Live · Model onlinev0.1 · Dixon-Coles
Generated May 11, 2026
Final Whistle · Issue 01

BACK/TEST

EPL 2024–25
League
EPL · E0
Method
Walk-forward
Window
6 seasons
Model
dc-xi0.0019-mg10-2024-25
§01 · The finding

Read this part.
Not the chart.

The model placed 521 bets at a yield of +1.13%. Across 521 settled picks the 95% Wilson interval is 33.94%42.25% — it brackets enough outcomes both clearly profitable and clearly losing that the headline is noise, not signal. The interesting result is below, inside the edge bands: the model's most confident picks lost the most money.

§02 · Scoreboard
Picks
521
198 W323 L
Win Rate
38.0%
95% CI 33.9% – 42.2%
Yield
+1.13%
flat 521u staked
P / L
+5.88u
net units · season-to-date
CLV
0.00%
degenerate vs closing
§03 · Equity curve

A bumpy flat line.cumulative P/L · 1u per pick

-20.0u-10.0u0.0u+10.0uAUG 16OCT 26DEC 16FEB 12APR 13MAY 25PEAK · +7.3uTROUGH · -22.8uCLOSE +5.88u
§04 · The finding

HIGHER EDGE, lower yield.

If the picks contained real signal, profitability would scale with conviction. It doesn't. The model's most confident bets (15%+ edge) lose the most.

This is the over-confidence pattern Pinnacle's closing line is designed to enforce against amateur models. The model has not yet beaten it.

-25%-20%-15%-10%-5%0%+5%+10%+15%+20%+25%NO PICKS1–3%+19.4%3–5%175 PICKSWR 42%-7.9%5–10%241 PICKSWR 37%-3.4%10–15%75 PICKSWR 33%-21.9%15%+30 PICKSWR 30%-21.9% ON 30 PICKSEDGE BUCKET · MODEL PROBABILITY − DE-VIGGED IMPLIED
1–3%
0 picks · WR
CI
3–5%+33.91u
175 picks · WR 42%
CI 35%–50%
5–10%-18.92u
241 picks · WR 37%
CI 31%–44%
10–15%-2.54u
75 picks · WR 33%
CI 24%–45%
15%+-6.57u
30 picks · WR 30%
CI 17%–48%
§05 · By market

1X2 is where it fails; totals is where it accidentally lives.

MarketPicksWin rateYieldP / LBrierLog lossN scored
1X2match result29831.2%-0.57%-1.70u0.19700.57841137
O/U 2.5total goals22347.1%+3.40%+7.58u0.24360.6803758
Informational — no picks made on these markets in v1.
BTTSboth teams to score0.25000.6985758
§06 · Ledger
Filter500/500
DateFixture·Mkt·PickOdds·Edge·ResultP / L·
May 25WolvesvBrentfordO/U 2.5Underunder2.634.9%won+1.63u
May 25TottenhamvBrightonO/U 2.5Overover1.493.7%won+0.49u
May 25TottenhamvBrighton1X2Hhome3.8320.8%lost-1.00u
May 25SouthamptonvArsenalO/U 2.5Underunder2.997.5%lost-1.00u
May 25SouthamptonvArsenal1X2Aaway1.254.1%won+0.25u
May 25Nott'm ForestvChelsea1X2Hhome3.504.9%lost-1.00u
May 25NewcastlevEvertonO/U 2.5Underunder2.7516.6%won+1.75u
May 25NewcastlevEverton1X2Aaway10.3610.5%won+9.36u
May 25NewcastlevEverton1X2Ddraw6.198.5%lost-1.00u
May 25Man UnitedvAston VillaO/U 2.5Underunder2.448.8%won+1.44u
May 25Man UnitedvAston Villa1X2Ddraw4.373.5%lost-1.00u
May 25Man UnitedvAston Villa1X2Hhome4.8214.4%won+3.82u
May 25LiverpoolvCrystal PalaceO/U 2.5Underunder3.4811.1%won+2.48u
May 25LiverpoolvCrystal Palace1X2Ddraw6.043.3%won+5.04u
May 25IpswichvWest Ham1X2Aaway2.237.0%won+1.23u
May 25FulhamvMan City1X2Hhome6.324.3%lost-1.00u
May 25BournemouthvLeicesterO/U 2.5Underunder2.678.6%won+1.67u
May 25BournemouthvLeicester1X2Aaway9.018.0%lost-1.00u
May 25BournemouthvLeicester1X2Ddraw5.674.9%lost-1.00u
May 20Man CityvBournemouthO/U 2.5Underunder2.663.6%lost-1.00u
May 20Crystal PalacevWolves1X2Hhome2.417.3%won+1.41u
May 19BrightonvLiverpool1X2Aaway3.0926.9%lost-1.00u
May 18ArsenalvNewcastleO/U 2.5Overover1.743.8%lost-1.00u
May 18ArsenalvNewcastle1X2Hhome1.946.2%won+0.94u
May 18LeicestervIpswichO/U 2.5Underunder2.303.6%won+1.30u
page 1 / 20 · 1–25
§07 · Methodology

For the curious.
Skip if you trust us.

For every calendar date in the test season, a Dixon-Coles model is refit on every completed match strictly prior to that date — including completed fixtures within the test season itself. No future data leaks into training. The fit uses exponential time-decay weighting (xi = 0.0019, calibrated to give recent fixtures roughly 3.5× the weight of matches eighteen months ago).

Each match emits an 11 × 11 score distribution which is then collapsed into 1X2 and Over/Under 2.5 market probabilities. Edge is computed against the de-vigged closing line — Pinnacle preferred, Bet365 as fallback. A pick is taken when edge ≥ 3% and the stake is 1u. BTTS is reported as informational only because football-data.co.uk does not carry a consistent BTTS odds feed.

Closing-line value is zero by construction here because edge already uses the closing price. The column is preserved against the day this harness is wired to a live odds source. Sample sizes inside the edge buckets are small; treat any band whose 95% Wilson CI crosses the break-even win rate as statistically unproven, regardless of point estimate.