Lotto Max: the four lines, and why the odds never move

What to play, and why

The four lines are below, in one play for $6.00. They are the model's own output, and they lean away from the numbers players crowd into.

• The model was measured, and it cannot predict. Across 973 draws it had never seen, its best strategy beat a random ticket by +0.0115 numbers per draw with a bootstrap p of 1.00 — noise, and the page prints that instead of dressing it up.
• So the ticket is built for the one thing that does change the arithmetic. A jackpot is split between everyone who matches it, and players crowd into dates and tidy shapes. These four lines lean away from both, which does not make a win likelier — it makes one worth more.
• The full working is on this game's own page: the backtest, the control, the ball and machine tests, and the one line the model ranks highest.
• And the same caution applies here. These four lines are one play. The model does not raise the chance of a match, nothing on this site claims it does, and the honest lever is how much you spend, not which numbers you mark.

The numbers for draw 1274 — Tuesday 29 September 2026

Four lines — one play, $6.00 — covering the model's twenty-eight strongest numbers with no number used twice.

1. Line 118283234384252
2. Line 204071015354851
3. Line 308142124263741
4. Line 405173640444750

The model's seven highest-scoring numbers, as a line on its own: 07 18 28 32 42 51 52. It is not one of the four above, and that is the dispersion rule working rather than an oversight. Line 1 already holds 5 of its seven and puts 34 and 38 in place of 07 and 51, because those four lines exist to spread across twenty-eight numbers: a number sitting beside one already taken is pushed down hard, and a number above 31 is nudged up, since low numbers look like birthdays and get shared more often. Here 07 was overtaken by the small nudge that favours numbers above 31; 51 sits next to 52 on Line 1. Nothing is thrown away — 07 on Line 2, 51 on Line 2.

Copy the four lines

What the backtest says

Features for draw t are computed from draws before t and nothing else, so no answer leaks backwards. The model is trained on the past, asked to pick the next draw's seven numbers, and scored on how many it got. It is retrained every forty draws, and scored over 973 draws it had never seen.

A shuffled-target control is run beside it: the same pipeline with the answers permuted. If the measurement were leaking, that control would still appear to beat random. It does not.

And a planted-signal check, because "no edge" only means something if the test could have found one: a feature equal to the answer half the time was added over 48 draws. A working measurement must report about 3.991 hits per draw. It reported 3.714.

Then the ablation, which is the only honest way to answer "did the extra detail help". Two blocks of features — one from the draw history, one from the conditions on the night — are run separately and then together. If the weather carried anything, the second block should lift the models. It does not — and all three models land on the same number when given only that block, which means it barely distinguishes one number from another at all.

One more figure is needed to read any of this honestly: the smallest edge this many draws could have seen. Over 973 out-of-sample draws the standard error is 0.0282 hits per draw, so an edge smaller than 0.0790 — about 8.0% of the base rate — could have been present and been missed. That is the resolution of the whole test, and it is why "no edge" here means "no edge larger than that".

And the simplest comparison of all: an arbitrary fixed set of seven numbers, chosen once and never changed, scores 1.0401 over the same draws. That is better than every model on this page, including the one with the most features.

Can the draw be influenced?

Lotto Max is not a ball draw, and has not been since 14 May 2019. Atlantic Lottery retired the machines that produced Lotto Max and Lotto 6/49 on that date and moved both to a random number generator — draw 504 in this history, the same draw at which the field of numbers went from 49 to 50. So 770 of the 1,273 draws here were produced by a computer program, including the one you are playing, and only 503 came from a machine, the last on 10 May 2019.

That splits the question rather than closing it. While there was a machine a draw really was a physical event: solid natural rubber balls about fifty millimetres across, tossed by counter-rotating arms inside an acrylic chamber. Manufacturers weigh every ball to a microscopic tolerance, X-ray them for uniform internal density, retire a set if a ball loses a fraction of a gram, and require that no bare hand touch one. The stated reason for the gloves is moisture, which could alter a ball's weight. That is a mechanism, so it can be tested rather than argued about — on the 503 draws where it existed.

Both eras were tested separately, with the same scan, on 503 machine draws and 770 generator draws. The machine era is the only one where any of this could act, and 0 of its 180 tests came out significant after correction, smallest corrected probability 0.35. The generator era gives 0 of 180, smallest 0.52. In both, fewer tests crossed the uncorrected 5% line than chance alone would produce.

