TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising

In online advertising, predicting whether a user will convert long after they click is a critical but difficult task. Systems face a two-clock, two-window problem: a short observation window (the click clock) provides timely but incomplete status updates, while a longer target conversion window (the conversion clock) captures the full conversion outcome but suffers from significant delays. Standard models often struggle to balance these two sources of information, as late-arriving conversions are often biased by historical traffic volume and shifting user intent.

TWICE addresses this by factorizing the conversion rate (CVR) into two components: a target-window conversion probability and a grouped elapsed-delay cumulative distribution function (CDF).

• Click-Clock Learning: The model uses clicks released after a short observation window to train the main CVR tower. A stop-gradient mechanism ensures that these fresh records update the CVR backbone without being distorted by incomplete delay information.
• Conversion-Clock Learning: Newly arrived conversions update a separate, lightweight delay module. To account for the fact that these conversions come from different historical cohorts, TWICE uses an arrival-conditioned temporal-convolution objective. This incorporates fixed historical predicted CVR (pCVR) mass, allowing the model to learn the timing of conversions without needing to re-process the original click features.

By separating these two learning paths, TWICE achieves high performance without the computational overhead of historical lookups or complex convolutions at serving time. The model produces monotone predictions for any horizon up to the target window, ensuring consistency. In industrial deployment at Kwai, this architecture significantly improved key business metrics—including revenue and conversion volume—while maintaining the strict latency requirements necessary for real-time advertising systems.