Methodology — Tizy Score and Tizy Nightlife Index

This edition — what the algorithm decided about itself

Edition v1-2026-10-11 · computed 2026-10-11 11:11 UTC · status published. Since v1 the component weights are not set by hand: every edition they are re-fitted so the score best ranks venues by an external market signal — how many people Google each venue by name — and an edition is published only if it passes the drift gate below.

Fit: fallback to prior weights (spearmanHoldout=0.075 < minSpearman=0.35) · not fitted (too little coverage): l. Label: log(1 + branded searches / month), residualised on log(1 + reviews).

705 venues carried a market label; graph 3519 matched venue nodes / 79193 artists.

Version v0, 2026-09. This page is the source of truth for how every grade on tizy.app is computed. Constants are provisional and frozen per edition; historical editions are never rewritten.

What is graded

A venue is eligible when its category is on the nightlife allowlist (night clubs, bars, lounges, live-music rooms, rooftops — restaurants, casino hotels and shops never get a nightlife grade), it has a public Google rating, and it has at least 50 reviews. Below that it is shown as unrated, not as a low grade. A grade is shown greyed as stale when the underlying rating snapshot is older than 180 days.

Tizy Score — five components, 0–100

Quality (Q). A Bayesian, category-normalised rating. We shrink the venue's star rating toward the mean rating of its own category with a prior worth 50 reviews (R_b = (n·r + 50·C_cat) / (n + 50)), then shift it onto the global scale (R_adj = R_b − C_cat + 4.452) and map 3.0→0 and 5.0→100 linearly. Night clubs average 3.99 stars on Google, cocktail bars 4.58 — without this step every club in the world would grade D.

Popularity (P). 80% log-scaled review count (saturating at 20,000 reviews) plus 20% log-scaled RSVPs across the venue's events in a ±90-day window (saturating at 5,000). Log scale so a casino with 140,000 reviews cannot dominate.

Programming (G). From events in [now−90d, now+90d]: 50% log-scaled event count (saturates at ~60), 30% distinct artists after a junk-token stop-list (saturates at ~80), 20% distinct genres (capped at 10).

Transparency (T). 25 points each for a published price range, opening hours, venue attributes and popular-times data. The one component a venue can fix itself.

Liveness (L). Hours streamed live on Tizy in the last 30 days, saturating at 40 hours. Tizy streams from venues, so this is a separate, labelled component with a hard 5% ceiling; it never touches Q or P.

Authority (R) — VenueRank. New in v1. Venues and the artists who play them form a graph; authority flows through it the way PageRank flows through links: a venue is strong when it books artists who also play other strong venues, and a name that plays everywhere carries less per booking than a curated one (artist edge weight 1 / log(2 + venues played)). Computed by power iteration over the global venue–artist graph with damping 0.85, then expressed as a percentile within the city (and globally, in the API). No hand-set weights inside R: the graph decides. Junk "artists" leaking from listings (payment methods, US states) are removed by a stop-list and a > 40-venues rule.

Freshness (F). New in v1. Recency of the venue's programming (days since its last listed event, half-life 45 days) and of its rating snapshot (half-life 180 days), so a venue that stopped booking decays instead of coasting on old reviews.

The label, not a component — market demand (D). Monthly Google searches for the venue by name ("club space miami"), measured through Google Ads volume. D never enters the score directly; it is what the weights are trained against, exactly as a search engine trains ranking features on what people actually look for. It is shown on the city page so anyone can compare the ranking with the market.

Weights — learned each edition, then governed

Prior weights (used when the fit is too weak): Quality 0.30 · Popularity 0.15 · Programming 0.15 · Transparency 0.10 · Liveness 0.05 · Authority 0.15 · Freshness 0.10. Each edition the engine fits non-negative weights (bounded 3–45 % each, ridge-regularised toward the priors) on a deterministic half of the labelled venues and reports the rank correlation on the other half; below 0.25, or with fewer than 80 labelled venues, it falls back to the priors and says so above. Governance, all automatic: a venue's score moves at most ±15 points per edition (0.7·new + 0.3·previous, then clamped; flagged dampened); a city whose new ranking correlates < 0.75 with the previous edition, or where more than a quarter of venues jump > 15 points, is held — the previous grades stay published and the edition is logged for review; a venue whose review count grows more than 4× the city-median daily velocity between snapshots has its popularity frozen and is flagged review-velocity. Every edition (weights, fit, drift, status) is stored in score_editions and rendered at the top of this page; historical editions are never rewritten.

Weights and coverage-aware renormalisation

Score = Σ wᵢ·Sᵢ / Σ wᵢ   over components MEASURED for the venue's city
w:  Quality 0.40 · Popularity 0.20 · Programming 0.25 · Transparency 0.10 · Liveness 0.05

A component is unmeasured — dropped from both numerator and denominator — when its data source has zero coverage for the city. In v0 Liveness is unmeasured everywhere (no streaming log yet) and Programming is unmeasured in Las Vegas (no event data ingested). Venues are not penalised for gaps in our ingest; the city page says which components are measured.

Grade bands

Tizy Nightlife Index (city)

CityScore = 0.35·mean(Q of graded venues) + 0.25·EventDensity + 0.15·Demand + 0.15·Diversity + 0.10·LiveShare

Event density and demand are normalised per 100k population and require an external population table; until it is loaded (v0) they are unmeasured and the remaining weights renormalise, so the v0 index is mean venue quality plus category diversity (Shannon entropy of category shares). Cities are never ranked by raw venue count — our venue sample is capped at a few hundred per city and says nothing about the real universe. A city is indexed only with at least 30 graded venues.

Data sources and limits

Public rating signals (Google Maps rating, review count, price range, hours, attributes, popular times; one snapshot, March 2026), Resident Advisor event listings with lineups, genres and RSVP counts (June–October 2026), and Tizy's own streaming logs. Known limits in v0: a single rating snapshot (no review velocity yet), one global programming cap for all categories, a hard ±90-day event window instead of decay, and RA-only event coverage. Review text is never republished; we publish the transform, not the raw table.

Anti-gaming

The Bayesian prior blunts small review bursts; RSVP inflation can move a grade by at most ~4 points because P_att is capped at 20% of P; a review-velocity guard (freeze Q and flag "under review" at >3× category-median velocity) ships with the second rating snapshot. Every grade stores its input vector and the JSON API returns the components, so anyone can recompute a score.

Cite this

"Tizy Nightlife Index, edition v0 (2026-09), tizy.app/rating/methodology". Data is available under CC BY 4.0 via /rating/api/v0/. To challenge a grade, email the venue's location id and the disputed input to hello@tizy.app.