Model lab
Serious model lab
Rolling Poisson, Elo/form, capped market overlays, totals, Asian handicap and longshot guard models tested in public.
Updated 2026-05-11 21:58 UTC. Paper-trading research only.
Generated: 2026-05-11T21:58:47+00:00 Train before: 2022-07-01 Test from: 2022-07-01
This run builds rolling pre-match Poisson, Elo/form, hybrid market overlays, Asian handicap, totals and longshot guard models. It is paper trading and backtesting only.
Anything using closing prices, closing lines or close-to-open movement is labelled diagnostic.
Market Baseline
- Opening 1X2 log loss: 0.9991 over 23967 matches.
- Closing 1X2 log loss: 0.996 over 23969 matches.
Strategy Results
| Strategy | Type | Bets | P/L | ROI | W | L | Push | Avg odds | Max DD | Verdict |
|---|---|---|---|---|---|---|---|---|---|---|
| Diagnostic AH line movement | diagnostic | 8525 | +367.69u | 4.3% | 4487 | 3451 | 587 | 1.84 | -36.02u | diagnostic |
| Diagnostic longshot with market move | diagnostic | 965 | +68.45u | 7.1% | 238 | 727 | 0 | 4.635 | -38.73u | diagnostic |
| Diagnostic totals price movement | diagnostic | 8720 | +96.08u | 1.1% | 4618 | 4102 | 0 | 1.937 | -128.25u | diagnostic |
| Elo/form guarded 1X2 | live-test | 4344 | -173.95u | -4.0% | 1778 | 2566 | 0 | 2.409 | -224.65u | unstable |
| Hybrid market/model 1X2 capped | live-test | 123 | -1.79u | -1.5% | 46 | 77 | 0 | 2.639 | -17.81u | unstable |
| Hybrid market/Poisson AH opening | live-test | 291 | +15.39u | 5.3% | 144 | 125 | 22 | 2.0 | -12.24u | candidate |
| Hybrid market/Poisson totals | live-test | 688 | +5.88u | 0.9% | 329 | 359 | 0 | 2.126 | -27.89u | unstable |
| Market price-shop 1X2 | live-test | 21 | +17.27u | 82.2% | 13 | 8 | 0 | 3.076 | -3.87u | exploratory |
| Poisson AH opening line | live-test | 18119 | -239.74u | -1.3% | 8467 | 8440 | 1212 | 1.969 | -276.97u | reject |
| Poisson guarded 1X2 | live-test | 4879 | -248.95u | -5.1% | 1753 | 3126 | 0 | 2.771 | -258.53u | unstable |
| Poisson totals value | live-test | 6994 | -199.77u | -2.9% | 3468 | 3526 | 0 | 1.981 | -204.55u | reject |
| Rough-patch fade proxy AH | live-test | 420 | -30.79u | -7.3% | 176 | 226 | 18 | 1.964 | -43.30u | unstable |
Read
- The market remains hard. Closing prices improve log loss over openers, which is what we should expect.
- No standalone live model deserves to be auto-promoted yet.
- Poisson and Elo are useful as small overlays, not as market-overruling brains.
- Asian handicap line movement remains the best research direction, but movement is diagnostic until openers/current prices are stored live.
- The longshot/favourite guard should stay. The rough-patch fade proxy exists to test the idea, not endorse it.
Recommendation
No standalone live model earns promotion yet. Promote the market router: store openers, paper-trade Asian handicap and DNB first, keep longshot catalyst gate, and use Poisson/Elo only as capped overlays.
Top candidates:
- Hybrid market/Poisson AH opening: 291 bets, ROI 5.3%, live-test
- Diagnostic longshot with market move: 965 bets, ROI 7.1%, diagnostic
- Diagnostic AH line movement: 8525 bets, ROI 4.3%, diagnostic
Rejected / not promoted:
- Elo/form guarded 1X2: 4344 bets, ROI -4.0%, unstable
- Hybrid market/model 1X2 capped: 123 bets, ROI -1.5%, unstable
- Hybrid market/Poisson totals: 688 bets, ROI 0.9%, unstable
- Market price-shop 1X2: 21 bets, ROI 82.2%, exploratory
- Poisson AH opening line: 18119 bets, ROI -1.3%, reject
- Poisson guarded 1X2: 4879 bets, ROI -5.1%, unstable
- Poisson totals value: 6994 bets, ROI -2.9%, reject
- Rough-patch fade proxy AH: 420 bets, ROI -7.3%, unstable
Raw report: serious-model-lab.md. Machine-readable data: serious-model-lab.json.