Pick a fixture, run the simulator, and explore the pitch in detail.
Forecast Record
Every forecast is saved before the game starts and scored against what actually happened, next to the sportsbooks’ prices. Here’s how we keep score, in public.
47Premier League games compared1.0160our error score0.9912sportsbooks’ error scoreLower is better. The sportsbooks are ahead so far.
One record, four sports.
| Sport | Games compared with the sportsbooks | Our forecast error score, lower is better | Sportsbooks error score, lower is better | Record |
|---|---|---|---|---|
| Premier LeagueHome / draw / away · 17 Oct 2026–22 Nov 2026 | 47 | 1.0160 | 0.9912 | |
| NFLHome / away · 9 Oct 2026–20 Nov 2026 | 83 | 0.6362 | 0.5727 | |
| NHLHome / away · 7 Oct 2026–22 Nov 2026 | 258 | 0.6692 | 0.6553 | |
| NBAHome / away · 20 Oct 2026–22 Nov 2026 | 239 | 0.6156 | 0.5712 |
Games compared counts finished regular-season games that had both our forecast and saved sportsbook prices 30 minutes before kickoff (for the Premier League, since the paper record began); every forecast in a row is scored on those same games. It is usually smaller than the full record’s count of games with a saved forecast, which also includes preseason games, games still to play and games with no saved prices. Error score is log loss: lower means the forecast gave more chance to what happened. In the tables below, Darius is our own forecast (the chances on our game pages), Arta is the sportsbooks’ average adjusted by past results (Arta 2.1, the forecast Mitra’s paper picks use), and Bahram is the sportsbooks’ consensus with the margin removed. Each sport is scored on its own; sports are not ranked against each other. Guard and the exit research cover the Premier League only.
Every game, scored.
Saved before kickoff, graded after the final, win or lose. Latest first.
Every scored game loads here.
Does the forecast add information beyond the odds?
Premier League · 17 Oct 2026 to 22 Nov 2026
47 common finished games · same games for every scored model · lowest log loss first
Columns
| DariusOutcome forecast from prior team performance. | 1.0160 | 45% · 21/47 | 0.6068 |
| ArtaMarket probabilities adjusted using past results. | 1.0037 | 51% · 24/47 | 0.5999 |
| BahramSportsbook consensus with the margin removed. | 0.9912 | 51% · 24/47 | 0.5923 |
No matching forecasts.
Lower log loss is better. Darius is our forecast, saved before kickoff; Arta (the market adjusted by past results) and Bahram (the sportsbooks’ consensus) are rebuilt from pre-kickoff prices. So far the sportsbooks’ consensus is ahead of our forecast on these 47 games (1.0160 against 0.9912).
What happened using my sportsbook’s prices?
Arta probabilities · a separate $1,000 paper replay at each book
| FanDuel | −$30.00 | 30 | 100% | 5.5% |
| Fanatics | −$133.00 | 27 | 100% | 13.3% |
| DraftKings | $12.00 | 35 | 100% | 6.7% |
| BetMGM | −$51.00 | 28 | 100% | 12.2% |
No matching sportsbooks in this saved comparison.
49 games · coverage = share with eligible prices. Each book is its own paper replay: different prices mean different picks, so rows are not identical tickets.
What changed when Guard skipped uncertain games?
A matched comparison, with holding kept beside both filters
49 saved inputs · retrospective replay · 2026-27 Arta fit · $1,000 starting funds per method
Columns
| Mitra Select · Hold | −$126.58 | 15.8% | 49 | 0 |
| Mitra Select Guard | −$56.50 | 10.9% | 42 | 7 |
| Mitra Select Guard Strict | −$10.50 | 5.6% | 15 | 34 |
No matching methods.
These methods share the annual forecast, saved prices, starting funds and allocation rules. Guard filters uncertain games; Strict adds a conservative probability check. Skipped includes excluded or unfunded inputs. These decisions were reconstructed afterwards, and the small sample does not establish lower future risk.
Show this comparison over time
Does the forecast add information beyond the odds?
NFL · 9 Oct 2026 to 20 Nov 2026
83 common finished games · same games for every scored model · lowest log loss first
Columns
| DariusOutcome forecast from prior team performance. | 0.6362 | 66% · 55/83 | 0.4446 |
| BahramSportsbook consensus with the margin removed. | 0.5727 | 72% · 60/83 | 0.3848 |
No matching forecasts.
Lower log loss is better. Darius is our forecast, saved before kickoff; Arta (the market adjusted by past results) and Bahram (the sportsbooks’ consensus) are rebuilt from pre-kickoff prices. So far the sportsbooks’ consensus is ahead of our forecast on these 83 games (0.6362 against 0.5727).
What happened using my sportsbook’s prices?
Darius probabilities · a separate $1,000 paper replay at each book
| FanDuel | −$263.93 | 59 | 84% | 26.4% |
| Fanatics | −$43.32 | 50 | 75% | 11.3% |
| DraftKings | −$121.01 | 58 | 86% | 18.2% |
| BetMGM | −$153.21 | 50 | 72% | 18.4% |
No matching sportsbooks in this saved comparison.
85 games · coverage = share with eligible prices. Each book is its own paper replay: different prices mean different picks, so rows are not identical tickets.
Does the forecast add information beyond the odds?
NHL · 7 Oct 2026 to 22 Nov 2026
258 common finished games · same games for every scored model · lowest log loss first
Columns
| DariusOutcome forecast from prior team performance. | 0.6692 | 60% · 154/258 | 0.4763 |
| BahramSportsbook consensus with the margin removed. | 0.6553 | 60% · 154/258 | 0.4631 |
No matching forecasts.
Lower log loss is better. Darius is our forecast, saved before kickoff; Arta (the market adjusted by past results) and Bahram (the sportsbooks’ consensus) are rebuilt from pre-kickoff prices. So far the sportsbooks’ consensus is ahead of our forecast on these 258 games (0.6692 against 0.6553).
What happened using my sportsbook’s prices?
Darius probabilities · a separate $1,000 paper replay at each book
| FanDuel | −$251.07 | 165 | 87% | 32.8% |
| Fanatics | −$188.36 | 136 | 74% | 22.3% |
| DraftKings | −$108.56 | 100 | 53% | 20.1% |
| BetMGM | −$76.29 | 124 | 65% | 18.0% |
No matching sportsbooks in this saved comparison.
