UK Lotto: AI prediction and backtest – what are "statistical" lines worth?
Two honest experiments on the real draws: we compare random Quick Pick lines with "hot" lines built from the most frequent numbers, and we let a learning algorithm (k nearest neighbours) predict each draw, always using only the data known before the draw. Both methods are compared with the exact expectation of pure chance.
Backtest: Quick Pick lines vs hot lines (12 months)
For each of the 105 draws from 18.10.2025 to 17.10.2026, each group plays 10 lines: random Quick Pick versus lines built from the 15 most drawn numbers over the previous 100 draws. The Bonus ball are random in both groups so that only the main numbers are compared. 1,050 lines per group in total.
| Matches | Quick Pick | Hot | Expected by chance |
|---|---|---|---|
| 0 | 520 | 557 | 535.0 |
| 1 | 421 | 372 | 401.2 |
| 2 | 97 | 108 | 102.4 |
| 3 | 11 | 12 | 10.9 |
| 4 | 1 | 1 | 0.5 |
| 5 | 0 | 0 | 0.0 |
| 6 | 0 | 0 | 0.0 |
| Winning lines | 109 | 121 | 113.8 |
Wins by tier
| Tier | Matches | Quick Pick | Hot | Expected by chance |
|---|---|---|---|---|
| 1 | 6 matches | 0 | 0 | 0.00 |
| 2 | 5 matches + Bonus ball | 0 | 0 | 0.00 |
| 3 | 5 matches | 0 | 0 | 0.01 |
| 4 | 4 matches | 1 | 1 | 0.48 |
| 5 | 3 matches | 11 | 12 | 10.92 |
| 6 | 2 matches | 97 | 108 | 102.36 |
Result: the difference between the two groups is not statistically significant (chi² = 4.93, p = 0.177). It stays within what pure chance produces for a sample of this size.
AI prediction (k nearest neighbours)
For each draw, the model looks for the 7 past moments whose preceding 3 draws most resembled the current ones and plays the numbers most often drawn next. It is reapplied after every draw and never sees the future. 300 draws evaluated from 06.12.2023 to 17.10.2026.
Matches per prediction
Average: 0.603 matches per prediction – chance gives 0.610.
Deviation from the expectation of chance: chi² = 1.70, p = 0.428. The model hits exactly as often as numbers picked at random.
Model line for the next draw
91026334554
For entertainment only: this line has exactly the same chance as any other.
Latest predictions
| Date | Prediction | Matches |
|---|---|---|
| 17.10.2026 | 132030353659 | 0 |
| 14.10.2026 | 8912313842 | 0 |
| 10.10.2026 | 61028525557 | 0 |
| 07.10.2026 | 122529414246 | 1 |
| 03.10.2026 | 91231364553 | 1 |
| 30.09.2026 | 41418202343 | 0 |
| 26.09.2026 | 223645475456 | 2 |
| 23.09.2026 | 71228343741 | 2 |
| 19.09.2026 | 22728323942 | 0 |
| 16.09.2026 | 173136414551 | 3 |
| 12.09.2026 | 142223325456 | 1 |
| 09.09.2026 | 1911154356 | 0 |
| 05.09.2026 | 31314344250 | 0 |
| 02.09.2026 | 131719274459 | 0 |
| 29.08.2026 | 222438414245 | 3 |
| 26.08.2026 | 81335374054 | 2 |
| 22.08.2026 | 121322395556 | 0 |
| 19.08.2026 | 141727314351 | 1 |
| 15.08.2026 | 104349545556 | 0 |
| 12.08.2026 | 151842464851 | 1 |
| 08.08.2026 | 51832404752 | 1 |
| 05.08.2026 | 2422323343 | 0 |
| 01.08.2026 | 21022374045 | 0 |
| 29.07.2026 | 252831354749 | 1 |
| 25.07.2026 | 5929405157 | 1 |
| 22.07.2026 | 121415162556 | 2 |
| 18.07.2026 | 91220274345 | 0 |
| 15.07.2026 | 21422354352 | 1 |
| 11.07.2026 | 121426343553 | 0 |
| 08.07.2026 | 41433365659 | 0 |
| 04.07.2026 | 3711232640 | 0 |
| 01.07.2026 | 63339405759 | 2 |
| 27.06.2026 | 4812343942 | 1 |
| 24.06.2026 | 111214182333 | 2 |
| 20.06.2026 | 151832353656 | 0 |
| 17.06.2026 | 71827284858 | 0 |
| 13.06.2026 | 71015203854 | 2 |
| 10.06.2026 | 249335657 | 1 |
| 06.06.2026 | 23637424547 | 1 |
| 03.06.2026 | 21213202738 | 1 |
How it works
A nearest-neighbour model looks for points in time whose last draws resembled the current ones and picks the numbers that followed most often. It is reapplied draw by draw without looking into the future; a backtest additionally compares Quick Picks with lines made of frequent numbers.
Example
For 6 from 59, 0.6 matches per line are expected on average; the model lands exactly there.
Limitations
No model can predict random draws. The page shows this transparently with a chi-squared test.
Frequently asked questions about this tool
- Can AI predict lottery numbers?
- No. With independent, uniformly distributed draws there is no pattern a model could learn.