Model guide
Three stories a curve can tell
Start with equally spaced observations. Look at what changes from one step to the next, then test whether the pattern keeps working.
Linear growth adds.
Consider 10, 15, 20, 25. Each step adds five. The percentage increase falls as the starting number grows. The lab fits a straight line to the early values and floors its predictions at zero because the observations here are nonnegative.
Exponential growth multiplies.
Consider 10, 15, 22.5, 33.75. Each step multiplies by 1.5. The absolute change grows. Fitting a straight line to the logarithms of positive observations estimates the multiplier. A zero has no logarithm, so this model is unavailable when the training window contains zero.
A plateau has a limit.
Early growth with a fixed limit can look like multiplication. Later, each new step gains less as the remaining room shrinks. The lab tries a range of limits above the largest training observation and chooses the logistic curve with the smallest squared errors on those early observations. The fitted limit can be poorly determined from a short window.
Try a forecast before looking at the answer.
- Open “A small audience.” Start with four observations. Notice that several models can follow the early points.
- Choose the model you expect to do best, then reveal the later weeks. Compare the mean absolute error: the average size of a forecast’s misses, in the original units.
- Raise the training window to nine observations and repeat. The fitted lines and the dates used to test them both change. Compare models within the same window.
A useful test keeps the later values out of the fit. A model can win this comparison and still fail when conditions change or you forecast farther ahead. These cases use fictional data for practice.
For the model shapes, see OpenStax Precalculus. For the distinction between fitting and testing, see Forecasting: Principles and Practice.
Open the growth lab