On this page
Write a claim that can be tested
Suppose a fictional process produces measurements whose population mean is being compared with a benchmark of 50. Specify what a measurement represents, how observations are selected and which population the claim concerns. A mean for all transactions is a different target from a mean for transactions selected after a filter.
Choose the alternative from the research question
A question about any departure uses a two-sided alternative. A question specifically about an increase uses an upper-sided alternative. Choose that direction before examining the sign of the estimate. Changing the direction after seeing the data changes the analysis procedure.
| Question | Alternative | Evidence direction |
|---|---|---|
| Any change from 50 | Mean ≠ 50 | Either tail |
| An increase above 50 | Mean > 50 | Upper tail |
| A decrease below 50 | Mean < 50 | Lower tail |
Record the model before calculating
For the known-variance normal example in this cluster, observations are independent normal measurements and the population standard deviation is known. Those are exercise assumptions. Observed financial returns may violate independence or distributional assumptions, so the example is not an automatic testing recipe for them.
Report more than a decision word
Retain the sample size, estimate, uncertainty, test statistic, reference distribution, p-value and planned significance level. Include filters and related tests that were attempted. A rejection addresses the stated null model; it does not by itself certify a profitable strategy or explain a causal mechanism.