Method leaderboard
Every prediction method we run, scored on real draws — including a pure-random control arm. This is the table no tipster site would ever publish.
| Method | Picks | Straight | Box | Box/pick | vs chance | Return |
|---|---|---|---|---|---|---|
| ProfileEdge | 72 | 0 | 38 | 0.53 | +0.20 | 153% |
| Cold endingsCollecting | 51 | 0 | 21 | 0.41 | +0.08 | 37% |
| Weighted ensembleCollecting | 57 | 0 | 23 | 0.40 | +0.07 | 124% |
| FollowerCollecting | 57 | 2 | 19 | 0.33 | +0.00 | 80% |
| AI (Claude)Collecting | 51 | 0 | 15 | 0.29 | -0.04 | 26% |
| RhythmChance | 57 | 0 | 16 | 0.28 | -0.05 | 84% |
| Own-cycleCollecting | 43 | 0 | 12 | 0.28 | -0.05 | 105% |
| DueChance | 78 | 1 | 21 | 0.27 | -0.06 | 31% |
| Trend-builtChance | 72 | 3 | 18 | 0.25 | -0.08 | 96% |
| In formChance | 51 | 1 | 12 | 0.24 | -0.10 | 90% |
| AnniversaryChance | 51 | 1 | 12 | 0.24 | -0.10 | 67% |
| HotChance | 78 | 1 | 18 | 0.23 | -0.10 | 111% |
| Near missChance | 57 | 0 | 13 | 0.23 | -0.10 | 67% |
| CompanionsChance | 72 | 2 | 16 | 0.22 | -0.11 | 102% |
| JiranChance | 63 | 0 | 14 | 0.22 | -0.11 | 29% |
| CoolingChance | 51 | 1 | 11 | 0.22 | -0.12 | 90% |
| Random controlcontrolChance | 57 | 0 | 11 | 0.19 | -0.14 | 107% |
| PositionalChance | 66 | 3 | 12 | 0.18 | -0.15 | 112% |
| DateCollecting | 11 | 0 | 2 | 0.18 | -0.15 | 9% |
| Head pairsChance | 51 | 0 | 9 | 0.18 | -0.15 | 26% |
| Repeat watchChance | 72 | 0 | 12 | 0.17 | -0.16 | 18% |
| EndingChance | 53 | 4 | 8 | 0.15 | -0.18 | 69% |
| ML modelChance | 57 | 1 | 7 | 0.12 | -0.21 | 28% |
The return column
What RM1 iBox Big on every pick, on every board, would actually have returned — priced with the real dividend table. Counting box hits treats a consolation and a 1st prize as the same event; this does not. A fair draw returns about 67%, and 100% is break-even.
Read this column last. A single RM1 bet has an average return of 67 sen but a standard deviation of RM5.75, because 38% of all expected return sits in the 1st prize alone — an event with a 0.24% chance. That makes return roughly 5 times noisier than the box-hit rate and in need of about 25 times more data to mean anything, which is why the verdict is decided on hits and the table is not sorted on money. For scale: the pure-random control arm is currently returning 107%.
From 20 Aug 2026 every row here is scored on the Malaysian-licensed boards only (Magnum, Toto, Da Ma Cai, Sabah 88, STC, CashSweep) — the same six the picks are built from. Fewer boards per night means fewer chances to hit, so both the methods and the random control sit lower than they did before, and the whole history is recounted on this basis.
How we scoreThe statistics behind each verdict, and the rules for retiring a method›
How the verdict is decided (and why it isn't decided yet)
Re-checking a leaderboard every night and calling the best row a winner is how false discoveries are made — repeated peeking inflates the error rate. The verdict column instead runs Wald's sequential probability ratio test per method: chance-level box rate vs a 1.5× edge, with 5% error caps that hold no matter how often we look. A method is only called an Edge or confirmed as Chance when the accumulated evidence crosses the threshold; until then it honestly says Collecting.
For scale: a method performing exactly at chance needs roughly 95 picks before the test can confirm it — verdicts here are earned slowly by design.
Method retirement policy
Pre-committedA method is retired from the nightly picks when its SPRT verdict is "chance" AND it has at least 1,000 scored picks. Retired methods keep their full record on this leaderboard, marked retired. Any change to the line-up is announced here with its date. The random control is never retired — a control removed for behaving like a control is not a control.
Status right now: no method meets this rule. That is the whole point of publishing it today — a rule written after a method fails is not a rule, it is an excuse.
Why a pick floor as well as a verdict: the SPRT can confirm "chance" fairly early on a method that simply had a quiet run. The floor makes sure nothing is benched on a short record, and keeping the retired rows visible means nothing quietly disappears from the scoreboard once it stops being flattering.
Relegation & promotion
The softer stage before retirement: a MAIN method (hot, overdue, own-cycle, ending) whose verdict is "chance" with at least 150 scored picks loses its main-pick slot and continues in the deep set — still locked, still scored, still on this board. Its slot goes to the deep method with the highest ensemble trust weight and at least 150 picks of its own. The random control and the ensemble itself are never eligible. Swaps happen only when this rule fires, and are announced here with their date.
Status right now: no main method meets the relegation rule. Published today, before any does.
Are “hot numbers” real? The shrinkage test
Empirical Bayes on the last 365 days of top-3 hits, fitted across all 10,000 numbers.
| Number | Hits (365d) | Honest estimate | Excess kept |
|---|---|---|---|
| 4251 | 6 | 1.04 | 7% |
| 2375 | 5 | 0.97 | 7% |
| 3264 | 5 | 0.97 | 7% |
| 6437 | 5 | 0.97 | 7% |
| 7623 | 5 | 0.97 | 7% |
| 7675 | 5 | 0.97 | 7% |
| 7713 | 5 | 0.97 | 7% |
| 7874 | 5 | 0.97 | 7% |
| 8047 | 5 | 0.97 | 7% |
| 8134 | 5 | 0.97 | 7% |
“Honest estimate” is the posterior expected hit count for the NEXT 365 days, after shrinking toward the all-number average of 0.69. “Excess kept” shows how much of a number's above-average performance the math believes; near 0% means its hot streak was luck.