Data Investigation · July 2026

The Lottery Predictor Everyone Asks For

I asked Claude, Anthropic's AI, to train a model on every Powerball drawing since 2015 — all 1,374 of them. It learned exactly one thing: there is nothing to learn.

Directed & published by Jeffry Arevalo Code, models & writing by Claude (Anthropic's AI) 11 min read Data through 06/29/2026
A single white lottery ball casting a long shadow on paper, trailed by a wandering red line through a field of scattered grey dots.

Every data scientist eventually gets the question. A friend, a coworker, an uncle at Thanksgiving: "You work with data. Can't you just predict the lottery?" The usual answer is a polite no and a subject change. I decided to give the question the full treatment instead: real data, a real train/test split, a real backtest — the same methodology any forecasting problem would get. One disclosure up front: I directed this experiment and set the questions, and Claude, Anthropic's AI, did the hands-on work — the code, the model training, the backtests, and the writing — following my instructions. If the predictor was going to fail, I wanted it to fail on the record, with receipts.

So here is the honest version of the project everyone imagines. At my direction, Claude pulled every drawing under Powerball's current format, trained models on 80% of history, tested them against the most recent 20%, and then let the final model call its shot on a live drawing. All of it is below, including the part where the model faceplants at 292-million-to-1 odds. The faceplant is the finding.

The short version

  1. Eleven years of drawings are statistically indistinguishable from a perfectly fair random draw. There is no bias to learn.
  2. On 275 unseen drawings, trained models performed identically to a random quick pick. Not slightly worse or better. Identically.
  3. The only real "strategy" in Powerball has nothing to do with the machine, and it argues against everything the model learned.

The Data

First, know your data-generating process

Powerball is not one dataset. It is seven. The game has changed its number matrix seven times since 1992, and each change creates a new statistical universe.[2][3] Mixing draws from different formats would poison any analysis, because a "59" drawn from a 59-ball pool is not the same event as a "59" drawn from a 69-ball pool.

In effectFormatJackpot odds
Apr 19925 of 45 + 1 of 451 in 54,979,155
Nov 19975 of 49 + 1 of 421 in 80,089,128
Oct 20025 of 53 + 1 of 421 in 120,526,770
Aug 20055 of 55 + 1 of 421 in 146,107,962
Jan 20095 of 59 + 1 of 391 in 195,249,054
Jan 20125 of 59 + 1 of 35 · $2 ticket1 in 175,223,510
Oct 20155 of 69 + 1 of 26 · current1 in 292,201,338

The October 2015 change is the one that matters here: the white-ball pool grew from 59 to 69 and the red Powerball pool shrank from 35 to 26, stretching jackpot odds to 1 in 292,201,338.[2][3] A third weekly drawing was added on Mondays in 2021, which produces more data but does not touch the process that generates it.[4]

That process, by the way, is not software. Drawings use physical, gravity-mixed ball machines at the Florida Lottery's studio in Tallahassee, supervised by an independent CPA firm, with certified ball sets.[3] This matters for modeling: if there were anything to exploit, it would have to be a physical bias, like a heavy ball or a worn machine. That gives us a testable hypothesis.

The dataset is therefore every drawing under the current format: 1,374 drawings, October 7, 2015 through June 29, 2026.[5][6]

The Method

An 80/20 split, no peeking

The split is chronological, the way any honest forecast backtest works: the first 1,099 drawings (October 2015 to September 2024) are the training set; the last 275 drawings (September 2024 to June 2026) are the holdout. The model never sees the future while learning from the past.

What does "training" even mean here? With no features to engineer, no weather, no momentum, no fundamentals, a model can only learn the frequency structure of the numbers themselves. So that is what Claude built: a hot-numbers model that always plays the five most-drawn white balls and the most-drawn red ball, and a frequency-weighted sampler that draws its picks in proportion to historical appearance rates. As a control, a pure random quick pick with no training at all. If the lottery is fair, the control should tie the trained models. If the machines have a physical bias, training should win.

