Fitting Is Cheap; Validation Is Earned

You'll learnto test which model a data set really supports, not just fit a curve through it.

Any model can be forced through a point or two; that is fitting, not validating. Two tests that can fail decide which family deserves the data: the rate signatures hiding in a table, and the shape of the residuals. Watch three candidates compete, and watch a near-perfect fit fail to say anything at all about year thirty.

8 STEPS · 6 QUICK CHECKS · Δ · Δ² · RATIOS · RESIDUAL PATTERNS · v2

residual = actual − predicted

Before you start

What you're looking at

One data set with two models through the same two points; points used to fit and points used to check are marked apart.

The question

Away from the points used to build it, does the model still predict the data?

Watch for

The residual — actual minus predicted — and how a near-perfect fit can still fail the check points.

Step 1 / 8