Ask "what's the RTP of this slot?" and you usually get one number, copied from a press release.
We capture every game from its official demo, and the paytable tells a different story: 618 of
1,333 games (46%) ship with more than one RTP configuration. Each one has a condition attached,
and the condition is printed right there in the game's own rules screen.
We just published the whole thing as an open dataset (CC BY 4.0): 1,333 games from 9 studios,
with RTP variants, bonus buy prices, themes and features as taxonomies, provider licences and
about 8,600 accessibility notes. This post is about what's in it and how it's captured.
How the data is captured
A slot demo is a canvas application: there's no DOM to parse, the paytable is drawn inside a
WebGL or Canvas2D context, and page.content() gives you an empty shell. So every game is opened
in its official provider demo in a headless browser, the rules, paytable and settings screens are
captured, and the values are read and verified with OCR.
Every record says how and when it was captured (capture_method, captured_at). And every record
says what we could not capture: the coverage_gaps table has a row per missing item. In this
release that's 5,762 gaps next to 2,502 explicitly confirmed items. A blank is a blank, not a guess.
That distinction ("this game has no bonus buy" vs "we failed to read whether it does") is the part
most datasets skip, and the part that makes this one usable.
What's interesting in the numbers
Most extra RTPs are per feature, not per operator. Across all 2,656 RTP rows: 1,353 are a game's
default, 824 belong to a bonus buy, 286 to an ante bet, 98 are operator configurations, 50 depend on
the player's own settings, and 45 on a feature. So when a game "has several RTPs", it's usually
because buying the bonus or switching on the ante bet plays at a different return, not because the
casino picked a worse version.
One game, thirteen rows. The Big Dog House lists 13 RTP rows: 96.43% for one bonus buy,
96.51% for an ante bet, 96.53% default, and so on. Small differences, but they're different numbers
and each one is real.
Rounding can cost more than the configuration. Some crash games are fine at a normal bet and not
at the minimum one. Crash X plays at 95.99% at a bet of 1.0 and "drops to 88.07% when betting 0.1:
wins are rounded down", to quote its own rules. That's 8 points, from rounding alone.
Accessibility is mostly about colour. The ~8,600 accessibility notes split into colour-blind
(3,674), motor (1,347), screen reader (1,250), hearing (1,226) and low vision (1,163). A typical
note: "No gameplay-critical information is conveyed by sound alone; all wins, triggers and jackpot
hits are also shown as on-screen text and numbers." I haven't found this kind of per-game review
anywhere else.
What's in the dataset
| table | one row per |
|---|---|
slots |
game: maths, grid, mechanics, themes, features, bonus economics, capture date |
rtp_variants |
RTP configuration, with its condition |
bonus_buys |
bonus buy option and its cost in bets |
accessibility |
accessibility note per game and group |
coverage_gaps |
captured / not captured item |
providers, licences
|
studio profile; licences with regulator and number |
themes, features
|
taxonomy with aliases (aztec also catches maya, inca) |
CSV and Parquet, refreshed monthly. Honest limits: 9 studios is not many. If you need RTP and
volatility for thousands of games, there are bigger open slot databases; this one is narrower
and deeper. Not included: screenshots, symbol art and game texts, which belong to the providers.
Getting it
- GitHub: https://github.com/igamingtools/igaming-slot-specs
- Hugging Face:
load_dataset("iGaming/igaming-slot-specs", "slots") - Kaggle, and Zenodo with a DOI: 10.5281/zenodo.24331735
The data comes out of iGaming Tools, a data API about the industry. It
does not run games or take bets. If you'd rather query it live, there's a free MCP server at
https://mcp.i-gaming.tools/mcp (no key, no sign-up) and a REST API with a free tier.
If you work with capture-and-verify data in another field, I'd like to hear how you handle coverage
accounting, i.e. telling "absent" apart from "not captured". That's the part I've seen the fewest
good answers to.
This article was written with the help of AI and checked by a person against the dataset before
publishing: every number above comes from the published CSVs.
Top comments (0)