
A baccarat simulator is a program that deals baccarat without a casino: it shuffles a virtual shoe, applies the drawing rules and records who won. People use the word for two different things. One is a practice table where you place virtual chips and play hand by hand. The other is a batch simulator that deals hundreds of thousands of hands in seconds, so you can test a betting idea before it costs money. This page is about the second kind, with a link to the first further down. We wrote a simulator, ran it for 100,000 eight-deck shoes — 8,027,824 hands — and publish the output here next to the exact figures, so you can see what a correct simulator produces, how many hands it takes before the output means anything, and what happened to seven common strategies inside it.
Baccarat has no player decisions after the bet, which makes it one of the easiest casino games to simulate — and one of the easiest to simulate slightly wrong without noticing. These are the parts that decide whether the numbers can be trusted:
| Part | Correct behaviour | Common mistake |
|---|---|---|
| Shoe | 8 decks, 416 cards, dealt without replacement until the cut card | Drawing every card at random from an endless deck |
| Burn | First card is turned over and that many cards are discarded (face cards count 10) | Skipped — harmless for the odds, but the shoe length comes out wrong |
| Deal order | Player, Banker, Player, Banker | Two cards to one side, then two to the other — changes nothing in the totals, but breaks any card-level check |
| Naturals | An 8 or 9 on either side ends the hand with no draw | Letting the other side still draw |
| Player's third card | Draws on 0–5, stands on 6–7 | Rarely wrong |
| Banker's third card | Depends on the Banker total and the Player's third card (the 3-4-5-6 table) | "Banker draws on 0–5" — the single most common bug |
| Payouts | Banker win pays 0.95, Player win pays 1, a tie returns both bets | Paying Banker 1 to 1, or counting ties as losses |
| Random numbers | A full-quality generator and an unbiased shuffle | A short home-made generator, or a shuffle that favours some orders |
The Banker rule deserves the warning. If the Banker simply draws on 0–5 whenever the Player has drawn, the Banker's advantage shrinks and every result downstream is off. The full table is on the third-card rule, and the shoe procedure — burn, cut card, number of hands — on the baccarat shoe.
Baccarat's probabilities are known exactly, because every possible ordering of the cards can be counted. So a simulator has an answer key. Run it long enough and it should land on the right-hand column below. Ours did:
| Quantity | Our simulator (8,027,824 hands) | Exact value, 8 decks |
|---|---|---|
| Banker wins | 45.852% | 45.8597% |
| Player wins | 44.620% | 44.6247% |
| Tie | 9.528% | 9.5156% |
| Banker share of decided hands | 50.680% | 50.6825% |
| A natural on either side | 34.30% | 34.305% |
| Player takes a third card | 50.35% | 50.341% |
| Banker takes a third card | 43.56% | 43.541% |
| Player pair (first two cards same rank) | 7.471% | 7.4699% |
| Banker wins with a total of 6 | 5.383% | 5.3864% |
| Return on a flat Banker bet | −1.061% | −1.0579% |
| Return on a flat Player bet | −1.231% | −1.2351% |
The two rows that expose bugs fastest are "Banker takes a third card" and "Banker share of decided hands". A simulator with the simplified Banker rule cannot hit 43.54% and 50.68% at the same time. If you are testing a simulator you did not write, check those two before you believe anything else it says. Where the figures on the right come from is laid out on baccarat odds.
Shoe-level output is a second check. Our shoes averaged 80.3 hands (shortest 73, longest 87) with the cut card 14 cards from the end, 7.65 ties per shoe, and a longest streak averaging 6.52 hands. A run of six or more appeared in 69.4% of shoes and seven or more in 42.9% — streaks that feel remarkable at the table are routine output from a shuffled shoe with no memory.
This is the part most simulator screenshots leave out. Below is the same run, read at six points along the way. The bracketed figure is the 95% range you should expect from chance alone at that sample size.
| Hands simulated | Banker win rate (expected range) | Flat Banker return (expected range) |
|---|---|---|
| 100 | 46.0% (±9.8 points) | +3.70% (±18.2 points) |
| 1,000 | 44.8% (±3.1) | −1.94% (±5.8) |
| 10,000 | 46.23% (±0.98) | −0.34% (±1.8) |
| 100,000 | 45.74% (±0.31) | −1.22% (±0.57) |
| 1,000,000 | 45.77% (±0.10) | −1.24% (±0.18) |
| 8,027,824 | 45.85% (±0.03) | −1.061% (±0.06) |
After 100 hands the simulator showed Banker betting winning 3.7%. After 10,000 hands — about 125 shoes — the Banker bet looked three times cheaper than it is. After a full million hands the measured return was −1.24%, which happens to be the Player bet's edge, not the Banker's. Only at eight million hands did the result settle within a few hundredths of a point of the true −1.058%.
The practical rule: the uncertainty on a flat-bet return is about 1.82 divided by the square root of the number of hands, in units per unit bet. To measure a return to within one percentage point you need roughly 33,000 hands; to within half a point, about 132,000; to within a tenth of a point, 3.3 million. The difference between Banker and Player is 0.18 points, so a simulation of less than a few million hands cannot even tell those two bets apart. Any "strategy tested in a simulator over 500 hands" is reporting noise.
