xgEdge

Research

Research note

A dated, frozen reading of the retrospective signal record by the research script: what is failing, why it probably is, what the tests can and cannot say, and what is being tested next.

Edition
Database
A copy of the production database (xg_prodcheck)
As of
Rule
Version 2

This edition is a snapshot. Live figures are on Performance and may have moved since.

Scenario: insufficient

The record's ROI is −12.6% (95% −20.9% to −3.4%), which is negative. Its mean retrospective CLV is +0.16% (95% −0.51% to +0.86%), which is inconclusive. A sign counts only when its interval excludes nought, so no scenario can be named yet: the forward ledger decides it.

What is failing

The retrospective record — every signal under rule version 2, each at its own last evaluation before kickoff — loses money: 1221 settled signals on 758 matches, 366 won, 845 lost and 10 void, −153.863 units on a stake of one. Its ROI is −12.6%, with a 95% interval of −20.9% to −3.4% that lies wholly below nought.

The model gave these signals a mean model probability of 0.397 and a mean model-implied edge of +18.1%; 30.2% won, a calibration gap of −9.5%. On the record's own 1211 decided rows the margin-free market forecast them better than the model did:

The model's and the margin-free market's Brier score and log loss on the record's decided rows
Forecast Rows Brier Log loss
Model 1211 0.200988342 0.587694861
Market, margin-free 1211 0.191130914 0.565159242

So the rows the rule chooses are worse forecasts than the price they are picked against: model 0.200988342 Brier and 0.587694861 log loss, market 0.191130914 and 0.565159242.

Why it is probably failing

The model is overconfident, it does not know more than the market once the price is known, and the rule chooses exactly the rows where the two disagree most. Four statements follow from the validated tests below:

  1. The model's probabilities run too high everywhere, not only on the signals. Across all 8647 settled forecasts the share won falls short of the mean model probability in every tenth of probability, by −4.0% to −14.0%. The bias is in the model, not something the rule creates.
  2. The disagreement with the market is not realised. On the record, the slope of the realised gap on the claimed gap is 0.148 (95% −0.525 to 0.824): indistinguishable from nought and below one (p against one 0.0090). Over every forecast it is −0.095 (−0.362 to 0.175). The realised share of the claimed gap over the record is −0.21 (−0.53 to 0.11).
  3. Adding the model to the market does not improve out-of-sample prediction in any market selection tested; for both teams to score and over 1.5 goals it makes the prediction worse, with intervals wholly above nought.
  4. A selection effect is not demonstrated. Weighted across tenths of probability, signal rows are calibrated −2.0% against other forecasts in the same selections, with an interval of −5.0% to +1.1%. Choosing them amplifies an existing bias at most mildly; it does not visibly create one.

A winner's curse, where the rule selects for model error, is consistent with these figures; they do not prove it. The more economical reading is that the model's disagreement with the market carries no information, so a rule that backs disagreement backs noise and pays the bookmaker's margin. That is what the ROI shows.

Validated results

The decisive tests, declared before the data was read, each with a 95% interval from a bootstrap that resamples matches rather than rows (2,000 replicates, a fixed seed). A row under the sample floor shows its count and no rate.

The selection effect

The calibration gap (share won minus mean model probability) of signal rows against the other forecasts in the same ten market selections. Weighted by the record's rows per tenth, the difference is −2.0% (95% −5.0% to +1.1%, p 0.1729).

Calibration gap of signal rows against other forecasts in the same market selections, by tenth of model probability
Tenth of model probability Signal rows Other forecasts Gap, signal rows Gap, other forecasts Difference 95% interval p
0.0–0.1 under the floor 4 37 — — — — —
0.1–0.2 131 367 −6.1% −4.6% −1.5% −7.7% to +4.9% 0.5967
0.2–0.3 260 523 −4.5% −6.1% +1.6% −4.4% to +7.7% 0.6517
0.3–0.4 268 410 −10.0% −7.4% −2.6% −9.3% to +4.5% 0.4588
0.4–0.5 221 339 −14.3% −6.7% −7.6% −15.6% to +0.4% 0.0760
0.5–0.6 149 314 −11.0% −11.8% +0.8% −8.9% to +11.0% 0.8616
0.6–0.7 124 395 −8.4% −7.9% −0.4% −10.2% to +9.3% 0.8986
0.7–0.8 under the floor 40 456 — — — — —
0.8–0.9 under the floor 14 231 — — — — —
0.9–1.0 under the floor 0 26 — — — — —

