The data behind every figure

Nothing on this site is computed while you wait, and nothing is quoted from another site. Every number comes out of one of the 37 datasets below, generated offline by a script and then checked against something outside this project before it shipped. 32 of them are exact rather than simulated.

They are here so the figures can be checked. Any claim to be right is worth what the working behind it is worth, and the working should be visible.

Exact and simulated are not the same claim. A dataset marked exact carries no sampling error at all — only the modelling assumptions stated alongside its tool. A simulated one comes from Monte Carlo runs and its low-order digits will move if it is regenerated. Only the poker figures are simulated; everything else is counted.

Poker

  • Preflop hand equity

    Simulated · 91 kB

    Win, tie and loss rates for all 169 starting hand classes, at every table size from 2 to 9 players, against random opponents played to showdown.

    Checked against: Every heads-up figure agrees with exact enumeration: rebuilt from the 47,086 exactly counted matchups behind the hand-vs-hand page, all 169 hands' win and tie rates against a random hand land within sampling error of these (worst 2.9 standard errors). AA heads-up near the commonly cited 85% and 72o near 32% are re-checked by npm run test:poker-math on every run.

    Download JSONGenerated 21 Oct 2026
  • Hand vs hand, exact

    Exact · 7 kB

    All-in preflop equity for every pair of the 169 starting-hand classes, from 47,086 distinct matchups each enumerated over all 1,712,304 boards, with the spread across suits and the deals at each end of it. One file per hand ships with the page; the summary carries the derived facts.

    Checked against: All 26 matchups in Deuces Cracked's exact tables reproduce to the two decimals printed, as do two of the three extremes in Timo Denk's; the third does not, and is recorded rather than tuned to. Equity against a random hand rebuilt from the matchups agrees with the site's independent Monte Carlo figures for all 169 hands.

    Download JSONGenerated 22 Oct 2026
  • Heads-up push/fold

    Exact · 142 kB

    The heads-up push/fold equilibrium at every stack from 1 to 20 big blinds in half-blind steps, no antes, chip EV: the small blind’s shove frequency and the big blind’s call frequency for each of the 169 classes, with each hand’s EV of shoving and of calling. Solved by fictitious play over the exact heads-up equities, with card removal.

    Checked against: Matches HoldemResources' published hand lists at 7, 10, 15 and 20bb to within four borderline classes, each within a tenth of a big blind of indifferent, and PokerLevelUp's shove and call shares at all six depths it prints to within half a point. Every solve ends within 0.005bb of an exact equilibrium for both players.

    Download JSONGenerated 22 Oct 2026
  • Set mining

    Exact · 35 kB

    For each pocket pair from 22 to KK against each higher pair and AK, AQ, AJ and KQ suited and offsuit: how often the flop gives you a set or better with both hands known, your share of the pot from that flop to the river, split by whether their hand improved on the same flop (a higher set, against a pair). Every flop and every turn and river counted; every opposing combo weighted equally. The page prices set over set separately and weights a mixed range by combos.

    Checked against: With only your two cards known, the flop odds reproduce Deuces Cracked's printed set or better (11.755%), full house (0.735%) and quads (0.245%), and PokerStrategy's 11.8% and 7.5 to 1. The flop-by-flop counts for a matchup are exactly ten times the whole-board counts of the engine behind hand vs hand, which itself reproduces Deuces Cracked's 26 exact matchups.

    Download JSONGenerated 30 Oct 2026
  • Hand-type frequency (realistic folding)

    Simulated · 1 kB

    How often each hand category — pair through straight flush — actually reaches showdown at a table, under a model where players fold weak holdings rather than calling to the river.

    Checked against: Checked against the standard five-card hand frequencies, which the no-fold variant below should approach and this one deliberately should not.

    Download JSONGenerated 19 Sept 2026
  • Hand-type frequency (everyone plays to the river)

    Simulated · 1 kB

    The same categories under the simpler assumption that nobody folds. Published together with the realistic model because the gap between them is the interesting part.

    Checked against: Approaches the standard published five-card hand frequencies, which is the check: with no folding, showdown hands are just random seven-card draws.

    Download JSONGenerated 15 Aug 2026
  • Positional profitability

    Simulated · 2 kB

    Simulated win rate by seat in big blinds per 100 hands, at 6-max and 9-max, under four models: flat versus positional entry ranges, each with and without a postflop in-position equity layer. The point is the decomposition, not one number.

