Advertising
Better campaign setups, decided before launch.
A ranking system that searches a vast campaign-configuration space and recommends the setups most likely to perform, while flagging the ones that only look good.
- Project
- Risk-aware campaign recommendation engine
- Client
- Global advertising & media group
- Capabilities
- Machine learning · Recommendation & ranking · Uncertainty estimation · Evaluation
- data points modelled
- 1B+
- of the best configurations found in the top 50 recommendations
- 97%
- CTR gap between our top pick and the best-known setup, in tested cases
- ≈ 0.1 pp
- confidence on every recommendation, so risky setups are flagged, not hidden
- Calibrated
See it work
campaign-builder
interactive demo · synthetic data
Top 5 · risk-adjusted
- 1
LinkedIn · Feed / Carousel
Retargeting 30d Low risk
2.05%1.88%–2.22% - 2
LinkedIn · Feed / Video 6s
Customer list Low risk
1.69%1.57%–1.81% - 3
TikTok · For You / Video 6s
Interest stack Low risk
1.60%1.53%–1.66% - 4
Meta · Stories / Static
Retargeting 30d Low risk
1.57%1.50%–1.64% - 5
Google · Display / Video 6s
Customer list Low risk
1.57%1.50%–1.64%
The problem
Before a campaign launches, a team has to choose an audience, geography, platform, placement, format and dozens of other settings. The number of possible combinations runs into the millions, and most teams decide from habit, gut feel and last quarter’s spreadsheet.
The cost of a weak setup is real: wasted media budget, slow learning, and campaigns that are “optimised” only after they have already underperformed.
Our approach
We treated set-up as a ranking problem, not a prediction problem. The question isn’t “what CTR will this exact configuration get?” but “out of everything we could do, which options should a planner look at first?”
- Representation learning places brands, audiences, placements and formats in a shared space, so the model can generalise to combinations it has rarely or never seen.
- Hard-negative sampling teaches the ranker the difference between a good setup and one that is almost good, which is where real decisions are made.
- Calibrated uncertainty sits on every prediction. A configuration with a high expected CTR but little supporting evidence is flagged as risky instead of being recommended first.
The result
An overwhelming configuration space became a short, ranked, decision-ready list. In testing, 97% of the best-performing configurations appeared in the top 50 recommendations, and the top pick landed within about a tenth of a percentage point of the best-known benchmark setup.
Just as important, every recommendation came with a confidence level, so the system was as clear about what it didn’t know as about what it did.
Client details are anonymised. Demos are illustrative, built on synthetic data and fictional brands.