Work out stratified split sizes instantly with clear inputs, formula shown and shareable results.
Stratified splitting keeps the class ratio identical in every partition, which matters most when positives are rare: a random split of an 8 percent positive set can easily land a test fold with half the expected positives, making recall estimates swing wildly. The key sanity check is the absolute number of positives in the test set, not the percentage.
Stratified counts
test rows = rows x test share; test positives = test rows x positive rate; train positives = total positives - test positives
As a rule of thumb at least 50, and ideally a few hundred, before precision and recall are stable enough to compare models. Below about 20, confidence intervals are wider than most model differences.
Yes. Stratifying on a sensitive attribute or on time bucket as well as the label helps ensure the test set represents the population you care about.