    Checked against: Seats sum to zero by construction, which is the arithmetic check. The interesting result is that the flat model shows almost no positional gradient at all — the button only pulls away once postflop in-position equity is layered in, which is the claim the dataset exists to test. The blinds lose in every variant.

    Download JSONGenerated 21 Oct 2026

Blackjack

  • Basic strategy

    Exact · 289 kB

    The expected-value-maximising action for every player hand against every dealer upcard, across several rule sets, from a recursive EV engine rather than a copied chart.

    Checked against: Spot-checked against the textbook chart: 5,5 doubles rather than splits, 8,8 always splits, 9,9 stands against a 7, hard 12 hits 2-3 and stands 4-6. Late surrender was checked the same way.

    Download JSONGenerated 21 Sept 2026
  • House edge by rule set

    Exact · 834 B

    What the house keeps under each combination of dealer, doubling, splitting and payout rules, and what each individual rule costs you.

    Checked against: Each rule’s isolated effect is checked against Wizard of Odds’ published rule-variations table, to within 0.05 points: dealer-hits-soft-17 +0.22, 6:5 payouts +1.39, no-DAS +0.14, late surrender −0.07, no-hole-card +0.11. An early version read 3.2% overall because dealer blackjack was double-counted — exactly the error this check exists to catch.

    Download JSONGenerated 21 Sept 2026
  • Third base, and what that card would have done

    Exact · 3 kB

    What the card a third-base player takes would really have done to the dealer, per up-card, plus what a seat costs per hour by how many people are at the table.

    Checked against: Built on a dealer recursion that removes each card from the shoe for real, rather than the frozen-composition model behind the rest of the site — a different model, and it reproduces the published six-deck dealer bust table to 0.004 percentage points, which is an independent check rather than a restatement. The expectation question itself is settled by exchangeability rather than measured: burning one card before the dealer draws moves her outcome distribution by 1.7e-16, zero to floating point.

    Download JSONGenerated 21 Sept 2026
  • Side bets

    Exact · 34 kB

    House edge, hit rate and full combination breakdown for Perfect Pairs, 21+3, Royal Match and Lucky Ladies — every published pay table, at one through eight decks.

    Checked against: All 58 published house edges reproduced to within 0.005 percentage points, across four bets and twelve pay tables, against Wizard of Odds. The raw combination counts match too: 144 suited king-queens and 11,868 other suited pairs in a six-deck shoe. Every enumeration is separately required to sum to C(52 × decks, n), which is what proves no hand is double-counted — the trap here being that a suited three of a kind is simultaneously a three of a kind and a flush.

    Download JSONGenerated 21 Sept 2026
  • Variance and session outcomes

    Exact · 16 kB

    The exact payoff distribution of a single hand, the covariance between two hands played at once, and the exact distribution of where a session finishes over a range of hand counts.

    Checked against: Checked against Wizard of Odds’ variance tables, which are themselves simulations of about ten billion hands per rule set. Three rule sets are compared rather than one: variance agrees to within 0.018 and covariance to within 0.004, and every one of ours sits just below its published counterpart — a residual with a consistent sign, explained by this engine never resplitting a pair. Per-hand standard deviation across one, two and three simultaneous hands lands within 0.008 of the published figures.

    Download JSONGenerated 21 Sept 2026
  • What each mistake costs

    Exact · 8 kB

    The house-edge cost of every common blackjack mistake, from playing like the dealer down to each individual contested decision, with the rarity of the decision already priced in.

    Checked against: Anchored on an identity rather than a published number, because the identity verifies itself: a player who mimics the dealer with a natural paying even money has no advantage left, so the entire house edge must be the probability that both hands bust. Those two figures are computed independently here and agree to 0.0008 points. The commonly published mimic-the-dealer figure of 5.48% is carried alongside for comparison.

    Download JSONGenerated 21 Sept 2026
  • Effect of removal and count indexes

    Exact · 12 kB

    What removing one card of each rank does to the house edge, how closely Hi-Lo’s three whole-number tags track those effects, and the true count at which each contested hand changes its correct play.

    Checked against: Effect of removal correlates at 0.998 with Wizard of Odds’ six-deck table and 0.999 with Griffin’s single-deck table in The Theory of Blackjack — two independent sources at two deck counts. The index numbers were checked against Schlesinger’s Illustrious 18: 14 of 17 playing indexes match exactly and all 17 land within one, with the insurance index of +3 falling out of pure arithmetic about a single fraction.

