Forecast · next 6 draws

Each upcoming date is read for its own three calendar attributes, then every ball is scored on how it has behaved under exactly those conditions: score = z(weekday) + z(day of month) + z(date root) + z(overall), where z = (seen − expected) / √expected and expected respects the ball-pool eras. A cell is ignored unless it expected at least 5 occurrences: without that floor the thinnest cell reads as the most significant, because √expected shrinks with the sample, and a ball from the 2026 pool expansion can be ranked on a single draw. The top seven are the cross-score set. The pair-anchored set starts instead from the pair with the most appearances under this date's weekday and root, then fills by score.

Track record so far. Every forecast above is written to a ledger when it is issued and scored once the draw lands — the earliest issue for each date, so nothing is graded on a call made after the fact.

• z(weekday)+z(dom)+z(root)+z(overall) — 5 settled, mean 1.00 hits against 0.94 by chance (+0.06). too few to mean anything.
• z(weekday)+z(dom)+z(root)+z(overall) | min-exposure 5 — 4 settled, mean 0.75 hits against 0.94 by chance (-0.19). too few to mean anything.

A handful of draws cannot separate this from chance in either direction; it takes hundreds. The powered comparison is backtest_forecast.py, which replays the method over 400 draws and finds it does not beat random selection.

draw	forecast	hits	alt	drawn
2026-10-06	7 18 19 26 28 38 39	0	0	11 12 22 31 32 46 50
2026-10-06	7 18 19 26 28 38 39	0	0	11 12 22 31 32 46 50
2026-10-02	7 19 24 28 30 36 39	0	0	2 4 20 23 25 41 47
2026-10-02	7 19 24 28 30 36 39	0	0	2 4 20 23 25 41 47
2026-09-29	6 19 22 26 28 38 39	2	1	6 10 13 22 23 43 49
2026-09-29	6 19 22 26 28 38 39	2	1	6 10 13 22 23 43 49
2026-09-25	7 19 24 28 31 39 52	2	2	8 24 38 41 43 50 52
2026-09-25	7 19 24 28 30 31 39	1	1	8 24 38 41 43 50 52
2026-09-22	19 22 28 30 34 36 38	1	1	1 10 12 17 30 37 42