Design Ad CTR Prediction

Ad click-through-rate prediction is the extreme-throughput calibrated classification question. Unlike a general recommendation problem, the output is not merely a relative ranking score: an auction may multiply the predicted probability by a bid, so a 2x calibration error can become a 2x pricing error. Interviewers use this problem to test whether we can connect biased impression logs, sparse clicks, fresh features, strict latency, and marketplace economics into one coherent system. This walkthrough follows the Interview Framework, with serving constraints pulled forward because they eliminate whole model families before modeling begins.