CSP · T1–T3 · Research
Profit research
Rules are fitted on older matches, then frozen for later and future matches. Every selected loss stays in the calculation.
All model selections: ROI →Search reproducibility
Run: cb4b3077d9432e2f · Declared rule space: 720 · Selections evaluated: 1597
Dataset rows: 877 · Data cutoff: 2026-10-29 18:24 UTC+3 · Dataset hash: cb4b3077d943
Chronological split; fitted and later samples are separate; future verification is measured separately.
One shared rule for all matches
A single global rule, not a separate rule for every match. No profitable result is invented: wins and losses are scored after selection. Full-history tuning and later validation are different samples.
Updated: 2026-11-21 04:32 UTC+3 · Rows with prices at least 1.30: 1472 · Matches: 324 · Holdout starts: 2026-11-03 10:00 UTC+3
Run: 454a20fc3409 · Dataset rows: 1472 · Data cutoff: 2026-11-20 21:40 UTC+3 · Dataset hash: 454a20fc3409
Search budget: 200000 selections / 180.0s · Spent: 7570 / 2.145s · Declared grid: 7560
Minimum unique matches per segment: 20 · Fixed stake: 1 units · T1–T3 · Winline
Historical optimum uses known outcomes and can overfit. A positive result on history does not guarantee future profit. Different markets on one match are correlated.
Overall ROI →
| Sample | Settled | Win rate | Profit, units | ROI | Maximum drawdown |
|---|---|---|---|---|---|
| All qualifying forecasts | 1190 | 54.4% | -88.63 | -7.4% | 94.71 |
| Whole-history fit: max ROI% | 24 | 62.5% | +5.46 | 22.8% | 3.00 |
| Whole-history fit: max profit | 76 | 59.2% | +9.34 | 12.3% | 6.07 |
| Older sample: max ROI% rule | 29 | 58.6% | +5.73 | 19.8% | 2.65 |
| Older sample: max profit rule | 76 | 59.2% | +9.34 | 12.3% | 6.07 |
| Later sample: max ROI% rule | 7 | 57.1% | +1.21 | 17.3% | 2.00 |
| Later sample: max profit rule | 0 | — | +0.00 | — | 0.00 |
Whole-history fit: max ROI%: Confidence ≥ 50.0% · Price 1.3–3 · EV ≥ 0.15 · S, A · BO—
Markets: Total maps
Whole-history fit: max profit: Confidence ≥ 50.0% · Price 1.3–3 · EV ≥ 0 · S · BO—
Markets: Total maps, Map round totals
Later sample: max ROI% rule: Confidence ≥ 50.0% · Price 1.3–3 · EV ≥ 0.15 · S, A · BO—
Markets: Total maps, Map round totals
Later sample: max profit rule: Confidence ≥ 50.0% · Price 1.3–3 · EV ≥ 0 · S · BO—
Markets: Total maps, Map round totals
Distinct rule selections tested: 2067 · Use the ROI pages to inspect every selected match, probability, coefficient and result.
The declared grid was searched exhaustively, so this is the best rule found inside it — not a global optimum over all possible rules. Best profit: +9.34 units · Runner-up profit: +8.85 units · gap 0.49
ROI: one pick →
| Sample | Settled | Win rate | Profit, units | ROI | Maximum drawdown |
|---|---|---|---|---|---|
| All qualifying forecasts | 253 | 49.8% | -44.23 | -17.5% | 48.43 |
| Whole-history fit: max ROI% | 20 | 70.0% | +5.26 | 26.3% | 1.27 |
| Whole-history fit: max profit | 42 | 71.4% | +5.68 | 13.5% | 2.43 |
| Older sample: max ROI% rule | 20 | 70.0% | +5.26 | 26.3% | 1.27 |
| Older sample: max profit rule | 42 | 71.4% | +5.68 | 13.5% | 2.43 |
| Later sample: max ROI% rule | 0 | — | +0.00 | — | 0.00 |
| Later sample: max profit rule | 0 | — | +0.00 | — | 0.00 |
Whole-history fit: max ROI%: Confidence ≥ 50.0% · Price 1.3–2 · EV ≥ 0 · S · BO—
Markets: Map round totals
Whole-history fit: max profit: Confidence ≥ 55.0% · Price 1.3–2 · EV No extra filter · B · BO—
Markets: Winner
Later sample: max ROI% rule: Confidence ≥ 50.0% · Price 1.3–2 · EV ≥ 0 · S · BO—
Markets: Map round totals
Later sample: max profit rule: Confidence ≥ 55.0% · Price 1.3–2 · EV No extra filter · B · BO—
Markets: Winner
Distinct rule selections tested: 1814 · Use the ROI pages to inspect every selected match, probability, coefficient and result.
The declared grid was searched exhaustively, so this is the best rule found inside it — not a global optimum over all possible rules. Best profit: +5.68 units · Runner-up profit: +5.68 units · gap 0.00
Future verification
Policy frozen at: 2026-10-22 10:39 UTC+3
Actual choices made after policy freeze. Choices keep their original probability, price and decision time.
Winner improvement
Same later matches, same observed pre-decision prices. Serving forecasts are retained for comparison.
Training rows: 122 · Holdout rows: 25
| Candidate | Win rate | Brier | Log loss | Profit, units |
|---|---|---|---|---|
| platt | 56.0% | 0.2424 | 0.6779 | -2.66 |
| market | 72.0% | 0.2104 | 0.6003 | +1.28 |
| serving | 52.0% | 0.2482 | 0.6911 | -4.61 |
| blend_half | 72.0% | 0.2217 | 0.6323 | +1.28 |
Preview: model + market · Future verification
Settled: 0 · Pending result: 0 · Profit, units: +0.00 · ROI: —
Only forecasts created after the freeze enter this sample; historical optimum is not recomputed for them.
Preview: market baseline
The original serving artifact is unavailable. This separate preview uses fresh paired bookmaker prices without margin; it does not recreate the missing model.
Model run: 5464 · Settled: 62 · Pending result: 18 · Profit, units: -7.27 · ROI: -11.7%
Upcoming forecasts: 0