268 games · coverage = share with eligible prices. Each book is its own paper replay: different prices mean different picks, so rows are not identical tickets.
Does the forecast add information beyond the odds?
NBA · 20 Oct 2026 to 22 Nov 2026
239 common finished games · same games for every scored model · lowest log loss first
Columns
| DariusOutcome forecast from prior team performance. | 0.6156 | 69% · 164/239 | 0.4276 |
| BahramSportsbook consensus with the margin removed. | 0.5712 | 71% · 169/239 | 0.3889 |
No matching forecasts.
Lower log loss is better. Darius is our forecast, saved before kickoff; Arta (the market adjusted by past results) and Bahram (the sportsbooks’ consensus) are rebuilt from pre-kickoff prices. So far the sportsbooks’ consensus is ahead of our forecast on these 239 games (0.6156 against 0.5712).
What happened using my sportsbook’s prices?
Darius probabilities · a separate $1,000 paper replay at each book
| FanDuel | −$354.96 | 200 | 97% | 41.8% |
| Fanatics | −$280.74 | 175 | 90% | 30.0% |
| DraftKings | −$310.75 | 163 | 80% | 36.7% |
| BetMGM | −$270.35 | 171 | 87% | 32.0% |
No matching sportsbooks in this saved comparison.
240 games · coverage = share with eligible prices. Each book is its own paper replay: different prices mean different picks, so rows are not identical tickets.
Try the comparison before choosing a plan.
Use the same forecasts on today’s games.
This record evaluates forecasts and paper selections, not lineup or tactical simulations. Inspect the worked lineup example.
Can I inspect the record?
Saved decisions, reconstructed comparisons and the underlying forecasts are kept separate.
One record, four sports.
| Sport | Games compared with the sportsbooks | Our forecast error score, lower is better | Sportsbooks error score, lower is better | Record |
|---|---|---|---|---|
| Premier LeagueHome / draw / away · 17 Oct 2026–22 Nov 2026 | 47 | 1.0160 | 0.9912 | |
| NFLHome / away · 9 Oct 2026–20 Nov 2026 | 83 | 0.6362 | 0.5727 | |
| NHLHome / away · 7 Oct 2026–22 Nov 2026 | 258 | 0.6692 | 0.6553 | |
| NBAHome / away · 20 Oct 2026–22 Nov 2026 | 239 | 0.6156 | 0.5712 |
Games compared counts finished regular-season games that had both our forecast and saved sportsbook prices 30 minutes before kickoff (for the Premier League, since the paper record began); every forecast in a row is scored on those same games. It is usually smaller than the full record’s count of games with a saved forecast, which also includes preseason games, games still to play and games with no saved prices. Error score is log loss: lower means the forecast gave more chance to what happened. In the tables below, Darius is our own forecast (the chances on our game pages), Arta is the sportsbooks’ average adjusted by past results (Arta 2.1, the forecast Mitra’s paper picks use), and Bahram is the sportsbooks’ consensus with the margin removed. Each sport is scored on its own; sports are not ranked against each other. Guard and the exit research cover the Premier League only.
Saved paper decisions Fixed before kickoff
Columns
| Mitra · Across booksSelection rules using Arta probabilities. | Saved | −$66.00 | -6.6% | 11.5% | 40 | — |
| Mitra · FanDuelSelection rules using Arta probabilities. | Saved | −$30.00 | -3.0% | 5.5% | 30 | — |
| Bahram · Across booksSportsbook consensus with the margin removed. | Saved | −$20.00 | -2.0% | 3.0% | 7 | — |
No scored methods in this group.
These are the saved Mitra and market decisions. Each portfolio began with $1,000. Different methods can select different games. Open selections are held at cost until settlement.
All outcome forecasts 7 models · common-game scores
Columns
| DariusOutcome forecast from prior team performance. | 1.0160 | 45% · 21/47 | 0.6068 | 47 |
| ArtaMarket probabilities adjusted using past results. | 1.0037 | 51% · 24/47 | 0.5999 | 47 |
| ArtabanTeam-strength ratings built from past results. | 1.0125 | 49% · 23/47 | 0.6069 | 47 |
| AnahitaPast results and injuries, selecting fewer signals. | 1.0158 | 47% · 22/47 | 0.6109 | 47 |
| RoxanaPast results and injuries, balancing signal weights. | 1.0335 | 45% · 21/47 | 0.6200 | 47 |
| TiridatesDecision trees combining performance signals. | 1.0639 | 47% · 22/47 | 0.6376 | 47 |
| BahramSportsbook consensus with the margin removed. | 0.9912 | 51% · 24/47 | 0.5923 | 47 |
No matching forecasts.
All 7 outcome models use the same 47 finished games. Artaban, Anahita, Roxana and Tiridates are statistical comparisons. Brier is available under Columns.
Every scored Premier League game loads here.
Paper replays by forecast Reconstructed from saved inputs
Columns
| Darius · Across booksOutcome forecast from prior team performance. | Replay | $31.27 | +3.1% | 9.7% | 43 | — |
| Arta · Across booksMarket probabilities adjusted using past results. | Replay | −$101.00 | -10.1% | 11.3% | 38 | — |
| Artaban · Across booksTeam-strength ratings built from past results. | Replay | $19.17 | +1.9% | 7.6% | 44 | — |
| Anahita · Across booksPast results and injuries, selecting fewer signals. | Replay | $18.59 | +1.9% | 9.8% | 43 | — |
| Roxana · Across booksPast results and injuries, balancing signal weights. | Replay | −$124.14 | -12.4% | 19.4% | 44 | — |
| Tiridates · Across booksDecision trees combining performance signals. | Replay | −$79.23 | -7.9% | 17.9% | 44 | — |
| Bahram · Across booksSportsbook consensus with the margin removed. | Replay | −$20.00 | -2.0% | 2.0% | 2 | — |
No scored methods in this group.
These were worked out afterwards from pre-kickoff forecasts and prices. Every portfolio starts with $1,000 under the frozen record allocation rules, while game coverage can differ. They are separate from saved decisions and the annual Guard rules.
Show curves Optional comparison
Choose the lines to display.
Paper balance, starting at $1,000. Overlapping curves can reflect identical decisions. These are the recorded values, without smoothing.