The realised share and the slope

Realised share = (share won − mean margin-free market) ÷ (mean model − mean margin-free market). One means the model's disagreement is realised in full; nought means it is noise. Below, the slope of the realised gap on the claimed gap, by ordinary least squares with an intercept.

The realised share of the model's disagreement with the market, by model-implied edge band
Model-implied edge band Rows Model probability Market (margin-free) Outcome rate Realised share 95% interval
5–10% 325 0.464 0.403 0.388 −0.24 −1.08 to 0.62
10–15% 240 0.405 0.336 0.342 0.08 −0.78 to 0.94
15–20% 183 0.373 0.298 0.257 −0.55 −1.38 to 0.27
20–30% 271 0.364 0.272 0.225 −0.52 −1.07 to 0.09
30–40% 192 0.346 0.241 0.260 0.19 −0.37 to 0.79
Total 1211 0.398 0.319 0.302 −0.21 −0.53 to 0.11
Slope of the realised gap on the claimed gap, by population
Population Rows Matches Slope 95% interval p against 1
The record 1211 752 0.148 −0.525 to 0.824 0.0090
All forecasts (decided, complete book) 8566 1639 −0.095 −0.362 to 0.175 0.0010
Other forecasts in the record's market selections 3098 1302 0.062 −0.297 to 0.411 0.0010

Market-conditional information, walk-forward

A logistic fit of the outcome on the margin-free market and the model against one on the market alone, in weekly folds by kickoff: each week is scored only by a fit on matches kicked off before it began. ΔLL is the combined fit's out-of-sample log loss minus the market-only fit's; below nought would mean the model adds information. No interval lies below nought.

Out-of-sample log loss of market plus model against market alone, by market selection
Market selection Rows Folds Scored ΔLL 95% interval
Home win 538 8 430 0.0171 −0.0047 to 0.0404
Draw 1249 7 1012 0.0011 −0.0006 to 0.0026
Away win 533 6 364 −0.0018 −0.0125 to 0.0097
Home team to score 373 6 236 −0.0004 −0.0043 to 0.0033
Away team to score 422 6 273 0.0042 −0.0027 to 0.0112
Both teams to score 718 9 571 0.0053 0.0010 to 0.0098
Home team 2+ goals 347 6 221 0.0124 −0.0015 to 0.0268
Away team 2+ goals 353 6 225 0.0067 −0.0022 to 0.0155
Over 1.5 goals 402 8 270 0.0034 0.0002 to 0.0065
Over 2.5 goals 375 8 256 0.0022 −0.0021 to 0.0070

Three calibrators — Platt, isotonic, and shrinkage toward the market — were fitted on the same folds and are reported in the script's output, not adopted. None beats the market in any market selection; shrinkage stays nearest it, with its fitted weight on the model mostly near nought or below, which says the most useful thing to do with the model's disagreement is discard it.

Calibration of every forecast

Every settled forecast, signal or not, by tenth of model probability. Every tenth is overconfident, and the gap broadly grows with the probability.

Model probability, margin-free market and share won for every settled forecast, by tenth of model probability
Tenth of model probability Rows Model probability Market (margin-free) Won Calibration gap
0.0–0.1 under the floor 46 — — — —
0.1–0.2 637 0.161 0.119 11.7% −4.4%
0.2–0.3 1752 0.259 0.202 20.8% −5.1%
0.3–0.4 1470 0.338 0.274 27.3% −6.5%
0.4–0.5 977 0.451 0.367 35.9% −9.2%
0.5–0.6 1248 0.550 0.459 43.6% −11.4%
0.6–0.7 1164 0.648 0.546 53.1% −11.8%
0.7–0.8 763 0.746 0.644 60.6% −14.0%
0.8–0.9 514 0.843 0.755 70.4% −13.9%
0.9–1.0 under the floor 76 — — — —
Total 8647 0.464 0.387 37.5% −8.9%

