BlyxDocumentation
Browse chapters
Documentation / Numerical programming

Measure prediction error

Evaluate a set of predictions, trace the residuals and interpret their mean squared error.

Compare paired observations

Take one prediction and the matching observed target. Their difference is a residual. Its sign tells you whether the prediction is high or low. Averaging signed residuals alone could hide errors that cancel one another.

Accumulate the loss

With predictions [2, 3, 5] and targets [1, 4, 3], the residuals are 1, -1 and 2. Squaring gives 1, 1 and 4. Add the squares, then divide 6 by the number of observations, 3. The mean squared error is 2.

Match types and lengths

The loop index is an integer for at. The floating-point count supplies a same-type divisor for squares in both execution paths. For inputs supplied by a caller, check that the lengths match and that at least one pair exists before you compute the mean.

Use held-out data

Training chooses a slope or other parameters. Evaluation asks how those fixed parameters perform on examples that did not determine them. Keep a second vector of test targets when comparing two models. A low training loss alone says little about unseen observations.

Try the program

Output Ready
Expected output
2
Open in the full playground →

Practice

Replace the predictions with the three actual targets. What is the best possible value of this loss?

Hint

Each residual should be exactly zero for these inputs.

Worked solution

All squared residuals are zero, so the mean squared error is zero.

Find

[2.0, 3.0, 5.0]

Replace with

[1.0, 4.0, 3.0]

Output after the change

0