Sportsbook probability scores 49 common games · separate cohort
Columns
| Bet Victor | 0.9951 | 49 |
| Betfred (UK) | 0.9951 | 49 |
| Betano (UK) | 0.9957 | 49 |
| BoyleSports | 0.9958 | 49 |
| Coral | 0.9959 | 49 |
| Unibet (UK) | 0.9960 | 49 |
| BetMGM | 0.9962 | 49 |
| Ladbrokes | 0.9970 | 49 |
| Casumo | 0.9973 | 49 |
| Bahram · Market consensusSportsbook consensus with the margin removed. | 0.9978 | 49 |
| FanDuel | 0.9981 | 49 |
| Fanatics | 0.9982 | 49 |
| DraftKings | 0.9985 | 49 |
| Betway | 0.9985 | 49 |
| Paddy Power | 0.9985 | 49 |
| LeoVegas | 0.9985 | 49 |
| Grosvenor | 0.9990 | 49 |
| Virgin Bet | 0.9996 | 49 |
| BetRivers | 0.9998 | 49 |
| Betfair Sportsbook | 0.9998 | 49 |
| Sky Bet | 1.0003 | 49 |
| Caesars | 1.0003 | 49 |
| LiveScore Bet | 1.0008 | 49 |
| William Hill | 1.0010 | 49 |
| Arta 2.1Market probabilities adjusted using past results. | 1.0068 | 49 |
No matching forecasts.
Offshore sites not licensed in any US state (Bovada, MyBookie and similar) are left out of this table.
Every row in this table uses the same 49 games, with the last saved pre-kickoff price and margin removed. This cohort differs from the outcome-model comparison; scores across the two tables should not be ranked together.
Guard forward record, versions and skipped-game audit
Research rules · annual-refit Guard comparison
A separate forward record; the matched replay is shown in the overview.
Arta’s 2026-27 fit uses 760 games from 2024-25 and 2025-26, ending 2026-05-24. The uncertainty cutoff is 8.56 percentage points, fixed from training. These rows use that annual refit; the earlier strategies above keep their original forecast and frozen rules.
Forward · saved decisions
A new, separate $1,000 record. Each match day is allocated together before its first kickoff; missed days are never backfilled. Initialized 2026-10-19T17:27:55.063288Z
Columns
| Mitra MatchNo uncertainty filter | — | — | — | 0 | 40 | Missing Bet365 reference40 decisions. 0 rejected for high uncertainty. 40 missing Bet365 references. missing Bet365 reference: 40 |
| Mitra Match GuardTraining uncertainty filter | — | — | — | 0 | 40 | Missing Bet365 reference40 decisions. 0 rejected for high uncertainty. 40 missing Bet365 references. missing Bet365 reference: 40 |
| Mitra Match Guard StrictUncertainty + lower-bound check | — | — | — | 0 | 40 | Missing Bet365 reference40 decisions. 0 rejected for high uncertainty. 40 missing Bet365 references. missing Bet365 reference: 40 |
| Mitra SelectNo uncertainty filter | −$52.50 | -5.2% | 6.6% | 20 | 20 | Recorded40 decisions. 0 rejected for high uncertainty. 0 missing Bet365 references. guard off: 20 no candidate: 20 |
| Mitra Select GuardTraining uncertainty filter | −$32.50 | -3.2% | 6.3% | 18 | 22 | Recorded40 decisions. 2 rejected for high uncertainty. 0 missing Bet365 references. high uncertainty: 2 accepted: 18 no candidate: 20 |
| Mitra Select Guard StrictUncertainty + lower-bound check | $2.50 | +0.2% | 3.9% | 7 | 33 | Recorded40 decisions. 2 rejected for high uncertainty. 0 missing Bet365 references. high uncertainty: 2 accepted: 7 no candidate: 20 no eligible funded offer: 11 |
No matching Guard methods.
Show Guard curves Forward · saved decisions
Paper balances start at $1,000. A method with no eligible selections holds cash. Missing inputs are reported as unavailable.
Guard rules, input differences and decision audit
Match first chooses at Bet365, then routes that same outcome. Select chooses across funded books. Guard rejects a match when the largest of the three bootstrap probability standard deviations exceeds the training 75th percentile. Strict also requires the bootstrap 10th-percentile probability to give positive expected profit at the offered price. These are parameter-instability estimates, not certified confidence intervals for the true probability.
The decision controller, $10 or 2% match cap, 10% daily cap, 25% account cap and no-transfer rules are the research rules. The historical experiment used a football-data average and UK execution books. Here, the recorded complete-book average and FanDuel, Fanatics, DraftKings and BetMGM are the inputs. No Bet365 quote means no Match selection; another book is never substituted. The forward snapshot commits all known day fixtures 20–100 minutes before the first kickoff, with offers at most 60 minutes old.
Model mitra-guard-research-2026-27-v1 · model hash 1b8ac8a5ae9c4efc · policy hash 74b38840c89b2c28. No current-season outcomes were used in training or choosing the cutoff. Changing rules or a season’s frozen model stops writes to this record.