    Download JSONGenerated 21 Sept 2026
  • Bust odds

    Exact · 3 kB

    The probability of going over 21 if you take one more card on each hard total, and the probability the dealer eventually busts showing each upcard under every rule variant.

    Checked against: The player figures are checkable by hand: with 13 equally likely ranks, a hard 12 busts on a 10, J, Q or K, so 4/13 = 30.77%, and each total up adds one busting rank. The dealer figures need the full recursion, and peak against a 6 as every published table shows.

    Download JSONGenerated 21 Sept 2026
  • Never-bust strategy

    Exact · 4 kB

    The house edge of never taking a card on a hard 12 or more and playing everything else by basic strategy, split by the dealer’s up-card and by stiff total, and under each of six rule sets.

    Checked against: Costs 3.52 points over basic strategy here against 3.48 in Wizard of Odds’ analysis of the same strategy, defined the same way, under slightly different rules. The ten per-up-card costs sum to the whole exactly, which is asserted, since the dealer shows one card a round.

    Download JSONGenerated 4 Nov 2026

Yahtzee

  • Category odds

    Exact · 2 kB

    Expected score and completion probability for all 13 categories under optimal single-category rerolling, plus the full score distribution and opening-roll value tiers behind each.

    Checked against: P(Yahtzee) over a turn comes out at 4.60%. First-roll probabilities match exact combinatorial counts out of 6^5 = 7,776 derived independently: 6 yahtzees, 240 large straights, 300 full houses.

    Download JSONGenerated 16 Sept 2026
  • Final-score distribution

    Simulated · 16 kB

    How a full game actually finishes under the solved optimal policy: the distribution of final scores, percentiles, and the expected contribution and zero-rate of every category. The policy is exact and only the dice are sampled.

    Checked against: Reproduces all seventeen figures Verhoeff publishes for the Optimal Solitaire Yahtzee Player -- grand total, standard deviation, median, and an expected contribution plus zero-rate for each of the thirteen categories and both bonuses. The mean also has to reproduce the exactly solved root EV of the table it plays from, which no amount of sampling error can hide.

    Download JSONGenerated 31 Oct 2026
  • Category detail

    Exact · 14 kB

    The per-category breakdown: expected score with one, two and three rolls available, the exact distribution of every final score, and how many genuinely distinct positions the opening roll can leave you in.

    Checked against: Each distribution sums to 1 and its mean reproduces the expected score computed by a separate recursion that never builds a distribution at all.

    Download JSONGenerated 16 Sept 2026
  • Full-game optimal expected value

    Exact · 8.2 MB

    The expected value of the rest of a game from each of 536,448 reachable scorecard states, solved by exact backward induction over about 21 minutes.

    Checked against: Optimal play from an empty card averages 254.5896, the figure Verhoeff (1999) publishes for his Optimal Solitaire Yahtzee Player, to all four decimals, under the same rules. Position by position too: at eight positions taken from his player, from a first roll to a joker with the bonus banked, all 184 options it ranks are valued within 0.01 of his figures, and the best play agrees at every one.

    Not offered as a file: 8.2 MB, and it already loads with the full-game tool that uses it.Generated 31 Oct 2026
  • Keep-or-reroll decision table

    Exact · 1.1 MB

    For all 252 distinct five-dice rolls, both reroll counts and all 13 categories: which dice to keep, what that is worth, and the next-best alternatives.

    Checked against: Derived from the same engine as the category odds, whose outputs are checked against independent closed forms.

    Not offered as a file: 1.1 MB, and it ships with the keep-or-reroll guide.Generated 16 Sept 2026

Monopoly

  • Square landing odds

    Exact · 6 kB

    Long-run share of turns a token ends on each of the 40 squares, and in Jail, as the stationary distribution of a Markov chain over dice, doubles, Jail and both card decks. This is where a token rests between turns; how often each square is arrived on, which is what rent is paid on, is in the property ROI dataset.

    Checked against: Reproduces the two most-cited published facts about this problem on the first correct run: Jail is the most-landed-on square by a wide margin, and Illinois Avenue is the top non-jail square. The board and decks were cross-checked against three independent sources.

    Download JSONGenerated 16 Sept 2026
  • Property return on investment

    Exact · 13 kB

    What each colour group costs to buy and build, what it earns per opponent turn at every development level, and how many turns it takes to pay for itself; the same for railroads and utilities by how many are owned; and how often each of the 40 squares is arrived on per turn, the rate rent is paid at.

