Is This Baccarat Roadmap Data Actually Real? 7 Statistical Tests We Ran On Ourselves

1. How the dataset gave itself away

The dataset was evo_data.csv, a simple format: one shoe per row, second column a string of 0s and 1s (0 = Player, 1 = Banker). Its scale:

At first glance this is a very presentable dataset: large enough, with a Banker/Player split close to the theoretical value (about 49.3% Banker given the house edge). Run the most basic sanity check and it passes.

But after 7 more detailed tests, 6 of them pointed the same way: this was generated by code.

2. Tests 1-3: Structural regularity

Test 1: Geometricity of the run-length distribution

Split each shoe into runs of identical consecutive results. Under independence, run lengths follow a geometric distribution - but real data fluctuates. Length-1 runs at 50% and length-2 at 25% is only the trend; real proportions wobble by several percentage points.

This dataset gave:

Ratios of 0.4960 / 0.4988 / 0.4932 / 0.4931 / 0.4973 - all six share-to-share ratios land between 0.49 and 0.50. A real sample of 1.8 million hands does not achieve that consistency.

Test 2: Run-length transition matrix

Next, look at the transition matrix: given the previous run had length L, what is the next run's length? Real shoes show all sorts of structure. This dataset showed strict geometric decay in every single row:

L=1 -> [227072, 112618, 56521, 27994, 13581, 6845]
       ratios: 0.496   0.502   0.495   0.485   0.504

L=2 -> [112935, 55759, 27943, 13652, 6717, 3311]
       ratios: 0.494   0.501   0.489   0.492   0.493

L=3 -> [55901, 28145, 13984, 6855, 3370, 1722]
       ratios: 0.504   0.497   0.490   0.492   0.511

Six rows, all the same. That's not a property of card shoes. It's a property of a generator.

Test 3: Cross-shoe variance of window Banker share

This one we deliberately designed to be easy to pass. We measured how much the Banker share in the first 10 and last 10 hands of each shoe varies across all 32,634 shoes, and compared it to what pure randomness would produce.

Why include a passing test? Because if all 7 failed, you'd reasonably suspect we were reverse-engineering a conclusion. The honest version shows the passing test too - real data should show real randomness, and this one behaved exactly as expected, which validates the framework itself.

3. Tests 4-5: Autocorrelation and streak frequency

Test 4: Positional autocorrelation

Compute P(hand i = hand i+k). In real shoes this sits near 50%, drifting slightly with sampling noise. Across 13 different k values:

Several confidence intervals exclude 50%, and the direction is systematic (k=2,3,4 low; k=8,9 high). At a sample size of 1.8 million, consistency like this cannot be sampling noise - it reflects a structural bias baked into the data.

Test 5: Streak frequency vs the independent expectation

Under full independence, the expected number of runs of length L has a known value. Actual versus expected:

Ratios sit at 0.25x for L=1 through L=6, then jump to 0.447x at L=7 and collapse to 0.015x at L=8. A tidy tail that suddenly fractures like that is a data-processing artefact, not card shoes.

4. Test 6: The decisive proof - a direct independence violation

If the earlier tests are only "highly suspicious", this one is not defensible.

The logic is simple. If hands are independent, then after N consecutive identical results, the next hand should still be about 50% likely to continue. Seven Bankers in a row doesn't make the eighth more likely to be a Banker - or less likely. Each hand is an independent coin flip.

Measured result:

Incidentally, this also explains the absurd "anti-streak wins 95% of the time" result we produced earlier. Running a strategy backtest on synthetic data measures how strong the generator's segmentation logic is, not how accurate a strategy is. Treating such output as evidence about the game is treating a data artefact as a market pattern.

5. Test 7: Position effects and shoe length distribution

Position effects within a shoe

If the data came from real shoes, the closer to the shuffle the composition gets more unusual and card counting becomes more valuable. Real data should show strategy performance differing between the start and end of a shoe.

Measured: continuation rate after 5 in a row, across the four quarters of a shoe, was 0.000% / 0.000% / 0.000% / 0.000%. Identical start to finish - consistent with "generated by a fixed rule", inconsistent with "harvested from real shoes".

Shoe length distribution

Shoe lengths run from 50 to 70 hands in a smooth bell shape (2,230 shoes at 50, 1,793 at 60, 5 at 70). Nothing obviously wrong here, but nothing that carries the signature of a real shoe either.

6. A 5-step audit workflow for any dataset

If you do any analysis on historical roadmaps - your own, someone else's, or bundled with a tool - run these in order. Every step is a plain statistical test. No AI, no toolkit, one Python script.

7. Why this matters as much as picking the right software

Our previous article covered how to choose baccarat prediction software (6 criteria). This one adds another, more fundamental one: the tool itself may be fine while the data it runs on is fake.

Returning to the previous article's core argument, with a correction to the wording. We said:

Those claims aren't wrong. But there was one thing we didn't do: verify the data was real in the first place. If the data is synthetic, then "win rates drop 8-15 points" could itself be a generator artefact - you cannot run a robustness test in a fake world.

So the correct order is:

1. Verify the data - 7 tests, 5 minutes
2. Then verify the method - does the backtest hold across disjoint windows?
3. Only then the tool - latency, privacy, free tier come last

Get the order wrong and you end up doing due diligence on a precisely calculated hallucination.