Sabermetrics for the resort: a documented library of guest-worth metrics that goes past ADT, behavior-first personas your hosts can actually remember, and an AI analyst that shows its SQL and refuses to guess.
100% synthetic data. Every guest, rating, and offer is generated by a seeded simulator. No real patron data appears anywhere in this project.
The demo is the real dashboard, running on the synthetic warehouse.
The industry still ranks guests on ADT — average daily theoretical — the way baseball once ranked hitters on batting average. It is a single number with no peer context, no sample-size discipline, and no way to tell luck from worth.
Every metric here is documented like a baseball card: a definition, the formula, why it beats the ADT view, and a line on how to act on it. The same registry that computes a metric writes its glossary entry, its tooltips, and the analyst’s description of it — so a CFO can audit any number down to the arithmetic.
Also in the library: LCK, a luck index in game-math standard deviations, which separates a guest who is running bad from a guest who is worth less (and flags the persistent negative tail as possible advantage play). And RE / PEL, which ask the question reinvestment percentage cannot: did the offer buy incremental play, or did it pay someone to do what they were going to do anyway?
Worth indexed to comparable guests. 100 is exactly peer par; 150 is 50% above.
100 × worth-per-active-day ÷ peer averageWhy it beats ADT — The OPS+ move. A $60/day local and a $60/day fly-in are not the same asset — peer context is what raw ADT never gives you.
Worth per day, shrunk toward the peer mean in proportion to how little we have actually observed.
(n · worth + k · peer_mean) ÷ (n + k), k = 8 active daysWhy it beats ADT — Empirical-Bayes shrinkage. Two visits at $400 theo is weak evidence — one hot weekend can’t mint a fake VIP.
Probability that the current absence already exceeds this guest’s own visit rhythm.
P(gap ≤ days_since_last) under the guest’s own Weibull fitWhy it beats ADT — 21 quiet days is an emergency for a three-times-a-week local and noise for a quarterly visitor. Fixed churn windows miss both.
Guests are clustered on roughly sixteen behavioral features — worth, cadence, game mix, bet level, time fingerprint, promo response — with a Gaussian mixture and soft membership, so a guest can be 60/40 across two personas instead of being forced into one bucket.
Each cluster comes out as a character with a memorable name: Midnight Floor Loyalists, Weekday Video Poker Regulars. Each ships a stat line as percentiles of the whole book, psychographic traits that are always labeled with the behavior they were inferred from, and a playbook of trigger → action → the metric it should move.
Demographics are profiled, never clustered on. Age and distance describe a persona after the fact; they never define one. Clusters are gated on bootstrap stability before they earn a card, and guests with too little observed play are told so rather than being quietly assigned.
mid-week grinders, low bet, long sessions
Bars are percentiles of the whole book; the tick is the median.
Ask in plain language. The analyst matches the question against a curated catalog over a semantic layer, extracts the slots it needs — dates, metrics, personas, top-N — runs parameterized SQL, and narrates the result deterministically.
The query is always on screen. Numbers come from the warehouse, never from a language model’s memory of what a number like that usually looks like.
Off-catalog questions get a straight refusal plus the nearest questions it can answer. A model that says “I don’t have that data” is worth more than one that produces a confident, unfalsifiable paragraph — especially when someone is about to spend reinvestment budget on the answer.
Which personas have the best reinvestment efficiency this quarter?
Weekday Video Poker Regulars lead at RE 1.42 — $1.42 of incremental worth per comp dollar. Midnight Floor Loyalists trail at 0.31: their offers are largely paying for visits that would have happened anyway.
select persona_name,
round(sum(incr_worth) / nullif(sum(comp_cost), 0), 2) as re
from marts.patron_360
where last_active between $start and $end
group by persona_name
order by re descWhich guests are most likely to book a spa package?
I can’t answer that. This warehouse has no spa or hotel folio data, and I don’t guess at numbers. Here is what I can answer:
Five behavioral flags — session escalation, loss chasing, frequency spikes, late-night concentration, stake drift — each carrying the evidence that triggered it, composited into a risk tier a human can review.
Guests at the Elevated tier are automatically excluded from frequency-stimulating offers. That exclusion lives inside the persona playbooks, in the pipeline, not in the memory of whoever pulls the list this month.
Flags are evidence for a conversation, not a verdict. Every one of them shows the behavior and the window that produced it, so the person acting on it can see exactly what the model saw.
Auto-excluded from frequency-stimulating offers. The exclusion is enforced in the persona playbook itself — it is not left to whoever pulls the mailing list.
The demo measures gaming worth, because gaming is what the simulator generates today. Everything you can click on is floor data: trips, ratings, theo, comps, offers.
The warehouse, the metric registry, and the persona engine were built to take hotel folios, F&B checks, entertainment, and retail on the same guest key — so the same metrics resolve to one unified guest worth across the resort rather than a gaming number and four spreadsheets.
That is the direction, and it is honest to call it a direction. Property-wide worth is designed for and not yet shipped.
Python, polars, scikit-learn, and DuckDB for the pipeline; FastAPI over a read-only warehouse; Next.js for the dashboard. Determinism is a test, not a promise: the generator’s math invariants, the metric golden values, and the persona stability gates all run in CI.
The demo is the whole thing: the glossary, the persona gallery, guest 360s, the floor and promotion views, the responsible-gaming queue, and the analyst — all over 25,000 synthetic guests.