    Checked against: Rents and prices come from the same cross-checked property data as the square odds. The payback ordering contradicts the folk belief that Boardwalk is the best buy, which is the point.

    Download JSONGenerated 31 Oct 2026

Scrabble

  • Tile distribution and draw odds

    Exact · 9 kB

    The standard English-language 100-tile set with, for every tile, the chance it reaches an opening rack and the full distribution of how many land there.

    Checked against: The distribution was taken from two independent published datasets and cross-checked letter by letter; they agree on every count and face value. Every single-copy tile comes out at exactly 7%, which needs no hypergeometric to verify.

    Download JSONGenerated 16 Sept 2026
  • Opening rack odds

    Exact · 5 kB

    The exact face value of an opening rack across every total it can take, the vowel and blank splits, and the probability of the racks players actually complain about.

    Checked against: The rack-value DP mean equals the closed form 7 x 187/100 = 13.09, and the DP never computes a mean. 187 is the published face-value total of a full set.

    Download JSONGenerated 17 Sept 2026
  • Bingo odds

    Exact · 4 kB

    How often an opening rack can play all seven tiles, overall and conditioned on holding each tile, measured against the public-domain ENABLE lexicon.

    Checked against: Bessoir & Pines (G4G13, 2018) get 12.92% against OWL2014 over the same 16,007,560,800 racks; ENABLE has 920 fewer seven-letter words so ours has to sit just below, and it does. The walk also asserts its own weights sum to exactly C(100,7).

    Download JSONGenerated 17 Sept 2026
  • Two- and three-letter words

    Exact · 31 kB

    Every two-letter word in ENABLE with the letters that hook it into a three-letter word front and back, the exact odds an opening rack can play it, its reverse where that is also a word, and its face value — plus the full three-letter list.

    Checked against: ENABLE's 96 two-letter words reconcile exactly against NWL2023's 107, counted from a published enumeration: 96 + the 11 ENABLE lacks = 107, with nothing in ENABLE that NWL does not have. Every hook is round-tripped against the three-letter list, and AA's odds reproduce P(a rack holds two As) from the separately-computed tile distribution.

    Download JSONGenerated 21 Oct 2026

Tennis

  • Tennis scoring by format

    Exact · 21 kB

    Match-win probability and expected match length for six formats at two serve strengths and five skill gaps, the better player’s chance in a tiebreak against a whole set, the tail of an advantage final set, and the deuce loop across every serve strength from 50% to 80%.

    Checked against: A symmetric matchup comes out at exactly one half at game, tiebreak, set and match level in all six formats. The deuce closed form p² / (p² + (1 − p)²) is rebuilt from a plain point-by-point walk to 1e-12, and a game between equals is checked at its exact 6.75-point average length. No published dataset exists to match figure for figure, so the checks are identities that cannot be tuned.

    Download JSONGenerated 29 Sept 2026
  • Most points you can win and still lose

    Exact · 2 kB

    The largest share of total points a player can win while losing, for each of six formats, with the scoreline that achieves it.

    Checked against: Solved exactly by Dinkelbach’s method over every game, tiebreak and set shape, and reproduced independently by hand: 106 of 168 points and a 6-7 6-0 6-7 loss in a best-of-three, 171 of 264 in a best-of-five. Both match OEIS A368009 and A368008 exactly. For best-of-five with a 10-point final-set tiebreak the solve beats OEIS A368146 (158 of 244 against its 174 of 270), by losing the decider 4-6 rather than in the tiebreak. The optimal lost game is asserted to be 2-4, with no deuce.

    Download JSONGenerated 29 Sept 2026
  • Point importance

    Exact · 20 kB

    How much every score in a service game swings the chance of holding, for serve strengths from 50% to 85%; the chance of facing no break point in a game, a set and a match; and match-level importance at three break points under Wimbledon’s rules of the 1990s.

    Checked against: Checked against Klaassen and Magnus (JASA, 2001), who model the server’s drop on a point as proportional to its match importance and print enough figures to work back to the importance each implies. At all six scores they quote, the importance computed here sits inside the range their rounding allows. Also asserted: 30-40’s importance equals the deuce win probability exactly, Morris’s game × set × match decomposition holds to 1e-12, and the no-break-point closed form is rebuilt from a plain point walk.