2026-10-19 · saved before kickoff · 2026-10-19T17:27:55.063288Z
2026-10-23 · saved before kickoff · 2026-10-23T17:22:47.591165Z
2026-10-24 · saved before kickoff · 2026-10-24T10:17:15.825587Z
2026-10-25 · saved before kickoff · 2026-10-25T12:34:30.070579Z
2026-10-31 · saved before kickoff · 2026-10-31T11:25:28.466196Z
2026-11-01 · saved before kickoff · 2026-11-01T14:50:25.671274Z
2026-11-02 · saved before kickoff · 2026-11-02T19:12:32.554599Z
2026-11-06 · saved before kickoff · 2026-11-06T18:50:40.150948Z
2026-11-07 · saved before kickoff · 2026-11-07T11:21:12.181114Z
2026-11-08 · saved before kickoff · 2026-11-08T12:29:55.098523Z
2026-11-21 · saved before kickoff · 2026-11-21T11:21:05.154463Z
2026-11-22 · saved before kickoff · 2026-11-22T12:25:00.815365Z
Columns
| Everton · Chelsea2026-10-17T11:30:00Z | 7.13 pp | 8.56 pp | $10.00 | accepted |
| Fulham · Hull City2026-10-17T14:00:00Z | 8.42 pp | 8.56 pp | $10.00 | accepted |
| Brentford · Liverpool2026-10-17T14:00:00Z | 7.11 pp | 8.56 pp | $10.00 | accepted |
| Manchester City · Ipswich Town2026-10-17T14:00:00Z | 10.25 pp | 8.56 pp | $0.00 | high uncertainty |
| Newcastle United · Aston Villa2026-10-17T16:30:00Z | 7.22 pp | 8.56 pp | $10.00 | accepted |
| Leeds United · Manchester United2026-10-18T13:00:00Z | 6.94 pp | 8.56 pp | $10.00 | accepted |
| Brighton and Hove Albion · Crystal Palace2026-10-18T13:00:00Z | 9.33 pp | 8.56 pp | $0.00 | high uncertainty |
| Bournemouth · Sunderland2026-10-18T13:00:00Z | 7.34 pp | 8.56 pp | $10.00 | accepted |
| Nottingham Forest · Arsenal2026-10-18T15:30:00Z | 7.36 pp | 8.56 pp | $10.00 | accepted |
| Tottenham Hotspur · Coventry City2026-10-19T19:00:00Z | 8.91 pp | 8.56 pp | $0.00 | high uncertainty |
| Ipswich Town · Nottingham Forest2026-10-23T19:00:00Z | 7.10 pp | 8.56 pp | $10.00 | accepted |
| Aston Villa · Manchester City2026-10-24T11:30:00Z | 7.73 pp | 8.56 pp | $10.00 | accepted |
| Coventry City · Fulham2026-10-24T14:00:00Z | 7.14 pp | 8.56 pp | $10.00 | accepted |
| Arsenal · Everton2026-10-24T14:00:00Z | 9.17 pp | 8.56 pp | $0.00 | high uncertainty |
| Chelsea · Tottenham Hotspur2026-10-24T16:30:00Z | 8.00 pp | 8.56 pp | $10.00 | accepted |
| Liverpool · Brighton and Hove Albion2026-10-25T14:00:00Z | 7.59 pp | 8.56 pp | $10.00 | accepted |
| Manchester United · Bournemouth2026-10-25T14:00:00Z | 7.11 pp | 8.56 pp | $10.00 | accepted |
| Hull City · Brentford2026-10-25T14:00:00Z | 7.49 pp | 8.56 pp | $10.00 | accepted |
| Crystal Palace · Newcastle United2026-10-25T14:00:00Z | 7.17 pp | 8.56 pp | $10.00 | accepted |
| Sunderland · Leeds United2026-10-25T16:30:00Z | 6.87 pp | 8.56 pp | $10.00 | accepted |
| Chelsea · Manchester United2026-10-31T12:30:00Z | 7.52 pp | 8.56 pp | $10.00 | accepted |
| Hull City · Ipswich Town2026-10-31T15:00:00Z | 6.98 pp | 8.56 pp | $10.00 | accepted |
| Manchester City · Brighton and Hove Albion2026-10-31T15:00:00Z | 8.46 pp | 8.56 pp | $10.00 | accepted |
| Bournemouth · Leeds United2026-10-31T15:00:00Z | 7.43 pp | 8.56 pp | $10.00 | accepted |
| Brentford · Nottingham Forest2026-10-31T15:00:00Z | 7.52 pp | 8.56 pp | $10.00 | accepted |
| Coventry City · Sunderland2026-10-31T15:00:00Z | 7.01 pp | 8.56 pp | $10.00 | accepted |
| Tottenham Hotspur · Crystal Palace2026-10-31T17:30:00Z | 8.50 pp | 8.56 pp | $10.00 | accepted |
| Aston Villa · Fulham2026-10-31T20:00:00Z | 7.64 pp | 8.56 pp | $10.00 | accepted |
| Liverpool · Arsenal2026-11-01T16:30:00Z | 6.99 pp | 8.56 pp | $10.00 | accepted |
| Newcastle United · Everton2026-11-02T20:00:00Z | 7.35 pp | 8.56 pp | $10.00 | accepted |
| Everton · Coventry City2026-11-06T20:00:00Z | 8.46 pp | 8.56 pp | $10.00 | accepted |
| Leeds United · Tottenham Hotspur2026-11-07T12:30:00Z | 6.98 pp | 8.56 pp | $10.00 | accepted |
| Fulham · Newcastle United2026-11-07T15:00:00Z | 7.63 pp | 8.56 pp | $10.00 | accepted |
| Arsenal · Hull City2026-11-07T15:00:00Z | 9.85 pp | 8.56 pp | $0.00 | high uncertainty |
| Nottingham Forest · Manchester City2026-11-07T17:30:00Z | 7.51 pp | 8.56 pp | $10.00 | accepted |
| Crystal Palace · Liverpool2026-11-08T14:00:00Z | 7.68 pp | 8.56 pp | $10.00 | accepted |
| Sunderland · Chelsea2026-11-08T14:00:00Z | 7.33 pp | 8.56 pp | $10.00 | accepted |
| Ipswich Town · Bournemouth2026-11-08T14:00:00Z | 7.34 pp | 8.56 pp | $10.00 | accepted |
| Brighton and Hove Albion · Brentford2026-11-08T14:00:00Z | 8.00 pp | 8.56 pp | $10.00 | accepted |
| Manchester United · Aston Villa2026-11-08T16:30:00Z | 7.47 pp | 8.56 pp | $10.00 | accepted |
| Manchester City · Fulham2026-11-21T12:30:00Z | 9.74 pp | 8.56 pp | $0.00 | high uncertainty |
| Coventry City · Crystal Palace2026-11-21T15:00:00Z | 6.93 pp | 8.56 pp | $10.00 | accepted |
| Chelsea · Leeds United2026-11-21T15:00:00Z | 7.61 pp | 8.56 pp | $10.00 | accepted |
| Aston Villa · Sunderland2026-11-21T15:00:00Z | 8.17 pp | 8.56 pp | $10.00 | accepted |
| Bournemouth · Nottingham Forest2026-11-21T15:00:00Z | 7.71 pp | 8.56 pp | $10.00 | accepted |
| Tottenham Hotspur · Ipswich Town2026-11-21T15:00:00Z | 9.10 pp | 8.56 pp | $0.00 | high uncertainty |
| Newcastle United · Arsenal2026-11-21T17:30:00Z | 7.45 pp | 8.56 pp | $10.00 | accepted |
| Hull City · Brighton and Hove Albion2026-11-22T14:00:00Z | 7.50 pp | 8.56 pp | $10.00 | accepted |
| Liverpool · Manchester United2026-11-22T16:30:00Z | 8.32 pp | 8.56 pp | $10.00 | accepted |
No matching saved games.
