Applied Statistician / Bayesian Modeler

Applied Statistician / Bayesian Modeler Engagement: Small fixed-scope project, ~1–2 weeks of work Location: Remote Compensation: Fixed price (propose your number) What this is We run a data analytics platform with 20+ years of proprietary outcome data in a specialized professional-placement domain. We've done the hard parts: the data is collected, cleaned, organized into labeled Excel files, and we know which signals should drive the prediction and roughly how they should stack. What we need built is the model itself — a defensible 1–5 fit score that turns our signals into a single number a client can trust. We can describe the logic in plain language; we need a statistician to turn that logic into a rigorous, validated, documented scoring model we can stand behind. This is a tight, well-defined job. You are not collecting or cleaning data, and you are not designing the strategy from scratch — the inputs and the signal logic are ready. You're building the math that makes the score honest and defensible. What you'll be handed Clean, labeled Excel files (the outcome data, the geographic signal, the composition percentages — all pre-organized and documented). A written explanation of which signals matter, and how we believe they should combine. What you'll build A scoring model that combines our signals (geographic receptivity, composition pipeline signal, trajectory-intent lift) into a calibrated 1–5 fit score. Proper handling of the small-sample problem (many cells have low counts) — the score must not be fooled by tiny samples that look extreme. Calibration: a "4 out of 5" must correspond to a real, defensible probability band, not an arbitrary cutoff. Validation against historical outcomes (backtest / holdout) showing the score actually predicts what we say it predicts. Required Strong background in hierarchical / Bayesian modeling on sparse count data. (If that phrase describes work you've actually done, say so specifically.) Fluency in Python (PyMC/Stan a plus) or R (brms/rstanarm). Comfort with calibration and validation, not just fitting a model. Ability to explain statistical reasoning in plain language. Strong plus Background with health-outcomes, placement, or admissions-type data. NOT what we're looking for Deep-learning / neural-network approaches. The dataset is small, interpretable, and high-stakes — a black box is the wrong tool and we will not deploy one. Anyone who can't explain their model simply. Two deliverables A documented, deployable scoring function our engineer can run in production. A plain-language write-up explaining every weight, cutoff, and assumption — so we can defend the score to a client. If you can't explain it in plain English, it isn't done. Screening question (please answer in your application) One program has 9 total members, 7 of a given type. Another has 280 members, 151 of that type. A naive rate says the first is "78%" and the second "54%." Why would it be a mistake to tell a client the first is more receptive — and how would you handle it? (We're looking for recognition of small-sample unreliability and a shrinkage / regularization / Bayesian-prior style answer. This separates the right candidate from the wrong one in one paragraph.) Confidentiality This role requires access to proprietary data under an NDA You will have to provide scanned ID with signed NDA The data may be used only for this engagement and may not be retained, reused, or disclosed afterward. Whatever you build is work-for-hire and belongs to us. Details provided before data access.