The honest limit is power, not physics. The machine era is 503 draws, and at that size the scan resolves only a 12% shift in how a set is shaped; the generator era at 770 resolves 10%. So the machine era can be cleared of any large sensitivity to the room, and not of a small one. Two independent mechanisms came out clean, which is a stronger statement than one — but a small effect would still be sitting under the noise floor.

So the conditions were collected for every draw night since 2009: 30 environmental variables — temperature, humidity, air density, pressure and how fast it was swinging, wind, cloud and rain, plus the sunspot number, the solar radio flux and the geomagnetic index. Humidity is carried as absolute humidity, the mass of water in a cubic metre, because vapour pressure is what a ball would soak up; and it is carried at four trailing timescales, because rubber does not equilibrate in an hour.

There are only four ways any of that could reach the numbers, and each was measured separately, on the whole history first. Either the draw leaves a mark on the shape of a set, or the weather changes which number comes out, or a whole night is different from an average night, or the data contains regimes — abrupt shifts, which is what rotating the ball set would produce. A mixed history weakens a test of the machine, so the same scan was then run on each mechanism alone.

A count of zero can hide a weak test, so the resolution is worth stating. Of the 180 shape tests, 0 came out significant after correcting for how many were run, and the smallest corrected probability was 0.09. The smallest shift in a set's sum this could have detected at all is 2.8 balls, and a set's sum varies by about 36 balls across the whole history. Of the 1,500 ball-by-ball tests, 52 came out significant at the raw 5% level — where chance alone produces about 75, so fewer than chance — and 0 survived correction.

The night-level test is the strongest of the four and it is worth spelling out, because it needs no weather data at all. On a MaxMillions night the draw is run many times over: one game is the main draw and every other set is a MaxMillions set, put through the same mechanism minutes apart. If anything about that night pushed the outcome — heat, moisture, a worn ball set, or simply a badly seeded generator — both would be pushed the same way. Across 496 such nights there is no agreement beyond chance (p = 0.93), and splitting the nights into the third with the most moisture in the air and the third with the least changes nothing (p = 0.72 and p = 0.58). That is a direct measurement of whether a night differs from an average night, and it says no — and it keeps working after 14 May 2019, when the mechanism stopped being a machine.

Every MaxMillions set is also seven more numbers put through the same mechanism, which multiplies the evidence about the mechanism itself. Including them, that mechanism has been run 61,117 times. Against that sample, no number comes up more or less often than it should: a goodness-of-fit test gives p = 0.97, the most extreme single ball sits at 2.28 standard deviations where chance alone produces values that large, and after correcting for testing fifty-two numbers at once the smallest surviving probability is 0.97. A number drawn today also tells you nothing about tomorrow (correlation +0.0170). With this many draws the test would see a skew of about 1.2% in the probability of a single ball.

One thing found while doing this, recorded because it nearly went the other way. The first run of the ball-by-ball test reported 212 significant pairs out of 1,550, which looks like a discovery. A control was run beside it: the same test against a variable with no relationship to anything, its values shuffled. It scored 48 of 50 balls significant — impossible. The fault was in the standard error, which was mis-scaled by a factor equal to the spread of each variable, so any wide-ranging variable manufactured false significance and any narrow one was suppressed. Corrected, the shuffled control scores zero and so does every real variable. Without the control the wrong answer would have been published here.

And even a real regime would not be usable. The ball set, and while there was one the machine, is chosen at random immediately before each draw, so which set is in play is unknown at the moment a ticket has to be bought. A bias you cannot know you are facing cannot be bet on.

Why zero significant is the right answer and not a failed test. Temperature, humidity, pressure and gravity are all real and all present. Set the mechanism aside for a moment and ask the question in the strongest form it can take — if a machine were drawing these numbers, could the room move them? The chamber would not be a vacuum and gravity would not be switched off. The point is that a force which acts on every ball equally cannot change which ball comes out. Gravity is the clearest case: every ball sits in the same field and falls at the same rate, which is why a hammer and a feather land together in a vacuum. Gravity does not prefer 7 to 31. Neither does the temperature of the air, because the chamber is one body of air at one temperature. They set the scale of the physics; they do not break the tie between one ball and another.

What would break the tie is a difference between the balls themselves — one a tenth of a gram heavier, or a millimetre wider, or holding a static charge. So the environment only matters if it is large enough to compete with the differences the balls are already permitted to have. It is not, and not by a small margin.

