Suppose a phone plan's bills are 20, 25, 30 and 35 dollars for 0, 1, 2 and 3 GB of data. Each extra GB adds the same 5 dollars, so the model is linear: C(g)=20+5g. Now suppose a video's views are 100, 300, 900 and 2700 on days 0 to 3. The differences grow (200, 600, 1800), but every ratio is 3, so the model is exponential: V(d)=100⋅3d. The data, not the wording of the story, decide the family.
Each family has a signature you can test in a table with equal input steps. Linear: equal first differences. Exponential: equal ratios. Quadratic: equal second differences, as in a ball's heights 5, 21, 29, 29, 21 feet every half second, whose differences 16, 8, 0, −8 fall by 8 each time. Once the family is chosen, the parameters come from the data: a starting value, a rate, a vertex.
A formula that fits is not automatically a model to trust. The ball's height formula gives negative heights after it lands, and the exponential model of views eventually predicts more viewers than there are people. Modeling includes stating the domain where the pattern holds. Fitting is also different from choosing: a calculator's regression will find the best line for curved data too, so the family has to be chosen first.
Think of it like this. Choosing a model is like picking a key from a ring: the linear, quadratic and exponential keys each match one lock pattern, and you test the pattern (differences, second differences, ratios) before turning. The analogy stops at one point: real data never fit perfectly, so you take the key that fits best and measure how far off it is.
Core idea. Functional modeling turns a real situation into a function by choosing the function type whose behavior matches the data.