Texas Two Step: 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.

🤖 AI lab (neural networks) ›

Backtest: Quick Pick lines vs hot lines (12 months)

For each of the 105 draws from 23.10.2025 to 22.10.2026, 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.

MatchesQuick PickHotExpected by chance
0623578631.0
1352384360.6
2748055.9
3182.5
4000.0

Wins by tier

TierMatchesQuick PickHotExpected by chance
14 matches + Bonus ball000.00
24 matches000.02
33 matches + Bonus ball000.07
43 matches182.42
52 matches + Bonus ball101.60
61 match + Bonus ball9910.30
70 match + Bonus ball171418.03

Result: the two distributions differ over this period (chi² = 8.76, p = 0.033). Across many repeated tests, a gap like this turns up from time to time; it is not proof of a better method, as every combination remains equally likely.

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 11.12.2023 to 22.10.2026.

Matches per prediction

ObservedExpected
0 Matches175 / 180.3
1 Matches106 / 103.0
2 Matches17 / 16.0
3 Matches2 / 0.7
4 Matches0 / 0.0

Average: 0.487 matches per prediction – chance gives 0.457.

Deviation from the expectation of chance: chi² = 0.56, p = 0.757. The model hits exactly as often as numbers picked at random.

Model line for the next draw

19222935

For entertainment only: this line has exactly the same chance as any other.

Latest predictions

DatePredictionMatches
22.10.2026 2171933 0
19.10.2026 4101429 0
15.10.2026 4203435 0
12.10.2026 14182932 0
08.10.2026 1101925 1
05.10.2026 12212326 1
01.10.2026 121929 0
28.09.2026 4172032 0
24.09.2026 6101218 0
21.09.2026 14182125 0
17.09.2026 262534 0
14.09.2026 1192022 0
10.09.2026 5202234 0
07.09.2026 13283335 1
03.09.2026 13212632 0
31.08.2026 2212633 1
27.08.2026 3111727 1
24.08.2026 1141829 1
20.08.2026 1121934 1
17.08.2026 3141723 0
13.08.2026 4131732 0
10.08.2026 35630 1
03.08.2026 231216 0
30.07.2026 271014 0
27.07.2026 13182223 1
23.07.2026 4142230 1
20.07.2026 8162230 0
16.07.2026 492526 0
13.07.2026 5222932 0
09.07.2026 21232526 0
06.07.2026 13242632 1
02.07.2026 12172729 1
29.06.2026 492126 1
25.06.2026 1142125 0
22.06.2026 691129 0
18.06.2026 682025 0
15.06.2026 1222829 0
11.06.2026 182729 0
08.06.2026 13632 0

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 4 from 35, 0.5 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.
▲