All outcome forecasts 6 models · common-game scores
Columns
| DariusOutcome forecast from prior team performance. | 0.6362 | 66% · 55/83 | 0.4446 | 83 |
| ArtabanTeam-strength ratings built from past results. | 0.6398 | 64% · 53/83 | 0.4477 | 83 |
| AnahitaPast results and injuries, selecting fewer signals. | 0.6100 | 71% · 59/83 | 0.4206 | 83 |
| RoxanaPast results and injuries, balancing signal weights. | 0.6096 | 70% · 58/83 | 0.4204 | 83 |
| TiridatesDecision trees combining performance signals. | 0.6170 | 72% · 60/83 | 0.4278 | 83 |
| BahramSportsbook consensus with the margin removed. | 0.5727 | 72% · 60/83 | 0.3848 | 83 |
No matching forecasts.
All 6 outcome models use the same 83 finished games. Artaban, Anahita, Roxana and Tiridates are statistical comparisons. Brier is available under Columns.
Every scored NFL game loads here.
Paper replays by forecast Reconstructed from saved inputs
Columns
| Darius · Across booksOutcome forecast from prior team performance. | Replay | −$157.60 | -15.8% | 20.3% | 70 | — |
| Artaban · Across booksTeam-strength ratings built from past results. | Replay | −$61.72 | -6.2% | 13.3% | 72 | — |
| Anahita · Across booksPast results and injuries, selecting fewer signals. | Replay | −$140.10 | -14.0% | 18.9% | 73 | — |
| Roxana · Across booksPast results and injuries, balancing signal weights. | Replay | −$128.38 | -12.8% | 18.7% | 73 | — |
| Tiridates · Across booksDecision trees combining performance signals. | Replay | −$161.47 | -16.1% | 22.2% | 70 | — |
| Bahram · Across booksSportsbook consensus with the margin removed. | Replay | −$40.77 | -4.1% | 5.6% | 7 | — |
No scored methods in this group.
These were worked out afterwards from pre-kickoff forecasts and prices. Every portfolio starts with $1,000 under the frozen record allocation rules, while game coverage can differ. They are separate from saved decisions and the annual Guard rules.
Show curves Optional comparison
Choose the lines to display.
Paper balance, starting at $1,000. Overlapping curves can reflect identical decisions. These are the recorded values, without smoothing.
Sportsbook probability scores 86 common games · separate cohort
Columns
| Fliff | 0.5811 | 86 |
| Hard Rock Bet (OH) | 0.5816 | 86 |
| Hard Rock Bet | 0.5819 | 86 |
| Hard Rock Bet (FL) | 0.5819 | 86 |
| Fanatics | 0.5821 | 86 |
| FanDuel | 0.5824 | 86 |
| ReBet | 0.5826 | 86 |
| BetMGM | 0.5827 | 86 |
| Bahram · Market consensusSportsbook consensus with the margin removed. | 0.5834 | 86 |
| theScore Bet | 0.5839 | 86 |
| Caesars | 0.5849 | 86 |
| Courtside | 0.5850 | 86 |
| DraftKings | 0.5855 | 86 |
| Bally Bet | 0.5858 | 86 |
| BetRivers | 0.5858 | 86 |
| betPARX | 0.5859 | 86 |
No matching forecasts.
Offshore sites not licensed in any US state (Bovada, MyBookie and similar) are left out of this table.
Every row in this table uses the same 86 games, with the last saved pre-kickoff price and margin removed. This cohort differs from the outcome-model comparison; scores across the two tables should not be ranked together.
All outcome forecasts 6 models · common-game scores
Columns
| DariusOutcome forecast from prior team performance. | 0.6692 | 60% · 154/258 | 0.4763 | 258 |
| ArtabanTeam-strength ratings built from past results. | 0.6652 | 61% · 158/258 | 0.4724 | 258 |
| AnahitaPast results and injuries, selecting fewer signals. | 0.6648 | 61% · 158/258 | 0.4721 | 258 |
| RoxanaPast results and injuries, balancing signal weights. | 0.6650 | 61% · 158/258 | 0.4722 | 258 |
| TiridatesDecision trees combining performance signals. | 0.6664 | 60% · 156/258 | 0.4737 | 258 |
| BahramSportsbook consensus with the margin removed. | 0.6553 | 60% · 154/258 | 0.4631 | 258 |
No matching forecasts.
All 6 outcome models use the same 258 finished games. Artaban, Anahita, Roxana and Tiridates are statistical comparisons. Brier is available under Columns.
Every scored NHL game loads here.
Paper replays by forecast Reconstructed from saved inputs
Columns
| Darius · Across booksOutcome forecast from prior team performance. | Replay | −$253.81 | -25.4% | 33.3% | 190 | — |
| Artaban · Across booksTeam-strength ratings built from past results. | Replay | −$329.95 | -33.0% | 35.7% | 181 | — |
| Anahita · Across booksPast results and injuries, selecting fewer signals. | Replay | −$303.10 | -30.3% | 31.2% | 176 | — |
| Roxana · Across booksPast results and injuries, balancing signal weights. | Replay | −$303.95 | -30.4% | 32.1% | 179 | — |
| Tiridates · Across booksDecision trees combining performance signals. | Replay | −$358.17 | -35.8% | 40.0% | 196 | — |
| Bahram · Across booksSportsbook consensus with the margin removed. | Replay | $15.03 | +1.5% | 1.8% | 4 | — |
No scored methods in this group.
These were worked out afterwards from pre-kickoff forecasts and prices. Every portfolio starts with $1,000 under the frozen record allocation rules, while game coverage can differ. They are separate from saved decisions and the annual Guard rules.
Show curves Optional comparison
Choose the lines to display.
Paper balance, starting at $1,000. Overlapping curves can reflect identical decisions. These are the recorded values, without smoothing.
Sportsbook probability scores 317 common games · separate cohort
Columns
| FanDuel | 0.6555 | 317 |
| Fanatics | 0.6556 | 317 |
| BetRivers | 0.6558 | 317 |
| betPARX | 0.6559 | 317 |
| Bally Bet | 0.6559 | 317 |
| BetMGM | 0.6561 | 317 |
| Bahram · Market consensusSportsbook consensus with the margin removed. | 0.6561 | 317 |
| Caesars | 0.6564 | 317 |
| theScore Bet | 0.6564 | 317 |
| Courtside | 0.6567 | 317 |
| Hard Rock Bet | 0.6567 | 317 |
| DraftKings | 0.6568 | 317 |
| ReBet | 0.6569 | 317 |
No matching forecasts.
