Projects
Case studies with a pre-registered claim, a baseline table, calibration evidence, a decision in the client’s units, and a repository a stranger can run.
Each case study is written in the same order, so once you have read one you know where to look in the next: the headline number in the client’s units; the decision it serves; the data and its problems; the claims I registered before evaluation and how each could have failed; the baseline table, then where the better-sounding model lost; the decision table; what runs on a schedule after handover; limitations; artefacts; and the published result I compared against.
Work in progress appears as a pre-registration, with the hypotheses and refutation rules posted before the results exist. There are no placeholders for work I have not started.
-
Monthly attendances and emergency admissions for every English provider, forecast one to six months ahead on the data as first published, with the answer expressed as escalation beds per trust.28–35% lower error than seasonal naive, and still behind ETS on interval score
-
Published cell-culture media campaigns replayed on their own measurements, model-guided against random; a planning table for the next campaign; a CLI that runs on a client’s CSV. Eleven pre-registered hypotheses with their verdicts.45–65 vs 95–105 experiments to a top-1% medium, model-guided against random, on enriched pools
-
Eight published studies catalogued to one schema, each with a dataset card recording units, design, replicate structure, licence and the loading decisions that move results as much as the choice of model.8 studies on one schema with a licence register; six mirrored, two as fetch instructions