We tested baccarat pattern reading.
It is real, it is statistically significant, and it still loses. Then we went looking for where the edge actually is.
The short version, in dollars
There is a genuinely profitable bet here. Whether it is reachable is a different question, and the answer is mostly no.
- about $18 a shoe
- what the edge is worth at the $100 table minimum, with perfect counting. Real money, and roughly minimum wage for an hour of unbroken concentration.
- about $200,000
- the bankroll needed before betting the $10,000 maximum is sensibly sized rather than reckless. That is where the headline numbers come from.
- 450 shoes
- roughly a year of weekends before you could tell whether you were winning or merely lucky.
- negative
- what the whole thing becomes if you miss even one card in twenty. The mistakes all push the same way, so you would never feel it going wrong.
Words we could not avoid
- house edge
- The share of every dollar you bet that the casino keeps on average. At 1.06% on Banker, a $100 bet costs you about a dollar — every time, forever.
- a shoe
- The box the dealer deals from. Here it holds six shuffled decks — 312 cards — and it is dealt down to near the end before being reshuffled.
- the big road
- The grid of red and blue marks on the screen at every baccarat table, recording who won each hand. It is the thing players read patterns into.
- bankroll
- The money you have set aside to play with — not the size of one bet, but everything you are prepared to put at risk before you stop.
- expected value
- What a bet is worth on average if you could make it thousands of times. Negative expected value means the average outcome is a loss, however the next hand lands.
- +EV
- A bet that makes money on average. Most bets are the opposite: you expect to lose a little every time, which is the house edge.
- percentage point
- A flat difference between two percentages. Going from a 1.2% cost to a 0.7% cost is half a percentage point, not half a percent.
- |z| and t
- Two names for the same thing: how many times bigger a result is than the wobble you would expect from luck alone. Under about 2 it is ordinary; over about 3 it starts to be worth believing.
- p = 0.015
- The odds this result is a fluke. Roughly one in sixty-six — low enough to call it real, which is not the same as calling it useful.
- the null
- What the same search produces on deliberately scrambled data. If your finding cannot beat that, you found nothing.
What this was computed on
Every number on this page is specific to these rules. Change the deck count, the cut, or a payout and the arithmetic moves — which is the whole point of the piece.
The shoe
- Six decks — 312 cards
- Shuffled by hand; no continuous shuffler
- Dealt deep — close to the whole shoe goes out before the cut
Drawing
- Either side showing 8 or 9 on two cards is a natural and ends the hand
- Player draws on 0–5, stands on 6–7
- If the Player stands, the Banker draws on 0–5 and stands on 6–7
- If the Player drew, the Banker follows the standard drawing table on its total and the Player's third card
Payouts
- Player 1:1 · Tie 8:1
- Banker 1:1 less 5% commission (the commission game)
- Banker even money except a win on 6, which pays 1:2 (the no-commission game)
- Small 1.5:1 · Big 0.5:1 · Player & Banker Pair 11:1
- Any Pair 5:1 · Perfect Pair 25:1 · Super 6 12:1
Limits
- Main bets and Small/Big: 100 to 10,000
- Pairs and Dragons: 25 to 500
- The 20× gap between those two lines is why Small matters and the pair bets do not
The Dragon Bonus ladder is assumed standard — 30:1 on a nine-point win down to even money at four, with naturals paying 1:1 and a natural tie pushing. That is the only payout schedule on this page we did not read directly, so treat those two rows as indicative.
Findings one to three are robust to the rules. They concern outcome sequences and shoe composition, and hold for any baccarat dealt from a shoe.
Findings four and five are not. They depend on the Small payout and on how deep the cut card sits — move the cut one deck forward and most of that edge is gone.
Finding one
The patterns are real. They are also 2.3× too small to bet.#
- non-tie hands
- 13,789,368
- patterns searched
- 79
- best cell
- last5 BBPBP
- its shift
- -0.292 pp
- after best-of-79 correction
- p = 0.015
We dealt a quarter of a million six-deck shoes and read the big road the way a player does — ties skipped — then measured whether the sequence predicted the next hand. The grid held every exact streak from one to eight, every chop, and every binary pattern of the last three, four and five outcomes.
