Graduate Marketing Scientist, Fospha

• Working proficiency in SQL — you can investigate a discrepancy yourself rather than asking someone else to pull the data
• Python, or a demonstrated ability to pick it up quickly. Most of our analysis tooling sits there
• Grounding in inferential statistics — hypothesis testing, uncertainty, statistical power, and what a null result means
• Some exposure to experimental design — randomisation, control groups, confounding, and why a badly designed test is worse than no test
• Strong AI fluency — you use AI tools to get moving on unfamiliar problems and plug gaps in your own knowledge, and you QA the output before you rely on it
• Clear, concise written communication — a large share of this job is explaining something technical to someone who isn't
• Composure in client-facing conversation, including when the client is unhappy with a number
• Real attention to detail, and the discipline to log things consistently even when it's dull
• High agency — you'll be given ownership as fast as you demonstrate you can hold it

Desirable

• We're looking for a Graduate Marketing Scientist to join Fospha's Marketing Science team in London
• Fospha builds marketing measurement products for ecommerce brands — attribution, marketing mix modelling, incrementality testing, and brand impact measurement. Marketing Science owns the applied end of that: designing and delivering incrementality tests and MMM engagements for clients, and standing behind the numbers when a client challenges them
• This is the entry point into the function, and it is a hands-on one. You'll work on live test and MMM delivery under supervision from the start, and you'll be the first person looking at a client's data when a number doesn't behave the way it should. It's a role for someone who wants to learn causal measurement properly, in a business where it's the product rather than a side project
• Assemble and validate test data — geo-level spend and conversion series, checking pre-period parity between treatment and control, spotting the coverage gaps that invalidate a design before it launches
• Support test design under review — market matching and control selection, power and minimum detectable effect sanity checks, and identifying contamination risks such as geo-targeting settings that don't behave the way the platform's documentation claims
• Run analysis and read the results honestly — pre-treatment fit diagnostics, lift estimates with their intervals, and what a null result does and doesn't tell you
• Qualify client data for MMM — spend coverage across channels, whether there's enough variation in spend to identify an effect at all, series length and granularity, collinearity between channels, and gaps that will bias the result
• Assemble and validate model input datasets, and investigate the discrepancies that surface when you do
• Support model runs and read the diagnostics — fit, residuals, convergence, and whether a channel's estimated contribution is plausible
• Contribute to output-extension work under review — building on an existing MMM result, for example forecasting or budget scenario work derived from it
• Compare results across methods — where MMM, incrementality, and platform-reported figures disagree, understanding why is the interesting part of the job
• First and second line on client trust queries — investigating why a number changed, working in SQL against client data to isolate the cause
• Distinguish a bug from a methodology change — attribution window changes, model recalibration, data feed gaps, and platform reporting shifts all look similar from the outside and have very different signatures underneath
• Triage PSPs on model trust, resolve what you can, and escalate what turns out to be a genuine model problem with a clear diagnosis attached
• Reconcile platform-reported figures against our measurement — why walled-garden ROAS disagrees with ours is the hardest recurring question in the business, and you'll be learning it from the inside
• Log and tag incidents consistently, so recurring failure patterns become visible and can be automated away rather than repeatedly handled
• Run templated explainer sessions under review, walking clients through how our measurement works
• Draft documentation and presentations above the core explainer content, and feed recurring query themes back into the source material
• Fact-check methodology claims in product marketing collateral before it goes out