Offshore sites not licensed in any US state (Bovada, MyBookie and similar) are left out of this table.
Every row in this table uses the same 317 games, with the last saved pre-kickoff price and margin removed. This cohort differs from the outcome-model comparison; scores across the two tables should not be ranked together.
All outcome forecasts 6 models · common-game scores
Columns
| DariusOutcome forecast from prior team performance. | 0.6156 | 69% · 164/239 | 0.4276 | 239 |
| ArtabanTeam-strength ratings built from past results. | 0.6082 | 69% · 164/239 | 0.4206 | 239 |
| AnahitaPast results and injuries, selecting fewer signals. | 0.6088 | 69% · 166/239 | 0.4217 | 239 |
| RoxanaPast results and injuries, balancing signal weights. | 0.6112 | 69% · 165/239 | 0.4240 | 239 |
| TiridatesDecision trees combining performance signals. | 0.6123 | 69% · 165/239 | 0.4232 | 239 |
| BahramSportsbook consensus with the margin removed. | 0.5712 | 71% · 169/239 | 0.3889 | 239 |
No matching forecasts.
All 6 outcome models use the same 239 finished games. Artaban, Anahita, Roxana and Tiridates are statistical comparisons. Brier is available under Columns.
Every scored NBA game loads here.
Paper replays by forecast Reconstructed from saved inputs
Columns
| Darius · Across booksOutcome forecast from prior team performance. | Replay | −$378.02 | -37.8% | 43.7% | 212 | — |
| Artaban · Across booksTeam-strength ratings built from past results. | Replay | −$303.58 | -30.4% | 44.3% | 214 | — |
| Anahita · Across booksPast results and injuries, selecting fewer signals. | Replay | −$418.58 | -41.9% | 47.1% | 210 | — |
| Roxana · Across booksPast results and injuries, balancing signal weights. | Replay | −$440.29 | -44.0% | 49.2% | 209 | — |
| Tiridates · Across booksDecision trees combining performance signals. | Replay | −$449.59 | -45.0% | 49.7% | 195 | — |
| Bahram · Across booksSportsbook consensus with the margin removed. | Replay | −$17.14 | -1.7% | 18.5% | 50 | — |
No scored methods in this group.
These were worked out afterwards from pre-kickoff forecasts and prices. Every portfolio starts with $1,000 under the frozen record allocation rules, while game coverage can differ. They are separate from saved decisions and the annual Guard rules.
Show curves Optional comparison
Choose the lines to display.
Paper balance, starting at $1,000. Overlapping curves can reflect identical decisions. These are the recorded values, without smoothing.
Sportsbook probability scores 179 common games · separate cohort
Columns
| Fliff | 0.5721 | 179 |
| Tipico | 0.5732 | 179 |
| theScore Bet | 0.5733 | 179 |
| 1xBet | 0.5741 | 179 |
| Hard Rock Bet | 0.5742 | 179 |
| PointsBet (AU) | 0.5742 | 179 |
| William Hill | 0.5743 | 179 |
| Virgin Bet | 0.5744 | 179 |
| Bally Bet | 0.5744 | 179 |
| Neds | 0.5745 | 179 |
| Unibet | 0.5745 | 179 |
| LeoVegas (SE) | 0.5745 | 179 |
| SportsBet | 0.5746 | 179 |
| betPARX | 0.5746 | 179 |
| Marathon Bet | 0.5746 | 179 |
| GTbets | 0.5746 | 179 |
| Unibet (SE) | 0.5746 | 179 |
| Paddy Power | 0.5746 | 179 |
| Unibet (NL) | 0.5746 | 179 |
| LiveScore Bet | 0.5747 | 179 |
| TABtouch | 0.5747 | 179 |
| DraftKings | 0.5748 | 179 |
| PMU (FR) | 0.5748 | 179 |
| Courtside | 0.5749 | 179 |
| Bahram · Market consensusSportsbook consensus with the margin removed. | 0.5752 | 179 |
| BetRivers | 0.5753 | 179 |
| Sky Bet | 0.5753 | 179 |
| Betr | 0.5753 | 179 |
| Ladbrokes | 0.5754 | 179 |
| Unibet (UK) | 0.5754 | 179 |
| Pinnacle | 0.5755 | 179 |
| BetMGM | 0.5755 | 179 |
| Caesars | 0.5757 | 179 |
| Fanatics | 0.5757 | 179 |
| Betano (UK) | 0.5759 | 179 |
| Coral | 0.5759 | 179 |
| TAB | 0.5760 | 179 |
| Winamax (DE) | 0.5762 | 179 |
| ReBet | 0.5763 | 179 |
| Dabble AU | 0.5763 | 179 |
| Codere (IT) | 0.5764 | 179 |
| FanDuel | 0.5764 | 179 |
| Bet Victor | 0.5767 | 179 |
| BoyleSports | 0.5768 | 179 |
| Picklebet (AU) | 0.5771 | 179 |
| Betsson | 0.5773 | 179 |
| Nordic Bet | 0.5773 | 179 |
| Unibet (FR) | 0.5784 | 179 |
| Winamax (FR) | 0.5786 | 179 |
| LeoVegas | 0.5799 | 179 |
No matching forecasts.
Offshore sites not licensed in any US state (Bovada, MyBookie and similar) are left out of this table.
Every row in this table uses the same 179 games, with the last saved pre-kickoff price and margin removed. This cohort differs from the outcome-model comparison; scores across the two tables should not be ranked together.
Method, score definitions and limitations
Correct counts the most likely home, draw or away outcome. Log loss is average negative log likelihood; lower is better. Brier is the mean sum of squared probability errors over the possible outcomes; lower is better. Ties use home, draw, away order.
A saved decision was fixed before kickoff. A replay is reconstructed afterwards from saved pre-kickoff inputs. Guard filters Arta decisions rather than creating a new probability forecast. All amounts are paper results; no real wagers, fees or taxes are included.
Recent inputs use recorded sportsbook prices. The historical research uses UK source books and limited early-season price coverage. Historical and synthetic results are separate evidence. The small recent record does not establish future profitability.