The Evidence

What eleven years of drawings actually look like

Before testing, the training data itself gets a physical exam. Across 1,374 drawings, each of the 69 white balls should appear about 100 times if the game is fair. Here is every ball:

6080100120140 Ball 1: drawn 93 timesBall 2: drawn 98 timesBall 3: drawn 110 timesBall 4: drawn 94 timesBall 5: drawn 89 timesBall 6: drawn 104 timesBall 7: drawn 94 timesBall 8: drawn 92 timesBall 9: drawn 91 timesBall 10: drawn 92 timesBall 11: drawn 98 timesBall 12: drawn 103 timesBall 13: drawn 75 times#13 · 75×Ball 14: drawn 91 timesBall 15: drawn 93 timesBall 16: drawn 109 timesBall 17: drawn 96 timesBall 18: drawn 102 timesBall 19: drawn 102 timesBall 20: drawn 100 timesBall 21: drawn 123 times#21 · 123×Ball 22: drawn 90 timesBall 23: drawn 116 timesBall 24: drawn 98 timesBall 25: drawn 87 timesBall 26: drawn 79 timesBall 27: drawn 116 timesBall 28: drawn 119 timesBall 29: drawn 89 timesBall 30: drawn 99 timesBall 31: drawn 99 timesBall 32: drawn 116 timesBall 33: drawn 115 timesBall 34: drawn 84 timesBall 35: drawn 91 timesBall 36: drawn 114 timesBall 37: drawn 109 timesBall 38: drawn 93 timesBall 39: drawn 107 timesBall 40: drawn 100 timesBall 41: drawn 90 timesBall 42: drawn 95 timesBall 43: drawn 98 timesBall 44: drawn 107 timesBall 45: drawn 99 timesBall 46: drawn 81 timesBall 47: drawn 111 timesBall 48: drawn 89 timesBall 49: drawn 81 timesBall 50: drawn 97 timesBall 51: drawn 90 timesBall 52: drawn 109 timesBall 53: drawn 112 timesBall 54: drawn 95 timesBall 55: drawn 90 timesBall 56: drawn 98 timesBall 57: drawn 96 timesBall 58: drawn 93 timesBall 59: drawn 108 timesBall 60: drawn 94 timesBall 61: drawn 123 timesBall 62: drawn 112 timesBall 63: drawn 114 timesBall 64: drawn 121 timesBall 65: drawn 88 timesBall 66: drawn 100 timesBall 67: drawn 100 timesBall 68: drawn 96 timesBall 69: drawn 113 times 110203040506069 95% chance band (81 to 118) expected: 99.6 white ball (1 to 69) · appearances since Oct 2015 · hover a dot for its count
Appearances of each white ball across all 1,374 drawings under the current matrix. The shaded band is the 95% range pure chance would produce. 63 of 69 balls sit inside it, and a handful of stragglers is itself exactly what a fair process predicts. Ball #21 (123×) runs hottest and #13 (75×) coldest: the same "hot/cold" gap that randomness manufactures for free.

The spread you see is noise wearing a costume. A chi-square test of uniformity on the training draws returns p = 0.20 for the white balls and p = 0.71 for the red ball, statistically indistinguishable from a perfectly fair draw. This is the entire "hot numbers" industry refuted in one test: the gaps between hot and cold balls are precisely the size that randomness guarantees, and they carry zero information about the next drawing.

The Test

275 drawings the models never saw

Still, maybe the holdout set knows something the chi-square missed. Each model predicted every one of the 275 unseen drawings. Under pure chance, a five-number ticket should match about 0.362 white balls per draw. Here is where everyone landed:

0.300.320.340.360.380.400.42 pure chance: 0.362 Hot numbers (trained)Hot numbers (trained): 0.360 white balls matched per draw0.360Frequency-weighted (trained)Frequency-weighted (trained): 0.367 white balls matched per draw0.367Random quick pick (untrained)Random quick pick (untrained): 0.345 white balls matched per draw0.345 avg white balls matched per draw · 275 held-out drawings
Nine years of training data bought the models a performance edge of zero. The differences between all three strategies are well inside sampling noise.
0
times any model matched even 4 of 5 white balls, in 825 model-drawing attempts
1
time a trained model matched 3 whites: a $7 ticket, once in 20 months
3.3 to 3.6%
red-ball hit rate for every strategy, vs. 3.85% pure chance

The trained models did not merely fail to beat the random control. They were indistinguishable from it, which is the stronger and stranger result. Powerball drawings are independent events: the machine holds no memory of Tuesday when it runs on Wednesday. Whatever the past 1,099 drawings contain, it is not a signal.

