drives a price, which helps when you prepare for a negotiation. A number you can defend to your CFO is a different job. It needs the price series you actually pay, a model fixed before the outcomes were known, and a record of how that model performed, including the months it got wrong.

Questions to ask before you act on a price forecast

Whoever produced it, these questions tell you how much weight it can carry:

  • How large are its errors at my horizon, and how large are the errors of a “no change” forecast over the same periods?
  • Which price exactly does it forecast, and is that the price in my contract?
  • Was it tested only on data available at each date, and how long has it run live?
  • Can I see every forecast and outcome for the past year or more, volatile months included?
  • Do the misses lean one way?
  • What range comes with it, and how often have outcomes fallen outside that range?
  • Which inputs drive it, where do their values come from, and when were they last updated?

From a forecast you can check to a decision

The checks tell you how far to trust a forecast. A buying decision then needs that forecast for your own materials, on the prices you actually pay, with its drivers and its uncertainty in view, at the moment a quote or an offer is on the table.

Sybilion helps industrial buyers decide what to pay for the materials they purchase, and when to commit: accept, counter, lock or wait. For the materials it covers, it gives a forward view of the price built from the cost drivers behind it, with the uncertainty stated as a range, so the inputs and the uncertainty are in front of you when you decide. It works alongside the systems you already use.

Check the forecast behind your next buy.

Frequently asked questions

What is a naive forecast?

The simplest possible forecast: the price at a future date will be the same as today’s. It is used as a benchmark: any price forecast should be compared with it at the same horizon, and one that can’t beat it adds nothing.

What is forecast bias?

A tendency to miss in the same direction, forecasting too high or too low again and again. You find it by looking at the sign of the errors over many periods, not only their size. A known bias can be corrected for; an unknown one quietly distorts every decision built on the forecast.

What is good forecast accuracy for a price forecast?

There is no single number, because it depends on the material, the horizon and how volatile the market is. A good price forecast has smaller errors than a “no change” forecast at your horizon, over a long period that includes volatile months, and comes with ranges that outcomes fall inside about as often as they claim.

Can ChatGPT forecast commodity prices?

It can explain what drives a commodity price and give a view, but that view has no track record and can change between runs. Use it to prepare questions for a supplier or analyst, not as a commodity price forecast to act on.

Explore more customer  stories

Frequently Asked Questions

What data do you use?

We use only the verified data from official institutions, market research companies, and other reliable sources vetted by us.

Each data source has to pass an extensive verification process before it is used in our analysis.
Each data source has to pass an extensive verification process before it is used in our analysis.Each data source has to pass an extensive verification process before it is used in our analysis.

How accurate are your trends?

We only provide forecasts that bring significant improvements (30%-70% relative error reduction) in comparison to established baselines.

What security measures do you use?

We use the latest and highest security standards in cloud architecture and access policies.

All data we used is anonymized and doesn’t contain any reference to customers or otherwise.

What do you mean by explainable?

Explainability means understanding why trends may unfold in a certain way and what external market factors influence them. Sybilion provides context and transparency to help you understand these factors.

Can I confidently share my data with you?

Yes. Our AI does not require data, that is significantly more sensitive than what you would anyway share in your annual reports.

We handle data with care and apply the latest security and hosting standards.