    Download JSONGenerated 29 Sept 2026

Paradoxes

  • Secret Santa

    Exact · 9 kB

    For groups of 2 to 30: the chance someone draws their own name, the expected number of redraws, and how often a working draw is one loop or contains a swapped pair.

    Checked against: Derangement counts match OEIS A000166 to twenty terms and both of Euler’s recurrences; the no-swap counts match OEIS A038205 to sixteen. The chance converges to 1 − 1/e, asserted.

    Download JSONGenerated 14 Oct 2026
  • The Monty Hall problem

    Exact · 2 kB

    Stick and switch for the classic game, a host who opens doors at random, a host with a door preference, more doors, and the four-door two-round game.

    Checked against: Enumerated case by case in exact fractions. Reproduces Wolfram MathWorld (2/3, and 2/8, 3/8, 6/8, 5/8 for the four-door plans) and Rosenthal (Math Horizons, 2008): 1/2 for the random host, and 1/(1 + p) and 1/(2 − p) for the host with a preference.

    Download JSONGenerated 14 Oct 2026
  • Penney’s game

    Exact · 3 kB

    The chance each three-flip pattern turns up before each other one, every pattern’s best reply, and every pattern’s average wait.

    Checked against: Every pairing computed twice, by Conway’s overlap numbers and by an exact Markov-chain solve, which must agree. Reproduces every row of Vallin’s table of best replies (Recreational Mathematics Magazine, 2020) and the rule that generates them.

    Download JSONGenerated 14 Oct 2026
  • Coin-flip streaks

    Exact · 12 kB

    The full distribution of the longest run of one face, and of the number of switches between faces, in 20, 50, 100 and 200 fair flips.

    Checked against: Exact counts over all sequences. Reproduces all four of Schilling’s figures for 250 tosses (Math Horizons, 1994): head runs of six, seven and eight in 87%, 63% and 38% of sequences, and runs of eight of either face in 63%. His either-face shortcut is checked against a direct count.

    Download JSONGenerated 14 Oct 2026
  • The 100 prisoners

    Exact · 6 kB

    The success rate of the loop strategy and of random guessing for every even group from 2 to 100.

    Checked against: Exact fractions. Reproduces Curtin and Warshauer (Mathematical Intelligencer, 2006) — 893/2520 for 10 players, 0.311828 for 100 — and Wolfram MathWorld to ten places; the rate falls with group size and never below 1 − ln 2, asserted at every size.

    Download JSONGenerated 14 Oct 2026
  • Riffle shuffles

    Exact · 1 kB

    Distance from random after 1 to 12 riffle shuffles of 52 cards, 52! in full, and the rising-sequence counts behind the explanation.

    Checked against: Computed exactly from Bayer and Diaconis’s Theorem 1 over all 52 Eulerian numbers, and reproduces their Table 1 (Annals of Applied Probability, 1992) to the three decimals printed. A shuffled deck’s mean of 26.5 rising sequences is checked exactly.

    Download JSONGenerated 14 Oct 2026
  • The birthday problem

    Exact · 39 kB

    The chance of a shared birthday for every group of 2 to 100 people, of three sharing one, and of someone sharing yours; the group sizes for an even chance of two to five sharing a day; and the same figures on real US births by date, with every date’s births against an average day’s.

    Checked against: Reproduces Wolfram MathWorld digit for digit: the exact reduced fraction for 23 people, and 23, 88, 187 and 313 as the group sizes for an even chance of two, three, four and five sharing a day, all from whole-number counts. The real-calendar figures must sit above an even 366-day calendar at every group size, which is a theorem, and the Social Security Administration’s 2000–2014 count agrees with the CDC’s 1994–2003 count to within 0.01 percentage points at every group size, an asserted bound.

    Download JSONGenerated 13 Oct 2026

Terms

These files are published so you can verify the figures on this site, not as a dataset to build on: they remain © Odds The Law, all rights reserved, and no licence to redistribute them is granted. If you want to use one for something, ask — the answer is likely yes.

If you quote a figure, please link the page you took it from. Two inputs are not ours and are credited where they are used: the Scrabble bingo figures are computed against the public-domain ENABLE lexicon, and board layouts, tile distributions and card decks are facts about commercially published games, cross-checked against independent public sources. Trademarks belong to their owners and are used descriptively.

For how each engine works rather than what it produced, see the methodology.

By Gabe · Updated · How every figure is checked

Cookies for analytics and ads; decline and both stop. Privacy