Model validation
Objective. Separate calibration, validation and extrapolation.
- 1. Learn
- 2. Worked example
- 3. Practice
- 4. Feedback
- 5. Continue
Learn: the key idea
Model validation checks whether a fitted model's predictions hold up against data it was not trained on, separating three distinct stages: calibration (fitting parameters to known data), validation (testing against independent data), and extrapolation (applying beyond the observed range, with reduced confidence). A model that calibrates well but fails on validation data is overfit, while good validation performance still doesn't guarantee reliable extrapolation far outside the data's range. Example: a growth model calibrated on five years of data might validate well on year six but should be used cautiously to predict year fifty, since unmodelled effects may dominate far outside the observed regime. Careful validation distinguishes genuine predictive power from mere curve-fitting.
Worked example
A model calibrated on data from 2010–2015 is tested against new data from 2016–2017 and performs well. According to model-validation stages, what is this testing step called, and is extrapolating to 2050 equally reliable?
- 1Identify the stage: testing against new, unseen data from 2016–2017 is validation, not calibration
- 2Recall calibration already used the 2010–2015 data to fit parameters
- 3Extrapolation to 2050 goes far beyond the data range used in both calibration and validation
- 4Conclude this step is validation, and extrapolation to 2050 is not equally reliable
Answer: Validation; extrapolation to 2050 is not equally reliable
Concept mastery
Based on your past attempts at this lesson — the question types to practise again come first.
Complete a practice set to see your concept breakdown here.
Practice
Practice happens on its own screen, one question at a time. Answers stay hidden until you submit yours, and you can stop and pick up at the same question later.
Start practisingCommon mistake. Students validate a model using the same data used to calibrate it, which only confirms the fit, not genuine predictive power; validation requires independent data the model has not already seen.
Your progress
Tick each step as you finish it. Your exact reading position is saved automatically, so Continue drops you back on the same line.
0 of 5 steps complete
Your place saves automatically as you read.