The bar is the model's probability; the caret on the same scale is the market's, with the margin removed. The distance between them is the disagreement.
OutcomeModel fill · market caretProbFairΔ mkt
CHE
55.8%1.79−22.7 pp
Draw
23.5%4.26+9.3 pp
HUL
20.7%4.84+13.4 pp
Ch 02 · Book
27 markets
dixon_coles xi001-20260903
Model fair prices, computed as 1 / probability from a Dixon-Coles score matrix. No bookmaker margin is included, and no bookmaker is quoting these. A real price will be shorter than every number here.
Thin sample
Hull City has only 2 completed matches in the training window, so its fitted strength is imprecise. Treat the long prices in this book as arithmetic rather than information: a real bookmaker would quote a much shorter price on the outsider.
Matches behind each fit: Chelsea 149 · Hull City 2
Match result
Chelsea55.8%1.79
Draw23.5%4.26
Hull City20.7%4.84
Double chance
Chelsea or draw79.3%1.26
Chelsea or Hull City76.5%1.31
Draw or Hull City44.2%2.27
Draw no bet
Chelsea73.0%1.37
Hull City27.0%3.70
A draw returns the stake, so these prices exclude it.
Win to nil
Chelsea29.5%3.39
Hull City10.0%9.96
Clean sheet
Chelsea35.4%2.83
Hull City15.9%6.29
Ch 03 · Reasoning
What is driving this
computed, not generated
Computed arithmetic, not generated text. Each driver is a stored quantity: the Elo difference replayed from results, points per game over the last five matches, rolling expected goals, and each absent player's share of team expected goals.
Underlying expected goals+0.34 net xG per match
Favours CHE. CHE +0.14, HUL -0.20 across the form window — results are noisier than the chances behind them
Home advantage+60 Elo equivalent
Favours CHE. the football default is around 100 Elo points, but Premier League home advantage fell during the crowdless 2020/21 season and settled lower afterwards, so the ratings carry a smaller term
Team rating gap+54 Elo
Favours CHE. CHE 1577 against HUL 1523, computed from stored results; an Elo-driven model scores RPS around 0.2015 on this league against roughly 0.194 for de-vigged closing odds
Recent form2.00 against 2.00 points per game
last 5 and 5 completed matches: WWLWD against WWWLD
Unavailable players0% home against 0% away xG missing
0 out and 3 doubtful for CHE, 6 out and 2 doubtful for HUL; measured as the share of recent expected goals those players contributed
Ch 04 · Comparison
Both sides
CHE
WWLWD
9 May – 30 Aug
HUL · incl. Championship
DLWWW
21 Apr – 29 Aug
2.00Points per game2.00
11Goals for8
9Goals against5
9.02Expected goals (2 of 5)1.96
1,577Elo1,523
Ch 05 · Market
Bookmaker prices
Real bookmaker prices, collected from published market data.
Market
Selection
Best
Book
Implied
N
h2h
away
13.40
1xBet
7.5%
33
h2h
draw
7.00
Betfair
14.3%
33
h2h
home
1.28
1xBet
78.1%
33
Ch 06 · Availability
Chelsea
0.0% xG out
Share of attacking output is what moves the number, so it is shown rather than implied.
CaicedoMIDdoubtful75%
Unspecified injury - 75% chance of playing
HendersonMIDdoubtful75%
Wrist injury - 75% chance of playing
PalestraDEFdoubtful75%
Unspecified injury - 75% chance of playing
Ch 06 · Availability
Hull City
0.0% xG out
Share of attacking output is what moves the number, so it is shown rather than implied.
ButlandGKPinjured0%
Arm injury - Unknown return date
GelhardtMIDinjured0%
Ankle injury - Expected back 5 Sep
GyabiMIDinjured0%
Thigh injury - Unknown return date
HughesDEFinjured0%
Groin injury - Unknown return date
MatazoMIDinjured0%
Knee injury - Unknown return date
ZambranoMIDinjured0%
Thigh injury - Unknown return date
CoyleDEFdoubtful75%
Unspecified injury - 75% chance of playing
CrooksMIDdoubtful75%
Thigh injury - 75% chance of playing
Ch 07 · Simulation
Re-price without a player
Removing a player scales that side's expected goals by their share of recent attacking output, then rebuilds the score matrix. Every market moves together.
Ch 10 · Calibration
How this model actually scores
Dixon-Coles bivariate Poisson fitted to Premier League results with exponential time decay. Statistical model, no language model involved.
loses to closing market2,281 out-of-sample fixtures
Metric
Model
Market
Δ
Ranked probability score
0.2047
0.1965
+0.0082
Log loss
0.9882
0.9623
+0.0259
Brier
0.5883
0.5711
+0.0172
Accuracy
52.5%
54.8%
−2.3 pp
Lower is better for the first three. These probabilities are for reading a fixture, not for expecting profit against closing prices. What these numbers mean