Idea
Model improving long-horizon conversion predictions to boost online ad revenue and conversion rates.
Research Paper
Core Innovation
This paper introduces TWICE, a framework that factorizes post-click conversion rate into target-window conversion probability and a delay distribution modeled via two clocks. It uniquely combines click-time supervision and conversion-time feedback to handle delayed conversions and cohort effects, enabling monotone, efficient predictions without historical data lookups.
Why It Matters
Accurate long-horizon conversion prediction is critical for optimizing online advertising spend and maximizing revenue. TWICE reduces uncertainty from delayed feedback, enabling advertisers to better allocate budgets and improve campaign performance. Its scalable design supports real-time deployment, transforming ad conversion forecasting workflows.
Market Size (TAM)
$20–50B TAM for online advertising conversion prediction; $5–10B SAM from digital advertisers and ad platforms. Driven by growth in programmatic advertising and demand for ROI optimization.
Potential Customers & Pain Points
- Online advertisers – Need accurate conversion forecasts despite delayed feedback
- Ad tech platforms – Require scalable models for real-time bidding and budget allocation
- E-commerce companies – Seek improved ROI from advertising spend.
Business Model
SaaS platform or API offering advanced conversion prediction models to advertisers and ad tech companies, priced by usage or subscription.
Competitive Landscape
- Google Ads conversion modeling
- Facebook Ads conversion prediction
- Criteo conversion optimization
- Appsflyer attribution models
Implementation Challenges
- Integration complexity with existing ad tech stacks
- Data privacy and compliance constraints on user tracking
- Competition from established ad platform models
Validation Strategy
- Conduct additional A/B tests across diverse advertiser verticals
- Benchmark against leading commercial conversion prediction tools
- Pilot integrations with major DSPs and ad exchanges
Research Paper Overview
TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising
Summary
TWICE addresses delayed feedback in online advertising by modeling conversion prediction with two complementary clocks: a short click clock for recent clicks and a longer conversion clock for delayed outcomes. It factorizes conversion rate into target-window probability and delay distribution, enabling accurate, monotone predictions without historical lookups. Tested on public and industrial datasets, TWICE improved revenue and conversions in live A/B tests and is deployed at scale.