What is selection bias?
Selection bias is a mismatch between the data a model learns from and the population it will make predictions about, caused by the way records were selected. The sample is skewed before any algorithm touches it. The model then learns the skew as if it were a fact about the world.
The classic business example is lending. A bank only observes repayment for the applicants it approved. A model trained on approved loans has never seen how rejected applicants would have behaved, so it inherits the judgment, and the errors, of the old approval policy. Survivorship bias is a related form: training on current customers alone omits everyone who already left.
Why selection bias matters
A model affected by selection bias can score well on its own test data, because the test data carries the same skew, and still fail or discriminate in production. That makes it a governance concern as much as a technical one. Fairness audits and bias detection start by asking who is missing from the training data.
A managed service provider that trains a ticket priority model on tickets from its largest client will see the model misjudge urgency for small clients with different systems. A hospital that trains a sepsis model on patients from its intensive care unit will see it underperform on the general wards. The remedy is usually to collect data from the underrepresented groups, reweight the sample, or state plainly which population the model is validated for and keep it there.