The question we set out to answer

Are numerical patterns just the product of random probability, or a structure that can be logically derived?

The question is old but precise. If a draw is perfectly fair, every digit shows up with equal probability, and the past tells you nothing about the future. If it is not fair, if the draw machine carries even the smallest bias, then that bias should be detectable by statistics once you have enough history.

We do not predict, and we promise no winning method. We do what science can do. We state a hypothesis, test it against the data, and report the result honestly, wherever it lands.

Data & method

The data is the official published draw results from Singapore Pools, covering draws 2900 to 5507 (Sat, 29 Aug 2009 to Sat, 11 Jul 2026), 2,608 draws in total. Each draw contributes 23 winning numbers: the 1st, 2nd and 3rd prize, plus 10 starter and 10 consolation prizes. That comes to 59,984 four-digit numbers and 239,936 individual digits.

Our null hypothesis is that the draws are fair and independent, so every position is a uniform digit from 0 to 9. If that holds, each test below should read "consistent with random." Only a real, reproducible deviation would be worth noticing.

Test 1. Digit frequency

How often each digit 0 to 9 appears across all 239,936 positions. If the draw is fair, each should land near 23,994 (red line).

Chart: observed digit frequency against the expected value (red dashed line). By eye, every bar sits very close to the expected line.

Test 2. Chi-square test for uniformity

The chi-square test measures how far the real distribution sits from perfectly uniform. A p-value above 0.05 means the gap is within random fluctuation, which counts as consistent with random.

Why we don't chase patterns

Across this full 17-year record of 2,608 draws, every digit's deviation sits within the ordinary range of random fluctuation. The largest is digit 9 (standardised residual z = +2.05). When ten digits are tested at once, one of them just crossing |z| = 2 is exactly what chance produces. The pooled chi-square (p = 0.52) shows the overall distribution is flat.

Here is the most important lesson in the whole study. Earlier, using just three years of data, we saw what looked like a "significant" deviation in the digit "2" (z about −3.1, p about 0.003). Extend the data to 17 years and that same digit's deviation falls to z = −1.04, and the "pattern" simply disappears. On a small sample, noise disguises itself as a pattern. Give it enough data, or a different window, and the false pattern dissolves. A real pattern reproduces. Noise does not. Dressing noise up as a "discovery" is pseudoscience, so we don't.

Digit-sum distribution

The sum of the four digits of each number, from 0 to 36. If the draw is random, this should form a bell shape peaking near 18, which is what we see.

Chart: the digit-sum distribution forms a textbook bell curve, matching the theoretical expectation for random, independent digits.

Most & least frequent numbers

For interest only, not for prediction. Under a fair draw, a number that appeared often in the past is no more likely to appear in the future.

Across 2,608 draws, each of the 10,000 possible numbers appears only about 6.0 times on average, so the gap between "most" and "least" frequent sits within random expectation.

Test 3. Can history predict the future? (out-of-sample backtest)

This is the most decisive test. Instead of looking at distributions, we actually try to predict. We use a slice of history to pick 10 numbers to "bet" on the next draw, then count how many hit. It never peeks at the future.

The method is called walk-forward. For every draw, we train only on the history before it and pick 10 numbers. Strategy 1 buys the hottest past numbers. Strategy 2 buys the most frequent digit at each position. Then we count how many land in that draw's 23 winners, and vary the training-window length from 3 to 120 months to ask one question. Does more history make the prediction more accurate?

"Lift" is actual hits divided by the hits you would expect from random guessing. 1.0× means no better than guessing. Only a lift clearly above 1, and statistically significant, would mean a real, exploitable pattern.

The result is unambiguous. Whether you use 3 months or 10 years of history, and whichever strategy you use, the hit rate stays around the random-guessing level (lift about 1.0), and no window ever reliably beats chance. The only statistically significant result is a strategy performing worse than random, which only rejects predictability more firmly. In short, no amount of history helps you predict the next draw.

We tested the 四方财神 (Money Box) method

四方财神 studies the past ten years of data and offers the 4 most likely digits for each position, claiming those 4 are more likely to appear. We put that claim straight into 17 years of data.

Step 1. What does luck score? Each position is one digit from 0 to 9, ten in all, and the tool gives you 4 of them. In a fair draw the winning digit has a 4-in-10 chance of being among those 4, which is 40%. So 40% is the zero-skill score. To be real, the tool must clearly beat 40%.

Step 2. What did the tool actually score? Below are the last 12 draws, worked out. For each draw we used only the ten years before it, picked the 4 digits per position by the same logic, then checked whether the real winning digit was among them (green means a hit).

How to read it: take 11 Jul, 1st prize 6452. The column 1 winning digit is 6, the tool offered 0 1 6 9, which contains 6, so a hit. The column 2 winning digit is 4, the tool offered 1 5 7 9, which does not, so a miss. That draw scored 1 of 4.

Step 3. Add it all up. Across the full 17 years and 1,062 draws, the winning digit landed in the tool's 4 suggestions 40.1% of the time. Random chance is 40.0%. The gap is about one tenth of one percent, which is pure noise.

Step 4. Why? The tool's "AI" simply counts which digits came up most over ten years. But in a fair draw every digit comes up almost exactly 10% of the time. So the "top 4" are just four ordinary digits, each near 10%, adding up to about 40%. There is nothing to lock onto, so there is no edge to win.

Method note: every figure on this page is computed directly from official draw data, and the analysis is reproducible. This research is a mathematical and statistical exploration only. It offers no betting advice and promises no winnings. If you take part in betting, please do so responsibly and within your means.