Empirical Probability Models
Build Probability Models from Observed Frequencies
Observed proportion is the model input
An empirical probability is estimated as the number of observed occurrences divided by the number of comparable trials. The numerator and denominator must refer to the same defined event and observation process. A relative frequency is an estimate, not a claim that the event has an immutable true chance equal to the observed decimal.
Comparable data matter
An empirical model transfers best when future cases resemble the conditions under which data were gathered. Changes in population, measurement process, or time can make a past relative frequency a poor predictor. More observations often make an estimate less variable, but sample size alone does not prove representativeness.
Use calibration language
When a model assigns probabilities to mutually exclusive and exhaustive observed categories, their relative frequencies should sum to one apart from rounding. A check of counts, denominators, and total probability catches common errors such as using a count as a probability or combining incomparable datasets.
Mark Yourself Question
- Grab a piece of paper and try this problem yourself.
- When you're done, check the "I have answered this question" box below.
- View the solution and report whether you got it right or wrong.
A quality log shows 39 delayed shipments among 120 shipments processed by one warehouse during a defined week. Calculate the empirical probability of delay for that warehouse-week process and give it as a decimal.
A retailer uses last month's observed return rate to estimate returns next month. Which condition most supports this use?