Data Scientist

Project Description

A B2B marketing analytics platform for LinkedIn advertisers wants to know whether the early signals of an ad campaign can predict its outcome months ahead. Small B2B advertisers close only a handful of deals per year, so standard attribution has too little data to work with. The proof of concept tests a different approach on the platform's historical data across many advertisers. Engagement and website signals from the first 7-21 days of a campaign are combined into a surrogate index that predicts later pipeline outcomes and states how confident the prediction is.

The PoC is time-boxed to 5-6 weeks and ends with a Go or No-Go decision backed by numbers.

Client Description
A B2B marketing analytics platform for LinkedIn

• A B2B marketing analytics platform for LinkedIn advertisers wants to know whether the early signals of an ad campaign can predict its outcome months ahead. Small B2B advertisers close only a handful of deals per year, so standard attribution has too little data to work with. The proof of concept tests a different approach on the platform's historical data across many advertisers. Engagement and website signals from the first 7-21 days of a campaign are combined into a surrogate index that predicts later pipeline outcomes and states how confident the prediction is.
   

Requirements:

• 5+ years of applied data science or statistics at a senior level.
• Strong command of regression modeling, including regularized and hierarchical (multilevel) models.
• Proven experience with small or sparse datasets and with probability calibration.
• Rigorous validation practice: temporal splits, leakage prevention, overfitting control.
• Python (pandas, scikit-learn, statsmodels) and SQL.
• Ability to explain uncertainty to non-technical stakeholders.
• English - upper-intermediate.

Nice to have

• Bayesian tooling such as PyMC, Stan or bambi.
• Familiarity with surrogate index and proxy metric methods, such as the work of Athey, Chetty, Imbens and Kang.
• B2B marketing analytics: attribution, account-based marketing, LinkedIn Ads, CRM pipeline data.
• Holdout design, controlled experiments and sequential testing.
  Statistical process control.
• Part-time, 0.5-0.8 FTE, about 20-32 hours per week.
  5-6 weeks, starting in October 2026, exact date to be confirmed.
• Possible continuation into productization if the PoC succeeds.

Responsibilities:

• Define, together with the client and a business analyst, what counts as campaign success, the outcome window and the cut-off between signals and outcome.
• Design the analytical dataset: candidate signals, outcome labels, exclusion rules and safeguards against future information leaking into the signals. A Python data engineer builds the dataset in ClickHouse to this design.
• Build the surrogate index as a regularized or hierarchical model, with signal weights shared across advertisers and adjusted per advertiser in proportion to its own data volume.
• Validate the model on campaigns it has not seen: temporal backtesting and leave-one-advertiser-out evaluation, with AUC, Brier score and calibration curves.
• Compare the model against the current practice of judging campaigns by CTR and CPC, and measure how prediction quality changes between day 14 and 21.
• Run error analysis, source ablation and learning curves to show which data is missing for a reliable forecast.
• Present weights, patterns and uncertainty to the team and reccomend Go or No-Go for the future phase.