Calibrating a pricing model against your own machines

Every estimate of machine time rests on coefficients that are, honestly, unknowable from outside your shop. How much of a print is travel rather than extrusion. How densely your nester really packs a cake. How much resin actually cures inside the envelope a support occupies.

A vendor who hard-codes those numbers is telling you their guess is better than your data. It is not.

What calibration is

You have printed hundreds of parts. For each one you know two things the estimator did not: how long it really took, and how much it really weighed. That is a dataset, and it is enough to fit the coefficients that are allowed to move.

The procedure:

1. Pick three or more parts you have actually made, on the machine you want to calibrate.
2. For each, record the real machine time and the real mass.
3. Let the fitter search the calibratable coefficients within their honest bounds.
4. Look at the error before and after, and at exactly which numbers it wants to change.
5. Apply it to a draft, look at the draft, and publish - or do not.

Step 5 matters. A fit is a suggestion, not a fact.

Why relative error, not absolute

If you fit on absolute error, a two-day build dominates every two-minute print in the dataset. You end up with a model that is excellent at the jobs you get one of a month and wrong about the jobs you get twenty of a day.

So fit on relative error - being 10 % out on a short job counts the same as being 10 % out on a long one. In practice this means minimising the mean squared error of the logarithm of the ratio, which has the pleasant property of treating “twice as long” and “half as long” as equally wrong.

The number reported back to you should still be a plain percentage, because nobody wants to read a log.

Which coefficients should be allowed to move

Not all of them. A coefficient is a reasonable calibration target when it is:

• genuinely uncertain - it depends on your machine, your firmware, your habits;
• bounded - there is a range outside which any value is obviously wrong;
• monotonic in its effect, so the search is well-behaved.

A travel factor qualifies. A nesting efficiency qualifies. The price of your filament does not: you know that number exactly, and letting a fitter “improve” it would be absurd.

Quotometry marks the calibratable settings explicitly, with their bounds, and refuses to fit anything else.

What to expect

Most shops start somewhere around 15–25 % mean error against their own history, and land in single digits. Two things usually explain a fit that does not improve much:

• The dataset is homogeneous. Three near-identical parts constrain very little. Include a tall one, a flat one, a dense one and a hollow one.
• The recorded times include something the model does not. If your “print time” includes bed heating or the twenty minutes the job sat waiting, the fitter will absorb that into a coefficient and make everything else slightly wrong. Record the machine’s number.

Recalibrate when the shop changes

New firmware, new slicer profile, a different filament brand, a new machine. Calibration is not a one-off set-up step; it is a thing you do again when something changes, which takes about five minutes because the dataset is your own job history.

How calibration works in Quotometry →