Fitting data versus making a prediction: what is the difference?

A curve can pass through the points on a graph and still leave the hardest question unanswered: will it work on observations that did not help shape it?

About this picture

Conceptual illustration of the distinction between fitting known observations and predicting beyond them. The curve is decorative, not scientific data.

Fitting data means adjusting a model to match observations. Testing a prediction means checking what that model says against evidence kept separate from those adjustments. Both matter. They answer different questions.

Think of studying for a test

Imagine practicing with ten questions and keeping the answer sheet beside you. You can learn something useful. You can also become very good at those ten questions.

A fresh set of questions tells you something the practice sheet could not: whether what you learned carries over.

Scientific models face a similar challenge. A model is a simplified description of how something behaves. Its adjustable settings can be chosen using measurements. That is a fit. A close match is worth investigating, but the measurements have already helped choose the answer.

What makes a prediction a fair test?

One useful approach is to hold some observations aside while choosing the model and its settings. Another is to state an expectation before new measurements arrive. In either case, the test needs clear rules about what will count as agreement, how uncertain the measurements are, and what would count against the model.

The word “new” can be misleading here. Evidence does not have to be collected tomorrow to test an idea. What matters is whether it was already used to build or tune the particular claim being tested.

There are other checks too. The NIST statistics handbook explains why a single impressive fit score is not enough: the pattern of differences between a model and the measurements also matters.

Why this matters for HD-Blast

In GOD PLAYS DICE, I explore HD-Blast, my unproved hypothesis about a higher-dimensional origin of the Big Bang. A curve that matches some observations would not, by itself, establish that proposed cause. Different physical explanations may produce similar patterns.

A useful question is therefore: which observations helped choose this curve, and which could challenge it afterward? A second question follows: could another explanation do as well or better?

That distinction belongs beside the one in my earlier guide to conditional mathematical results. A calculation can be correct under its assumptions while the physical explanation remains open.

I want readers to have a way to ask those questions without needing a physics degree. Explore GOD PLAYS DICE and its free preview for the longer journey.

Source checked September 24, 2026: NIST/SEMATECH e-Handbook, “How can I tell if a model fits my data?” This article explains a testing distinction; it announces no new HD-Blast discovery.