15 draws
Short-term frequency signal — 50% weight · optimal at 15 from backtest
Numbers in pair ranking + backtest evaluation
Digit Probability Scores — Next Draw
Score = 50% recency (last W draws) + 30% all-time base rate
+ 20% overdueness (draws since last ÷ avg gap, capped at 2×).
Normalised to 100%. The bar marker = 10% (fair random baseline per digit).
Note: overdueness is the weakest signal — it scored 13.8% alone vs 15% random in the backtest.
Do not over-interpret high overdue ratios.
Normal (<17%)
Calibrated 17–22% · zone hit rate 22.3% vs 19% baseline
Elevated >22% · rare zone, treat with caution
Top Unordered Pairs
Each row covers two tickets — buy both AB and BA.
Pair {5,9} = tickets 59 and 95. Symmetry confirmed: no positional bias.
Cumulative P — Next N Draws
P(digit appears ≥1× in N draws). Useful for tracking a digit across several upcoming draws rather than betting a single draw.
Walk-Forward Backtest
Model retrained on all prior draws only at each step — zero look-ahead bias.
Pair hit = drawn number or its mirror is in the top-N list.
+2 / +4 = did a pair hit occur within 2 or 4 draws from the prediction point?
Chips outlined in amber = digit was in elevated zone (>22%) at prediction time.
Model Calibration — What the Backtest Actually Found
Grid-search on 458 draws (2006–2026) · walk-forward · TWO prize only
✓ What holds up
- Recency (W=15) is the best individual signal — 16.4% hit rate alone.
- Base rate is second best — 17.8% alone. More stable than overdueness.
- Combined model best combo: 50/30/20 split → 18.2% vs 15% baseline (Z=1.87).
- Calibrated zone 17–22%: zone analysis shows 22.3% actual hit rate vs 19% baseline.
- Symmetry confirmed: "06" = "60", no positional bias. Always buy both tickets.
- Draw is fair: digit chi²=15.09 < 16.9 critical. No systematic number bias.
✗ What doesn't hold up
- Overdueness alone: 13.8% — worse than random. Was 45% weight in old model.
- The "OC correction" was fiction: digits above 24% appeared in 0.5% of draws (2 data points). No valid conclusion.
- High overdue ratio ≠ high hit rate. Gaps are highly variable (CV ≈ 0.9) — no reliable mean-reversion.
- Z=1.87 is not statistically significant (need 1.96). No confirmed edge at p<0.05.
- Year variance is extreme: 8% to 40% year-on-year. The model is not stable across time.
- Elevated zone (>22%): only 3% of draws reach this, insufficient sample to characterise.
⚠️ No statistically confirmed edge. The draw is indistinguishable from fair random over 458 draws.
The model provides structured probability ranking — a rational shortlist, not a prediction.
Use it to narrow a watch list over multiple draws, not to call a specific result on a specific draw.