Model Lab
This page replays filter, staking and calibration rules against the same earliest pre-event system records with settled outcomes. It does not use synthetic fixtures, and the variants are not separate trained production models. No money is attached; promotion requires the real-data gates below.
Variants
· 12 rule variants · same real pre-event recordsCumulative profit (staked units)
· Same fixtures · different rulesSide-by-side metrics
| Variant | Picks | Settled | W–L | Win% | Flat ROI | Staked ROI | +/− units | Avg edge | CLV | Brier |
|---|---|---|---|---|---|---|---|---|---|---|
Baseline (live) | 728 | 282 | 179–103 | 63.5% | -6.76% | -5.48% | -37.92u | -1.75pp | -1.75pp | 0.2161 |
Baseline · edge ≥ 3pp | 1 | 1 | 0–1 | 0.0% | -100.00% | -100.00% | -1.11u | +3.23pp | +3.23pp | 0.3120 |
Baseline · edge ≥ 7pp | 0 | 0 | 0–0 | — | — | — | +0.00u | +0.00pp | — | — |
Baseline · half stake | 728 | 282 | 179–103 | 63.5% | -6.76% | -5.48% | -18.96u | -1.75pp | -1.75pp | 0.2161 |
Baseline · ≥ 4 books | 671 | 272 | 172–100 | 63.2% | -6.72% | -5.66% | -37.46u | -1.70pp | -1.70pp | 0.2176 |
Edge ≥ 5pp | 0 | 0 | 0–0 | — | — | — | +0.00u | +0.00pp | — | — |
Market-shrunk | 1 | 1 | 0–1 | 0.0% | -100.00% | -100.00% | -1.00u | +2.26pp | +3.23pp | 0.3013 |
Market-shrunk · light | 4 | 4 | 1–3 | 25.0% | -58.25% | -58.25% | -2.33u | +2.41pp | +2.68pp | 0.3790 |
Favourites only | 505 | 235 | 159–76 | 67.7% | -4.93% | -4.93% | -11.58u | -2.05pp | -2.05pp | 0.2093 |
Value underdogs | 0 | 0 | 0–0 | — | — | — | +0.00u | +0.00pp | — | — |
Calibrated (Platt-lite) | 11 | 3 | 1–2 | 33.3% | -13.33% | -13.33% | -0.40u | +2.43pp | +0.88pp | 0.2683 |
Calibrated · strong shrink | 38 | 5 | 2–3 | 40.0% | 3.20% | 3.20% | +0.16u | +2.93pp | -0.03pp | 0.2378 |
Methodology
· How a variant gets promotedEvery variant is scored on the same auto-logged picks the live baseline model sees. Variants only differ in their filter rule (which picks they take) and staking rule. Outcomes (won / lost / void) come from the same reconciled results — no rewriting history.
- Baseline is whatever you see on the public track record today (recommended stake, no filter).
- Challenger variants paper-trade alongside. No money is attached.
- A challenger is promoted to baseline only when it beats the live model on staked ROI over ≥100 settled picks with positive CLV. Single-week noise doesn't move the needle.
- Brier score is computed on each variant's emitted probability (calibrated/shrunk variants use the transformed prob), so calibration improvements show up directly.
Research only. Past simulated performance is not indicative of future results.
Monte Carlo simulator
· Stress-test any model: ours or yoursRuns thousands of simulated betting seasons against the inputs below to show the full distribution of outcomes — not just the average. Use it to see how often a model bankrupts, what a realistic worst-case looks like, and whether the median path is actually profitable. Upload your own picks as CSV to bootstrap from real history instead of assumed parameters.
CSV with columns edge_pct, odds, stake (header optional, stake optional). When uploaded, the simulator samples real picks instead of using the edge% / odds inputs below. Staking rule from inputs is used when stake is 0 or missing.