Checking an experiment daily and stopping at the first significant result inflates the false-positive rate several-fold, but the desire to stop early is legitimate: large effects deserve fast decisions and harmful ones fast rollbacks. Sequential designs make the monitoring legal. Group sequential tests, imported from clinical trials, pre-plan a few interim analyses with strict early thresholds that relax toward the final look (O’Brien-Fleming boundaries, generalized by alpha-spending functions). Always-valid methods like the mixture sequential probability ratio test go further, producing p-values that hold simultaneously at every moment, so continuous dashboards are simply safe.
The practical selection rule, from Spotify’s published comparison: group sequential tests are more powerful when the sample size can be planned in advance, while always-valid inference suits open-ended monitoring for regressions.
