How accurate is a fertiliser price forecast?

The baseline any forecast must beat

Fertiliser prices are persistent: tomorrow usually looks a lot like today. That makes one very dumb strategy surprisingly hard to beat: assume the price simply stays where it was last printed. Forecasters call this the flat-carry baseline, and any model that cannot beat it is adding noise, not information.

So that is the bar we set. A grade earns the right to publish a modelled figure only by beating flat-carry by a clear margin, at least ten percent lower error, on months it has never seen. Where a grade cannot clear that bar yet, we say so on the chart rather than dress a flat line up as a forecast.

Tested only on months it has never seen

Accuracy measured on the data a model was trained on is fiction. Every grade is therefore tested walk-forward: the model is fitted only on months that came before the month it must predict, one month at a time, with at least three years of history behind the first test. The errors that come out of that process are the honest kind, out-of-sample, with no future information leaking backwards. The same discipline applies to changing the model itself: a candidate driver or fitting method runs the same walk-forward gauntlet against the live model, and ships only if the improvement survives statistical significance testing.

Grades earn their label, and can lose it

Not every grade has enough history to pass the test. Newer or thinner grades publish with an explicit early stage label carried in the data itself, and a deliberately reduced response, so the model moves them gently until the evidence firms up. The test re-runs with every calibration: a grade that accumulates enough evidence is promoted automatically, and a grade whose edge decays steps back down. Nothing about the label is editorial; it is the test's own verdict, and the API carries it on every figure it publishes.

What the shaded band really tells you

The band around each line is computed from that grade's own recent volatility, at a fixed one-week horizon. A calm grade wears a tight band; a jumpy grade wears a wide one. It is the model stating how far it trusts its own figure, and it is policed: a daily move larger than three standard deviations is held rather than published, and derived figures such as price per kilogram of nitrogen publish only while the latest UK print falls inside the band. The band-hits column in the scorecard shows how honest that band has recently been, grade by grade.

The model has to keep earning its place

A forecast validated once and then left alone rots quietly. Ours is recalibrated and re-scored continuously: every daily estimate is compared against the next realised weekly print, the walk-forward test re-runs on the growing record, and candidate improvements are trialled against the live model under the same rules. Most candidates lose, and the failures are recorded; the few that win with statistical significance ship. The scorecard above is the running output of that loop, not a snapshot taken the day this guide was written.

How to use a forecast well

Read the band before the line: the line is the centre of a range, not a promise. Use the direction and the band together to time purchases inside your own window, check the status label so you know how much validation stands behind the grade you buy, and compare any quote you receive against the going range rather than a single number. Used this way a forecast is a negotiating instrument; used as gospel it is a disappointment waiting to happen.