Open the complete forecast ledger (free account) · Download the weekly record (CSV) (free account): every finished game our forecast was saved for, by week and sport. Its model_brier and model_log_loss columns cover only the games that also had a saved sportsbook price; model_brier_all and model_log_loss_all cover every game. The weekly record scores each game’s last saved percentage, which is kept to one decimal place of a percent; the forecast rows keep the saved snapshots to five decimals, so the two files can differ from the fourth decimal. In the forecast rows, team names are stored as each source gave them, so some NBA, NFL and NHL rows use abbreviations (BOS) and others full names; and a game’s 24h, 1h and close rows often carry the same forecast, because the model’s inputs are earlier final scores. Both files need a free account.
Research
Ideas we are still testing, on simulated or already-inspected data. Kept apart from the record.
Cash-out research Simulated researchShould you cash out early?
Mitra keeps every bet to the final whistle. Six models try to cash out at the right moment. Same 848 bets of $10 on 1,000 computer-generated games.
Not proven. The extra profit could be luck, and it disappears when sportsbooks update their cash-out offers quickly. Simulated games, not real sportsbook offers.
Showing Mitra, the best and the worst. Tick the others to compare.
Profit on fixed $10 bets, starting at $0, not a $1,000 bankroll. Mitra keeps every bet.
| Model | Profit | vs holding |
|---|---|---|
ShapurWatches how the ticket's book lags the other books and predicts its next price move; exits when the offer has caught up. Method: Lead-lag ridge. Early cash-outs: 578. If offers update faster: +$717 (below holding). | +$1,002 | +$258 |
PerozExits when the offer reaches what the model thinks the ticket is worth. No learned forecast of the next move. Method: Fair-value exit. Early cash-outs: 586. If offers update faster: +$610 (below holding). | +$908 | +$163 |
BoranBlends the learned models into one forecast of the next move. Method: Ensemble. Early cash-outs: 575. If offers update faster: +$619 (below holding). | +$878 | +$133 |
CyrusA small attention network that weighs game state, the ticket's price, peer prices and recent moves together (seed 42; seeds 43 and 44 made $864 and $823). Method: Attention network. Early cash-outs: 534. If offers update faster: +$613 (below holding). | +$874 | +$130 |
KavadTells quiet markets from fast-moving ones and uses a separate lead-lag forecast for each. Method: Two-regime (HMM) lead-lag. Early cash-outs: 560. If offers update faster: +$628 (below holding). | +$867 | +$122 |
Mitra · HoldKeeps every ticket until the game ends. The reference line. Method: Hold. Early cash-outs: 0. If offers update faster: +$745. | +$745 | Reference |
SasanWaits until the other books agree and hold steady, then exits. No model value. Method: Stable consensus exit. Early cash-outs: 409. If offers update faster: +$77 (below holding). | +$227 | −$518 |
vs holding: profit above or below Mitra on the same bets. Tap a name for how it works.
Data needed for real-game evaluation
These methods currently have simulated research results. Real-game exit decisions require the original fitted artifacts and matching inputs.
Mitra · Hold
A matched real-entry cohort shared with the exit methods. The delivered seven-method results use synthetic entries. Output: Hold to settlement.
Shapur
The original selected ridge weights, 29 causal game/quote features, five-minute targets and one-minute decision/recheck histories. Output: Price movement / exit.
Peroz
As-observed game state, validated live probabilities and timestamped exit offers for the same entries. Output: Fair-value exit.
Boran
Validation predictions and selected blend weights for the original target, plus matching causal features and exit observations. Output: Ensemble exit.
Cyrus
The original selected attention weights and 29 causal game/quote features at the research decision resolution. Output: Attention / price movement.
Kavad
The selected HMM/ridge parameters and independent price paths at the original minute-level resolution. Output: Market regime / exit.
Sasan
Three simultaneous peer books over five consecutive one-minute observations and recorded exit offers. Output: Consensus exit.
Assumptions, source identities and reproducibility
- Simulated games, not real prices: generated from a Poisson goal model with assumed book price lags (AR(1), coefficient 0.85).
- Exits use a made-up cash-out offer (5 percent haircut on the simulated book price). Real sportsbooks set their own cash-out offers, and this experiment does not observe them.
- Fixed $10 per ticket, never compounded: this is profit on tickets, not a funded bankroll, and is not comparable with the $1,000 replay.
- Every exit model's 95 percent paired bootstrap range for extra profit over holding includes zero (Shapur / lead-lag: +$0.30 per ticket, range -$0.28 to +$0.86).
- When cash-out offers are repriced promptly (stress test), every exit model ends below holding.
- None of these models has been tested on real minute-by-minute price histories. Names were given after testing and have no effect on results.
Experiment model-family-v1. Names follow the research note; the delivered source keys, weights, entry decisions and profit values are unchanged. Each name stays tied to its original method and source key.
- Mitra · source key: mitra
- Shapur · source key: cyrus
- Peroz · source key: sina
- Boran · source key: rumi
- Cyrus · source key: khayyam
- Kavad · source key: hafez
- Sasan · source key: kavad
Mitra Gold research (early-exit models). Mitra Guard, the separate historical filter, is not in this chart. Source CSV SHA-256: simulated_curves.csv 40cb2272f0129d180790d6c37e9a3d7ac19466c90159d95ba5487ab0e0c0e99d.
More research
Historical replay Three Premier League seasons
2023-08-10 to 2026-05-24. These seasons were already inspected; selecting a method after seeing its result flatters the comparison. UK source-book prices differ from current US apps, and early price coverage is limited.
With Guard
Columns
Mitra GuardMitra Guard: Adds the uncertainty filter to Mitra. | $701.70 | +70.2% | 22.5% | 349 | $1,701.70 |
Mitra Select GuardMitra Select Guard: Adds the uncertainty filter to Select. | $241.54 | +24.2% | 41.5% | 430 | $1,241.54 |
No matching methods.