Something is there. The best cell shifts the Banker rate by 0.292 percentage points and survives correction for having searched 79 of them, at p = 0.015. That is not noise, and it should not be surprising: cards leave the shoe and do not come back, so a small amount of order information has to exist. The physics requires it.
It is still not enough to bet on. The Player wager breaks even when it wins half the non-tie hands, which needs a shift of 0.684 points. The best pattern in thirteen million hands delivers 0.292. Betting it takes the house edge from 1.237% down to 0.709% and never crosses zero — you lose more slowly, on 2.8% of hands, and that is the whole prize.
The null matters here. We shuffled each shoe's own outcomes, preserving its exact composition and destroying only the order, and re-derived every pattern from the shuffled sequence. Across 200 replicates the best of 79 cells reached |z| = 2.37 at the median and 3.27 at the 95th percentile. A t of 2.5 is what searching noise this wide looks like.
Finding two
The derived roads cannot help, and this is arithmetic rather than a result.#
- grid depth
- complete at 5
- possible last-5 patterns
- 32, all tested
Four boards hang above a baccarat table and only one of them records what happened. The bead plate is that one: a cell per hand in the order they came, ties included, filling downward and then across. The big road is the history: a column for each streak, red for Banker and blue for Player, a new column every time the side changes. When a streak runs past six rows it turns right along the bottom and keeps going — that sideways tail is the dragon people mean when they tell you to follow one.
The other three boards are not history. They are computed from the big road by fixed rules, and each asks the same question from further away. Big eye boy asks, of every new entry: did what just happened match what happened one column earlier? Red if it did, blue if it did not. The small road asks it looking two columns back, and the cockroach three. That is the entire difference between them.
The mechanism, exactly: when a result starts a new column, the road compares the depths of the two columns behind it and marks red if they match. When a result extends a column downward, it looks back the same distance to see whether that earlier column had already reached this row, and marks red if it had. Published sources disagree about which cell each road is allowed to start on; that moves the first mark and nothing else.
Dragon means two unrelated things at the same table, which is worth untangling before you bet on the wrong one. The dragon on the board is the tail above. The Dragon Bonus on the felt is a side bet on your side winning by a wide margin, and it is priced: 2.716% on Player against 9.560% on Banker — the same bet on the other hand, three and a half times worse.
Big eye boy, the small road and the cockroach road are computed from the big road by fixed rules. They are re-encodings of history you already have. No function of a variable can carry information the variable does not — so if the big road does not predict, nothing derived from it can. That is not a claim we tested; it is a claim that does not need testing. We implemented all three to be certain rather than to argue it: across 3,972 distinct big roads, two shoes that produced the same big road produced the same three derived roads every time.
The pattern grid makes the same point for anything else at that depth. Every possible pattern of the last five outcomes is one of 32 cells, and any predictor built on that history is just a rule for which cells to bet. Its expected value is a weighted average of those cells and cannot exceed the best one. None reaches break-even, so no function of the last five outcomes can profit — not a sample of systems, all of them.
Finding three
With perfect knowledge of every card left, there is still nothing there.#
- shoe states enumerated
- 700
- Player +EV
- 2.71% of states
- average edge when positive
- +0.675%
- states with 45+ cards left that are +EV
- 0 of 198
Rather than try more patterns, we computed the ceiling. For each of 700 shoe states we enumerated every reachable deal from the true remaining composition and got the exact next-hand probabilities. Nothing — no pattern, road, count or model — can beat this, because nothing can know more than what is left in the shoe.
The main bets do turn positive, and only at the very end: 25% of the time with fourteen cards remaining, 1.9% with thirty to forty-four, and not once in 198 states with forty-five or more cards left. For roughly 95% of the shoe the ceiling itself sits below zero. There is no information to find because there is no information there.
This reproduces Thorp and Walden's 1966 result on a modern six-deck game. It has been the answer for sixty years.