Only the part of gravity that differs between two balls could matter. Over half a metre of chamber that is 0.0000157 of a percent of gravity. The Moon and Sun together add a tidal pull of 0.0000164 of a percent, and across the whole one-metre chamber the difference between two balls is 7.70e-12 of a percent — seventeen billion times smaller than the tolerance. The atmosphere has more leverage, but it only pushes through the ball's buoyancy, which is itself 0.1035 of its weight; a five-degree change moves that by 0.001753 of the ball's weight.

The draw data bounds the outcome independently of all of this. Across 61,117 drawn numbers, a ball whose true chance differed by 5.4 or more from the others would have shown itself. None did. That is what "not significant" means here: not that the forces are absent, but that they are far too small to move a ball's chance by the smallest amount this much history could reveal.

The mechanisms that sound more promising than weather deserve the same treatment — static electricity, ball wear, airflow, machine calibration, and the rotation of ball sets. These belong to the machine era, and each is a real possible source of bias there. Each one, if it were doing anything, would leave one of a small number of signatures in the data — and the scans above, on the 503 draws where any of it could act, found none.

Two honest limits. Three of these cannot be measured from outside at all — the chamber's static charge, the calibration records, and the airflow — because none of them is published; what can be said is that all three would surface through the same signatures above, and those are empty. And "not significant" always means "not significant at this resolution", which is why the resolution is stated on this page rather than assumed.

Should you add Encore?

Encore is a different kind of game and worth separating from everything above. Its number is generated rather than drawn — which, since 14 May 2019, is also true of Lotto Max, so that is no longer the difference between them. The differences that remain are these: its prizes are fixed — one million dollars is one million dollars however many people win it, so the anti-sharing reasoning that justifies choosing unusual numbers in Lotto Max does not apply — and there is nothing to decide, because the number is assigned to the ticket, not chosen.

That leaves the arithmetic, and OLG publishes everything needed to do it. Each tier contributes its prize divided by its odds, and the three ways to win — right to left, left to right, and a run at each end — pay independently, so their contributions add.

So a dollar returns about $0.5141, a 48.6% margin to the house, and the chance of winning anything at all is 1 in 9.17 — which is exactly the figure OLG publishes, so the reading of the table is confirmed. A dollar spent on the main game returns about $0.4998. The two are within a penny of each other.

Where the return comes from is worth seeing. The million-dollar top prize supplies only $0.1000, while the two smallest tiers — two dollars for the last digit and five for the last two — supply $0.2229 between them. Most of what an Encore player gets back is small change.

One asymmetry, and it is the only argument for the main game over Encore: Lotto Max has an anti-sharing lever, because you choose the numbers and most players choose dates. Encore has no such lever and cannot have one. So a marginal dollar buys slightly more on the main game — both are losses, but one of them is a loss that can be shaved. Note also that the Lotto Max figure above is the best case, because it counts the full advertised jackpot; if the jackpot is shared, the real return is lower, and Encore's prizes are never shared.

Where the money actually moves

Every combination of seven numbers from fifty-two is equally likely — there are 133,784,560 of them — so no choice of numbers changes the chance of matching. What a choice does change is the number of people sharing the prize. A jackpot is divided between every winning ticket, so the same seven numbers are worth more to one player than to a thousand.

The lines above are therefore scored on two things: what the model thinks, and how many other players are likely to hold the same set. Numbers up to 31 are dates and are over-picked. Consecutive runs and tidy arithmetic patterns are over-picked too. The lines carry a distinctiveness score below, where a lower number means fewer people are likely to be holding it.

The penalty is a heuristic about other players, not a measurement of the balls. It is stated here so the reasoning is visible rather than buried in a choice of numbers.

Where the numbers came from

1,273 draws, from 25 September 2009 to 25 September 2026, taken from the British Columbia Lottery Corporation's public results feed — the same endpoint the winning-numbers page on PlayNow calls for itself. There are no gaps in the draw numbering.

The game changed twice inside that history, and the data shows it. The pool was 7 of 49 until 10 May 2019, 7 of 50 from 14 May 2019 — the same week Tuesday draws were added — and 7 of 52 from 14 April 2026. A number that only entered the pool in 2026 has not been "overdue" for seventeen years, so every recency feature is measured from the day that number became drawable. Without that, numbers 51 and 52 rank top of the board on an artefact rather than on evidence.

The feature set is not published. It is built from the draw history and from the conditions on the night, it is measured from the day each number entered the pool rather than from 2009, and it is ordinary rather than clever. A feature list is a recipe, and the recipe is not what this page is for — what the recipe produced is the backtest above.