Without Guard
Columns
MitraMitra: Keeps the reference-book pick and compares prices across books. | $104.15 | +10.4% | 49.7% | 574 | $1,104.15 |
Mitra SelectMitra Select: Chooses the pick using prices across books. | $22.74 | +2.3% | 44.0% | 723 | $1,022.74 |
Arta at Bet365Arta: Forecast-based betting at one reference book. | $427.10 | +42.7% | 34.3% | 627 | $1,427.10 |
Arta at BwinArta: Forecast-based betting at one reference book. | −$117.10 | -11.7% | 48.7% | 474 | $882.90 |
Arta at BetfredArta: Forecast-based betting at one reference book. | −$368.50 | -36.8% | 47.9% | 303 | $631.50 |
Arta at BetMGMArta: Forecast-based betting at one reference book. | −$106.53 | -10.7% | 43.4% | 307 | $893.47 |
Arta at PinnacleArta: Forecast-based betting at one reference book. | −$842.00 | -84.2% | 90.5% | 694 | $158.00 |
Arta at William HillArta: Forecast-based betting at one reference book. | −$432.45 | -43.2% | 57.4% | 232 | $567.55 |
Arta at VC BetArta: Forecast-based betting at one reference book. | −$339.90 | -34.0% | 37.3% | 175 | $660.10 |
Arta at InterwettenArta: Forecast-based betting at one reference book. | −$29.00 | -2.9% | 36.6% | 96 | $971.00 |
Arta at Betfair SportsbookArta: Forecast-based betting at one reference book. | −$254.70 | -25.5% | 40.3% | 173 | $745.30 |
Arta at 1xBetArta: Forecast-based betting at one reference book. | −$352.50 | -35.3% | 43.6% | 235 | $647.50 |
Arta at Betfair ExchangeArta: Forecast-based betting at one reference book. | $204.59 | +20.5% | 57.8% | 723 | $1,204.59 |
Arta at BetVictorArta: Forecast-based betting at one reference book. | −$35.00 | -3.5% | 28.0% | 254 | $965.00 |
Arta at CoralArta: Forecast-based betting at one reference book. | $123.80 | +12.4% | 23.0% | 188 | $1,123.80 |
Arta at LadbrokesArta: Forecast-based betting at one reference book. | $147.60 | +14.8% | 12.8% | 148 | $1,147.60 |
No matching methods.
Show curves Optional comparison
Choose the lines to display.
Paper balance, starting at $1,000. Overlapping curves can reflect identical decisions. These are the recorded values, without smoothing.
Training splits and historical rules
- Three Premier League seasons, 2023/24 to 2025/26, replayed with one $1,000 paper portfolio ($250 at each of four source books), never topped up.
- Before each season the Arta forecast was refitted using only the two seasons before it (documented NumPy implementation). The fold forecasts were reused as delivered, not retrained for this export.
- Mitra takes Arta's pick at Bet365's prices, then the best funded price across the books. Mitra Select picks across all the books.
- Guard (Mitra Guard 1.0, filter mode) is the instability filter: it skips a game when Arta's bootstrap uncertainty is above the 75th percentile of the training seasons (cutoff frozen before each season).
- Stakes: $10 or 2% per match, 10% daily and 25% per-account caps, no transfers between books.
- Source books: Bet365, Bwin, Betfred and BetMGM (provider market), from football-data.co.uk. Not US app prices. The first two seasons have limited price coverage.
- No real bets, no verified fills or quote timestamps, no fees or taxes. The data had been inspected before. Both filters' ROI intervals include losses.
Historical probability scores Tuning and retrospective test seasons
2023/24
Used for model tuning.
Columns
| Pinnacle | 0.9269 | 196 |
| Bahram · Market consensusSportsbook consensus with the margin removed. | 0.9281 | 196 |
| Bet365 | 0.9283 | 196 |
| VC Bet | 0.9297 | 196 |
| Bwin | 0.9299 | 196 |
| Interwetten | 0.9303 | 196 |
| William Hill | 0.9307 | 196 |
| Arta 2.1Market probabilities adjusted using past results. | 0.9362 | 196 |
No matching forecasts.
2024/25
Used for model tuning.
Columns
| Arta 2.1Market probabilities adjusted using past results. | 0.9718 | 230 |
| William Hill | 0.9758 | 230 |
| Bet365 | 0.9762 | 230 |
| Betfair Exchange | 0.9762 | 230 |
| Betfair Sportsbook | 0.9764 | 230 |
| Pinnacle | 0.9764 | 230 |
| Bahram · Market consensusSportsbook consensus with the margin removed. | 0.9767 | 230 |
| 1xBet | 0.9784 | 230 |
| Bwin | 0.9792 | 230 |
No matching forecasts.
2025/26
Previously inspected retrospective test; not a new forward record.
Columns
| BetMGM (provider market) | 0.9859 | 210 |
| Betfred | 0.9865 | 210 |
| BetVictor | 0.9876 | 210 |
| Bahram · Market consensusSportsbook consensus with the margin removed. | 0.9878 | 210 |
| Betfair Exchange | 0.9880 | 210 |
| Pinnacle | 0.9888 | 210 |
| Bwin | 0.9890 | 210 |
| Ladbrokes | 0.9891 | 210 |
| Coral | 0.9891 | 210 |
| Bet365 | 0.9910 | 210 |
| Arta 2.1Market probabilities adjusted using past results. | 0.9919 | 210 |
No matching forecasts.

How the record works
We keep score in public.
Saved before kickoff.
Each forecast is saved before the game begins and only ever added to, so nothing can be adjusted after the fact.
- 4sports: Premier League, NBA, NFL and NHL
- 1record, each sport scored on its own games
- Add-onlya saved forecast is never edited
Sports are not ranked against each other.
Scored after the final.
Once the game ends, every forecast is scored against the official final and added to the record. We score our forecasts against the sportsbooks’ prices with the margin removed, on the same games.
Score definitions, cohorts and limits sit beside every table, and with a free account the saved rows download as CSV: the forecast benchmark and the weekly record.
Shown win or lose.
Every scored game stays on the record, including the forecasts that missed. With a free account you can open the full record and download the saved rows. Saved decisions and replays worked out afterwards are labelled and kept apart.
What the record shows so far: our testing has not yet established a reliable forecasting advantage over market prices. We publish that openly, because a second opinion is only useful if you know how much to trust it.
What the numbers show, and what they don’tUse the record wisely.
Model estimates can help you think about a game, but they are never a guarantee or betting advice. All amounts here are paper results: no real bets are placed.
- Check the record before trusting a number
- Compare prices across sportsbooks
- Play responsibly: 21+ where required (18+ in the UK, Europe and Australia)
Create your free Scoutics account today
Start free with every app’s price and our forecast for the NFL, NBA, NHL and Premier League, plus two free simulation runs. Upgrade when you want more simulations and research tools. Plans pay for software access, never stakes.