The Math

The impossible loop

My original test protocol had one more rule: do not advance to the next holdout drawing until the model predicts the current one exactly, all five whites plus the Powerball. It sounds like rigor. It is actually a proof by contradiction, because the probability of an exact match is fixed at:

1 / 292,201,338
per attempt, and no amount of history improves it

Since draws are independent, the model's odds on retry five million are identical to its odds on retry one. We gave the trained model five million attempts at just the first test drawing. Exact matches: zero, precisely as expectation dictates. Here is what those five million attempts look like against the cost of one expected jackpot match:

1 million attempts (given to the model)1 million attempts (given to the model)1 million attempts (given to the model)1 million attempts (given to the model)1 million attempts (given to the model)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)1 million attempts (still required)
Each square is one million attempts. The five brick squares are what the trained model got at a single drawing (it matched nothing). The full grid of 292 is the average cost of one exact match. Clearing all 275 test drawings in sequence would take about 80 billion attempts, or roughly 2,500 years at one guess per second.

Any product, script, or seller claiming to have passed a "predict each draw before advancing" test is describing a bug or a lie. The loop does not terminate. That is not a limitation of this model; it is a property of the game.

The Edge

The one edge that actually exists

Here is the twist the frequency crowd never mentions: number selection does have a strategic dimension, just not against the machine. It is against other players.

You can't change your odds of winning. You can change your odds of sharing.

Millions of players bet birthdays (1 through 31), lucky sevens, and published hot numbers. Play those and a jackpot win is more likely to be split three ways. The only mathematically defensible "strategy" in Powerball is to play numbers other people avoid, which is, delightfully, an argument for the random quick pick and against everything this model learned.

A crumpled lottery ticket on a wooden table beside a cup of coffee in soft morning light.
A $2 purchase of a few days of daydreaming: a fair trade priced as entertainment, never as an investment.

The Receipt

Epilogue: the model calls its shot

Analysis is cheap; predictions are receipts. On June 30, with all 1,374 drawings in hand, the model produced its highest-conviction pick for the July 1, 2026 drawing: the five hottest white balls in history and the hottest red. Then the balls dropped in Tallahassee.[7]

Model pick212832616414
July 1 drawing262639686

Matches: zero. Not one ball. Eleven years of training data, outdrawn by gravity.

That is the finding, and it is worth stating without hedging: Powerball cannot be predicted. Not with frequency analysis, not with machine learning, not with more data. The dataset is not hiding a pattern that a bigger model would find; the chi-square test shows there is no pattern to hide. Every combination in every draw carries identical 292-million-to-1 odds. A lottery ticket is a $2 purchase of a few days of daydreaming, and that is a fair trade only if you price it as entertainment, never as an investment.

The real lesson is not about lotteries. It is that a rigorous negative result, a model that provably cannot work, is worth more than a sloppy positive one. Half of data science is knowing when the answer is "there is no signal here," and saying it plainly.

Sources & references

  1. Powerball / Multi-State Lottery Association. Official game rules, odds, and drawing procedures. powerball.com
  2. Louisiana Lottery Corporation. "Powerball's Matrix Changes to Build Bigger Jackpots." Announcement of the October 2015 format change (69 white balls, 26 Powerballs). louisianalottery.com
  3. Wikipedia. "Powerball." Matrix history, jackpot odds of 1 in 292,201,338, and drawing operations at the Florida Lottery studio under independent CPA supervision. en.wikipedia.org/wiki/Powerball
  4. Lotto America. "History of the Powerball Lottery." Timeline including the 2021 addition of Monday drawings and the Double Play feature. lottoamerica.com/powerball/history
  5. USA Mega. Powerball statistics for the current matrix, confirming the count of drawings since October 7, 2015. usamega.com/powerball/statistics
  6. Draw results dataset. Machine-readable Powerball results compiled from official drawings, updated after each draw (jbaranski/jeffs-lottery-utils). github.com/jbaranski/jeffs-lottery-utils
  7. Democrat and Chronicle (via AOL). "Powerball winning numbers for Wednesday, July 1, 2026": 2, 6, 26, 39, 68, Powerball 6. aol.com