Fitting and refining a function model
Introduit| English | Sens |
|---|---|
| residual/rɪˈsɪdʒuːəl/ | résidu |
A curve on a volcano, since 1958
- In March 1958, Charles Keeling began measuring carbon dioxide in the air at Mauna Loa in Hawaii. The annual mean was 316.9 ppm in 1960, 354.5 ppm in 1990 and 414.2 ppm in 2020. A straight line through the first and last points misses 1990 badly; a quadratic fits all three. Choosing and testing a model like this is what scientists do with every data set.
- A model is judged by its residual 残差 values (observed minus predicted) on data it was not built from.
A model predicts 368 ppm; the observed value is 369.7 ppm. Find the residual.
Difficulté 2/5 · 140 pièces + un petit bonus lors de la première réussite aujourd'hui
Observed minus predicted.
Fitting a model through points
- A line needs two points; a quadratic $y = a + bt + ct^2$ needs three, giving three equations.
- Choose a sensible variable: $t$ = years since 1960 keeps the numbers small.
How many points are needed to fix a quadratic model $y = a + bt + ct^2$ exactly?
Difficulté 2/5 · 140 pièces + un petit bonus lors de la première réussite aujourd'hui
Three unknowns.
Testing with a new point
- A model that passes through every point used to build it has residual 0 there — that proves nothing.
- Predict a point you held back, and compare the residuals of the competing models.
Fitting and refining a function model
State inputs, outputs, units and the modelling domain
Check the evidence and conditions behind a transformed or fitted function.
Worked example: line or quadratic?
Let $t$ = years since 1960. Compare the line through $(0, 316.9)$ and $(60, 414.2)$ with the quadratic through all three points, using the 2000 value of 369.7 ppm.
- Line: gradient $\tfrac{97.3}{60} = 1.62$: $C = 316.9 + 1.62t$. At $t = 30$ it gives 365.6 — residual $354.5 - 365.6 = -11.1$.
- Quadratic: $30b + 900c = 37.6$ and $60b + 3600c = 97.3$ give $c = 0.0123$, $b = 0.885$: $C = 316.9 + 0.885t + 0.0123t^2$.
- Test at $t = 40$: line 381.7 (residual $-12.0$); quadratic 371.9 (residual $-2.2$). The quadratic is much better: the growth is speeding up.
A line passes through $(0, 20)$ et $(10, 50)$. Predict $y$ at $t = 4$.
Difficulté 2/5 · 140 pièces + un petit bonus lors de la première réussite aujourd'hui
Gradient 3: $20 + 12$.
At a held-back point, model A has residual 6 and model B has residual $-1$. Which is better there?
Difficulté 1/5 · 100 pièces + un petit bonus lors de la première réussite aujourd'hui
Smaller absolute residual.
The model's domain
- State where the model is meant to hold: here $0 \le t \le 60$, the range of the data.
- A model for a ball's height or a tank's depth also has a physical domain, such as $0 \le x \le 6$.
Refining the model
- If residuals show a pattern (all negative in the middle, positive at the ends), the model has the wrong shape.
- Refine by changing the form — line to quadratic, or quadratic to exponential — not just the constants.
If all the residuals in the middle of the data are negative and all at the ends are positive, the model probably has the wrong shape.
Difficulté 1/5 · 100 pièces + un petit bonus lors de la première réussite aujourd'hui
A pattern means the form is wrong.
Three exact fits do not prove a model. Any three points fit a quadratic exactly. Only predictions for new data, or a reason from science, show that the model is right beyond those points.
A quadratic that passes exactly through three data points must be the correct model.
Difficulté 1/5 · 100 pièces + un petit bonus lors de la première réussite aujourd'hui
Any three points fit a quadratic.
Build a model from some data, then test it on data you did not use. Compare residuals, look for patterns in them, refine the form of the model if needed, and state the domain where it is valid. Exact fits to the building data prove nothing.
A ball's height is recorded as 1.0 m at $x = 0$, 3.5 m at $x = 2$ and 2.0 m at $x = 5$ (horizontal metres). Find a quadratic model, use it to predict where the ball lands, and give one reason the model might fail.
Fitting and using a model.