Exemplar analyses

Worked examples on real generated arms

One full grocery-sim engagement, split across three pages, following the Analysis catalog:

• Case description, the client-facing brief exactly as an analyst would receive it: a neighborhood grocer three years in, deciding whether to renew a lease after a record-revenue year that somehow netted him nothing.
• Stakeholder report, the plain-language answer, in the client’s own question order. Descriptive P&L, trend/seasonality decomposition, shrinkage forensics, an observational competitor-impact estimate, a full-cost staffing counterfactual, customer-panel churn, and a 2028 forecast (Layers 0, 1, 3, 4, 5, and 7 of the catalog) in one pass.
• Technical report, the review copy, on the same 3y_baseline run and the same eight questions as the stakeholder report, restated as an analyst review would expect: model specifications, coefficient tables with robust standard errors, confidence intervals, and diagnostics behind every plain-language claim above (Layers 1, 3, 4, 5, and 7 of the catalog).

And separately:

• Advanced methods demonstration, a standalone stress test on its own, separately generated run: a negative-binomial GLM with HAC-robust errors, a difference-in-differences cost pass-through design (and an honest account of where it fails to identify a clean number), gradient-boosting demand forecasting and stock-out classification judged on precision/recall, and a hierarchical Bayesian partial-pooling model with full posterior diagnostics (Layers 2, 3, and 6 of the catalog), not part of the Malm’s Market engagement, shown to demonstrate what the package’s output can support beyond what one ordinary engagement calls for.
• Auto-generated briefs, the raw, unedited output of sim.describe() for the two other engagement types it produces (optimize and invest, alongside the case description’s diagnose), with no accompanying analysis, shown to demonstrate that the brief generator itself varies the kind of engagement, not just the story.

More pages will be added here over time. In the meantime, archive/analyses/ in the main repository already has a full-depth, graded notebook for every layer of the catalog, and the package’s own sim.create_analysis(path) scaffolds a starting marimo notebook against any run you generate yourself.