Finding four
The edge is in the bet nobody looks at.#
- Small pays
- 1.5:1
- it needs
- P(4 cards) > 0.4000
- full shoe gives
- 0.37893
- deep-shoe range
- 0.1455 to 0.6732
- +EV on
- 10.9% of all hands
The Small side bet pays 1.5 to 1 if the hand finishes on four cards, so it needs that to happen more than 40% of the time. A fresh shoe gives 37.893% — a gap of barely two points. And whether a third card gets drawn depends heavily on what is left, so as the shoe depletes that probability swings from 14.6% to 67.3%.
It crosses the line often. Across 1,200 fully-dealt shoes the true probability beat break-even on 10.9% of all hands, at an average edge of 5.25%. Unlike the pair bets — which only come alive in the last fifty cards — this is live with 155 cards still to come, half the shoe away.
We did not trust that number. The probability function was validated on six shoes with known answers (a shoe of only eights and nines must give 1.0; only aces, sixes or tens must give 0.0), cross-checked against a separate enumeration, and then tested end to end by actually playing it: 7,293 bets, 42.55% hit rate against a 40% break-even, +6.369% realised against +5.226% predicted. The control is the part that convinces — flat-betting the same wager every hand returned −5.538% against a theoretical −5.269%.
Finding five
And it breaks on ordinary human error, in one direction.#
- optimal count
- two ranks: 8s and 9s
- Hi-Lo tags
- ten of thirteen ranks
- with a ±0.5 deck estimate
- 72% of the edge gone
- missing 5% of 8s and 9s
- negative
The effect of removal is concentrated almost entirely in eights and nines — every other card behaves the same as every other card. So the optimal system is not ten counts, it is count the eights and nines: two of thirteen ranks, against Hi-Lo's ten. Simpler than blackjack, and it recovers around 90% of the theoretical maximum.
Then it falls apart. Estimate the remaining decks to within half a deck — normal human performance — and most of the edge is gone. Miss five percent of the eights and nines and the whole thing goes negative.
The mechanism is what makes it dangerous. Every eight or nine you fail to count makes you believe more of them remain, so you overestimate and bet when you should not. In Hi-Lo, missed cards roughly cancel; here every error pushes the same way. The failure is systematic, not noisy, which is why a small miss rate does not shrink the edge — it flips the sign. You would be betting large, confidently, and losing.
The bankroll closes the argument. The swing on this bet is sixteen times the edge; it takes roughly 450 shoes to be two standard errors clear of zero, and a $10,000 wager is only correctly sized against a bankroll near $200,000. At a $100 table minimum the same edge is worth about eighteen dollars a shoe.
Every bet on the table, priced exactly
Six decks, by full enumeration rather than simulation — these are not estimates. The commission and no-commission games look nearly identical and disagree about which bet is correct.
| Banker — commission game (0.95:1) | 1.056% |
| Player — either game (1:1) | 1.237% |
| Banker — no-commission / Super 6 | 1.455% |
| Player Dragon | 2.716% |
| Small (1.5:1) | 5.269% |
| Big (0.5:1) | 6.839% |
| Banker Dragon | 9.560% |
| Player / Banker Pair (11:1) | 11.254% |
| Tie (8:1) | 14.438% |
| Any Pair (5:1) | 14.536% |
| Perfect Pair (25:1) | 17.071% |
| Super 6 side bet (12:1) | 30.002% |
What to do at the table
Every number here is read straight off the priced table further down this page, which was produced by full enumeration of a six-deck shoe rather than simulation.
scripts/make-playbooks.mjs · derived from this study’s own priced table
Before you sit down
Two decisions before a card is dealt. One of them is the most expensive thing on this page.
While you are playing
There is one bet worth making and a lot of expensive scenery around it.
The one real edge
It exists, it is large, and this is what it actually asks of you.
In the commission game Banker is correct at 1.056%. In the no-commission game the same bet costs 1.455% and Player becomes the better wager. The two tables look almost identical and nothing on the felt tells you which one you are at.
Player Dragon is 3.5× better than Banker Dragon — the same side bet on the other hand.
The bet a no-commission table is named after carries a 30% house edge.
None of this is advice to go and play. The one wager with a real edge requires conditions most tables do not offer, execution most people cannot sustain, and a bankroll most people do not have — and it turns negative in the hands of anyone slightly sloppy, without telling them. We published it because the question gets asked constantly and nobody answers it with numbers, in either direction.