Cross-Market Sweep — Top Results
Live top performers from the running sweep. Refreshes every 30s. Filter by min bets to drop noise — most signal is in 100+ bet rows.
Test window: Jan-Feb 2026 (2 months, out-of-sample, same for every cutoff). ROI/stake is time-independent (per €1 staked). BANKROLL Δ & P&L are over the 2-month window. ANN. = annualized.
To trust a row: look for FLAT € (real edge after stripping compounding), t-stat ≥ 5.4 (survives multiple-testing correction), and REPL ≥ 3/4 (works across cutoffs — not lucky on one). A high BANKROLL Δ alone means little.
📖 How to read this leaderboard (columns + rules of thumb)
Quick rule: A row is a candidate "real edge" only if it has n > 100 bets, t-stat ≥ 5.4, and REPL ≥ 3/4. Anything else is probably random luck (the sweep tries millions of combinations — top P&L in pure noise easily looks +20%).
| column | what it is | what to look for |
|---|---|---|
| MARKET | H2H = 1X2, OU25 = over/under 2.5 goals, OU15 = OU 1.5, AH = Asian Handicap (line in cfg) | any — leaderboard is unified across markets |
| OUT | Outcome: home/draw/away (H2H), over/under (OU), home/away (AH at the line) | — |
| SIDE | BACK = bet for outcome to occur (lose stake on loss). LAY = bet against (lose stake×(odds−1) on loss; win stake×(1−2%comm)). | compare BACK vs LAY for same market — model edge often only one-sided |
| SLOT | Decision time before kickoff: T-1440 (24h), T-720 (12h), T-60 (1h), T-10 (10min) | longer slots → more liquidity but stale info; T-10 is freshest but risk lockup overlap |
| CUT | Train cutoff: c1 (12mo Jan 24), c2 (16mo), c3 (20mo), c4 (22mo). Test is always Jan-Feb 2026 | "more data is better" — c4 should be ≥ c1 |
| MODEL | lgbm / rf / xgb / catboost / gbm (gradient-boosted, random forest, etc.) | — |
| FEATURES | Feature set: own (target market only), plus_ah/plus_ou/plus_h2h (cross-market), all_xmarket | cross-market sets often add edge for free |
| STAKE | Sizing algo: FRACTIONAL, KELLY, FLAT, VOL_TARGET, RIDGE, CPPI, RISK_PARITY, DD_AWARE, CVAR | same edge with different sizing → ROI/stake stays similar; BANKROLL Δ varies |
| EDGE | Min edge filter the model required to bet (e.g. 0.04 = 4 percentage-points above implied) | tighter edge usually = fewer bets, higher quality |
| CONF | ≥X.XX = min P(outcome) for BACK; ≤X.XX = max P(outcome) for LAY (i.e. lay only short-prob outcomes) | too tight → 0 bets; too wide → noise |
| N | Number of bets placed in the test window (Jan-Feb 2026) | ≥100 ideal for trust. <30 = ignore, even if ROI looks great |
| WR | Win rate from the side's perspective (LAY wins when outcome doesn't happen) | compare to break-even WR for that side+odds (e.g. LAY at odds 2.0 needs WR > 50%) |
| ROI/stake | P&L ÷ total amount staked. Time-independent — €X profit per €1 staked. After 2% commission + vig correction. | +5-15% is realistic edge; +30%+ is suspect (small N or selection bias) |
| FLAT € | n_bets × €100 × ROI — what the strategy earns at fixed €100/bet (no Kelly compounding fantasy) | this is the realistic € number for actual deployment. Beats BANKROLL Δ which compounds wildly. |
| t-stat | ROI × √n / σ_per_stake. Statistical significance. After Bonferroni correction for the 1.8M-trial grid, threshold is |t| ≥ 5.4. | green ≥ 5.4 = significant; grey < 5.4 = could be luck. Higher always better. |
| REPL | n_pos / n_cutoffs_done. How many train cutoffs (out of 4) of the same hyperparams had positive ROI on the same OOS test window. Asterisk = some cutoffs still pending. | ≥ 3/4 = robust real edge. 1/4 = lucky on one cutoff. 0/N = not edge. |
| BANKROLL Δ | (final - €10,000) / €10,000 over Jan-Feb 2026 with full Kelly compounding | flashy but inflated — same edge with different staking can give 10× different numbers. Don't deploy on this alone. |
| ANN. | (1 + bankroll Δ)(12/n_months) − 1 — annualized return. Comparable to S&P / index returns. | only meaningful if t-stat + REPL pass. Otherwise math on noise. |
| P&L | Absolute € profit/loss over the 2-month window with Kelly compounding | same caveat as BANKROLL Δ |
| DD | Max peak-to-trough drawdown during the test window | < 15% is healthy; > 25% means survivable but unpleasant; > 50% means don't deploy |
Worked example — top H2H result (in current snapshot):
H2H/home/LAY/T-1440/c1/lgbm/own/RIDGE — 59 bets, ROI +55%, FLAT € +€3,251, t-stat 8.51, REPL 2/2*
Reading this: model lgbm trained on 12 months of data (c1) decided that home outcomes at T-1440 odds were over-priced; LAYed home with edge ≥ 0.02. 59 LAY bets won 78%, ROI/stake 55%. At a flat €100/bet that's +€3,251 over 2 months. t-stat 8.51 is far above the 5.4 multiple-testing threshold → significant. REPL 2/2 means both cutoffs done so far were positive — but 2 cutoffs are still pending so it's not yet 3+/4 robust. Watch this row — if c2/c3/c4 also come up positive, this is a candidate for live deployment.