We then replayed the same 8,027,824 hands under seven betting rules. Ties are pushes and the bet carries over. "Return" is profit divided by total money staked.
| Strategy | Hit rate on decided hands | Units staked | Net result | Return |
|---|---|---|---|---|
| Always Banker, flat 1 unit | 50.680% | 8,027,824 | −85,205 | −1.061% |
| Always Player, flat | 49.320% | 8,027,824 | −98,839 | −1.231% |
| Follow the last result, flat | 50.024% | 8,027,824 | −89,844 | −1.119% |
| Bet against the last result (chop), flat | 49.976% | 8,027,824 | −94,200 | −1.173% |
| Bet against streaks of 5 or more only | 49.898% | 504,526 | −6,307 | −1.250% |
| Banker with a 7-step Martingale | 50.680% | 27,567,553 | −318,668 | −1.156% |
| Player with a 7-step Martingale | 49.320% | 29,169,859 | −369,184 | −1.266% |
Three things stand out. First, the pattern rules — follow, chop, break the streak — all finish between the Banker and Player edges, because all they do is choose which of two slightly losing bets to hold on each hand. The streak-breaker waited for a run of five, which happened on 6.3% of hands, and then won 49.9% of the time: a long streak told it nothing about the next card. Second, the Martingale did not change the return per unit staked; it changed how many units were staked. Doubling after losses pushed the average bet to 3.43 units, so the Banker Martingale lost 3.7 times as much as flat Banker betting over the same hands. Third, the progression's bust was not rare bad luck: 26,520 of 3,707,399 cycles ran to seven straight losses — one in 140 — each costing 127 units.
The same strategies measured on real casino shoes instead of simulated ones are on baccarat pattern strategy, and the exact Martingale tables on baccarat martingale. The simulated and the real results agree, which is the point: a correctly built simulator and a properly run table are the same machine.
Long-run percentages are not how a session feels. So we cut the run into consecutive sessions of flat 1-unit Banker bets and looked at where each one ended.
| Session length | Sessions simulated | Average result | Finished ahead | Middle 95% of sessions | Worst / best |
|---|---|---|---|---|---|
| 100 hands | 80,278 | −1.06 units | about 45% | −19 to +17 | −40 / +39 |
| 1,000 hands | 8,027 | −10.6 units | 35.5% | −67 to +47 | −119 / +104 |
| 10,000 hands | 802 | −106 units | 14.2% | −287 to +78 | −406 / +242 |
A hundred hands is close to a coin flip with a slight lean: nearly half of all sessions end in profit, and the swing of ±18 units is seventeen times the average loss of one unit. That is why short experience teaches the wrong lesson. Stretch the session and the lean takes over — one in three 1,000-hand runs still ends ahead, but only about one in eight 10,000-hand runs is expected to. Our figure of 14.2% for the last row comes from just 802 sessions, which carries its own error of about ±2.3 points around the expected 12.7%; even a session table needs a sample-size note. How to use these ranges when you set a stake is covered in baccarat strategy.
Ours is about a hundred lines of Python and deals 8 million hands in under half a minute on an ordinary PC. The structure is the same in any language:
Betting rules are then a second pass over the recorded string of results, which makes it cheap to try many ideas on identical hands. One caution from our own testing: try enough rules on the same simulated hands and some will show a profit by chance. Re-run any winner on a fresh seed before believing it.
| You want to… | Use | What it can and cannot tell you |
|---|---|---|
| Learn the flow of a hand and the payouts | A play-money practice table | Teaches rules and pace; a few hundred hands say nothing about any strategy |
| See whether a betting system holds up | A batch simulator, 1 million hands or more | Gives the return per unit staked and how often the system busts |
| Check a claim about streaks or patterns | A batch simulator, or a large archive of real shoes | Shows how often the pattern appears by chance and what follows it |
| Feel what variance does to a bankroll | Either — play 100 hands several times over | Each replay ends somewhere different; that spread is the lesson |
For the first kind, our free baccarat game deals from an 8-deck shoe with the full third-card rules and 10,000 virtual chips, no sign-up. If you would rather record real shoes and count for yourself, the free Baccarat Scoreboard app keeps the roads offline, and the live baccarat terminal shows real tables forming the same streak and tie counts the simulator produces.
Is a baccarat simulator accurate? It is if it reproduces the known probabilities: Banker 45.86%, Player 44.62%, Tie 9.52% over several million hands, with the Banker drawing a third card on 43.54% of hands. A simulator that misses those has a rule or shuffle error.
Can a simulator find a winning baccarat strategy? It can show you that none of the common ones is. In our 8-million-hand run every rule, flat or progressive, returned between −1.06% and −1.27% of the money staked. A simulator is a tool for ruling ideas out cheaply.
How many hands should I simulate? At least a million to compare betting systems, and several million to see differences of a tenth of a percentage point. Below 30,000 hands the return is uncertain by more than a full point either way.
Why did my strategy win in the simulator? Usually a short run, a rule tuned on the same hands it was tested on, or a bug that pays Banker 1 to 1. Fix the payout, change the random seed and run it again on ten times as many hands.
Does a simulator behave like a real live table? For results, yes. Real shoes we have logged show the same Banker, Player and Tie rates within sampling error. What a simulator leaves out is speed, table limits and how you behave when the money is real.
Is practising on a free simulator useful? For learning the rules, the commission and the pace, yes. It will not train you to predict hands, because there is nothing in a shuffled shoe to predict from the scoreboard.
Simulation: Baccarat Terminal, 12 November 2026 — 100,000 eight-deck shoes (8,027,824 hands), burn applied, shoe ended when 14 or fewer cards remained, standard punto banco drawing rules, 5% Banker commission, fixed random seed. Exact values: full enumeration of an 8-deck shoe. Expected ranges use a normal approximation. Simulated results describe the game's mathematics and do not predict any future hand. 18+. Information only — not a gambling service and not betting advice.