Three standard models are fitted to those features, and they are not named here either; the result is the part that matters, and the result is that they lost. What is worth knowing is that the measurement that rejected them is a real one: a walk-forward backtest on draws the model had never seen, a shuffled-target control that had to fail, a planted-signal check that had to succeed, and an ablation that separates the history from the conditions. Beside the models run plain rules that fit nothing at all — and the plain rules did better, which is reported above rather than quietly dropped.

The other games

A draw does not have to use balls to belong here. What a game needs is a published history — every draw with its date — because that is what can be measured. Lotto Max is the worked example above: 1,273 draws, drawn by a machine until 14 May 2019 and by a random number generator since. These are the histories held beside it, each with the source it came from, so the same measurement can be run on each game rather than argued about in general.

Named on this site and not held yet, with the reason recorded rather than glossed over: euromillions — the community API answers 429 on every attempt — it exists, it throttles; eurojackpot — lottoland.com answers 403 to a plain request — needs another route; superenalotto — no host resolves, and the one that does answers 403; uk-lotto — both sites answer JavaScript shells — the draws are fetched in the browser; lotto-649 — playnow.com/lottery/lotto-649/ answers a JavaScript shell, and its services2 draw endpoint answers 404 for every code but LMAX — the draws have to be read in a browser, the same route the European games need.

Where this page's own history comes from. Every one of the draws above is read from the operator's own results page — the same endpoint fetch_draws.py in this project calls, one draw at a time.

What this does not claim. Only Lotto Max has been taken all the way through the measurement above. The rest are histories — enough to ask the question of each game, not yet enough to answer it for each. What the method expects to find is what it found here: the numbers cannot be called, and the only thing a choice changes is how many people share the prize.

Disclaimer

This site is a model and its own backtest. It is not advice and it is not a promise. The backtest on the worked example says the model cannot predict a draw, so nothing here improves the chance of matching one: every set of numbers has the same chance as every other set, and the only thing a choice changes is how many people share the prize.

Every lottery named on this site is a losing proposition by design. A ticket always costs more than the prizes it can win, in every game, on purpose — that is how the money for the prizes is raised. Lotto Max returns about $0.50 of every dollar staked and Encore about $0.51, and the same shape holds for the others. Play only with money you would not miss, never to recover a loss, and never on credit. You must be of legal age where you live — 18 or older in Ontario — and in a jurisdiction that allows you to play. If it stops being fun, ConnexOntario can help — call 1-866-531-2600.

No responsibility is accepted for any loss, in any game. The operator of this page is not responsible for a losing ticket, for a missed draw, for a misread selection slip, for a number copied from this page wrongly, or for anything done or not done because of it. The lines here are worked examples of a choice, nothing more. Check your numbers at the counter before you pay, and keep the printed ticket — it is the only record of what you are playing, and the slip is not.

No affiliation. Not endorsed. This page is independent. It is not produced, endorsed, sponsored or approved by the Ontario Lottery and Gaming Corporation, the Interprovincial Lottery Corporation, or the operator of Powerball, Mega Millions, EuroMillions, Eurojackpot, SuperEnalotto, Mega-Sena, Lotto 6/49 or the UK National Lottery — and no lottery operator anywhere has seen it. Every game named on this site is named only to describe what was measured, and each mark belongs to its owner. For official rules, winning numbers, prize amounts and claims, go to the operator of the game you play.

Privacy

Nothing is asked of you and nothing is stored about you. There is no account, no form, no comment box and no upload. The page does not read your clipboard and does not know which numbers you play.

Analytics is off until you say otherwise. The Google Analytics tag is not in this page. The consent script adds it only after you press Accept, so a visit that declines makes no request to Google and receives no cookie. Your answer is kept in this browser's local storage rather than in a cookie, and declining is remembered — the bar does not return to ask again.

Error reports. If the page throws an error, a short report is sent through a small relay on this domain so the fault can be fixed. It carries the message, the file, the line, the stack, the page path without its query string, and the browser's own user agent. It carries no cookie, no identifier, no session replay and no record of a person — and if your browser asks not to be tracked, it sends nothing at all.

No third-party code runs here. The styles, the script and the icons are all served from this domain. The only outside request this page can make is the Google Analytics tag, and only after you accept it.

What this page is not. It is not gambling advice and it is not a prediction service. It cannot tell you what will be drawn — it says so above, in numbers. A lottery ticket is a bad bet and this page will not pretend otherwise. If gambling stops being fun, the Ontario problem-gambling helpline is 1-866-531-2600, free and open at any hour.