Retrospective CLV and the scenario

Each record row's price against the same bookmaker's latest near-kickoff capture of the same outcome. Captures began on 20 October 2026, so 953 of the record's rows (606 matches) predate them. On the 266 rows with a CLV the mean is +0.16% (95% −0.51% to +0.86%); 45 / 169 / 52 are above, exactly at and below nought. The record's price is its last evaluation before kickoff, often close to the capture itself, so most CLV is exactly nought: retrospective CLV cannot name a scenario, and the forward ledger, measured from a signal's first appearance, can.

The record's rows by near-kickoff capture status
Status Rows Matches
Captured 266 152
Short book 0 0
Another bookmaker 0 0
Outside the window 0 0
No capture 2 1
Before captures began 953 606
Total (the record) 1221 758
Rows with a CLV
266 on 152 matches
Mean CLV (95%)
+0.16% (95% −0.51% to +0.86%)
Median CLV
+0.00%
Above / exactly / below nought
45 / 169 / 52
Mean price CLV (95%)
+0.18% (95% −0.49% to +0.86%)
Median minutes, capture to kickoff
60.0

The power method, as a robustness check

Every other figure removes the margin multiplicatively, which spreads it evenly where books load it onto longshots. The power method scores slightly better and widens the model's disagreement a little; no reading in this note turns on the method.

The 1X2 headline under the multiplicative and the power margin removal
Population Rows Model probability Market, multiplicative Market, power Log loss, multiplicative Log loss, power Disagreement, multiplicative Disagreement, power
The record, 1X2 418 0.313 0.251 0.245 0.5110 0.5094 +6.2% +6.8%
All forecasts, 1X2 2320 0.299 0.238 0.230 0.5218 0.5213 +6.1% +6.8%

Exploratory findings

Not a basis for acting. These cuts were declared before the script ran and are reported as they came out. With this many of them, some will look striking by chance; none is a reason to back or avoid a market, a competition or a band.

One-way cuts of the record and a family of two-way cells, all declared in the script before it ran.

By model-implied edge exploratory

Every band loses, and larger model-implied edges do not do better.

The record by model-implied edge band
Model-implied edge Rows Matches Won Calibration gap Units ROI ROI 95% Brier, model Brier, market
5–10% 325 289 38.8% −7.7% −45.80 −14.1% −27.2% to −0.6% 0.2156 0.2111
10–15% 242 219 34.2% −6.3% −5.26 −2.2% −21.2% to +17.2% 0.2100 0.2054
15–20% 187 168 25.7% −11.6% −26.66 −14.3% −36.5% to +7.2% 0.2094 0.1928
20–30% 274 246 22.5% −13.9% −66.62 −24.3% −43.2% to −3.9% 0.1864 0.1665
30–40% 193 171 26.0% −8.5% −9.53 −4.9% −29.5% to +22.4% 0.1775 0.1726
Total 1221 758 30.2% −9.5% −153.86 −12.6% −20.9% to −3.4% 0.2010 0.1911

By odds exploratory

Every band from 1.20 up loses and the market beats the model on Brier in each: there is no sign of a failure confined to longshots.

The record by odds band
Odds Rows Matches Won Calibration gap Units ROI ROI 95% Brier, model Brier, market
under 1.20 under the floor 0 0 — — — — — — —
1.20–1.50 under the floor 25 25 — — — — — — —
1.50–2.00 168 158 53.0% −12.1% −15.66 −9.3% −22.6% to +3.3% 0.2664 0.2477
2.00–3.00 336 279 36.0% −12.8% −51.35 −15.3% −27.0% to −2.6% 0.2420 0.2270
3.00–5.00 425 374 24.0% −9.0% −52.02 −12.2% −27.5% to +3.2% 0.1890 0.1819
5.00+ 267 241 14.8% −4.5% −33.39 −12.5% −37.3% to +16.4% 0.1242 0.1222
Total 1221 758 30.2% −9.5% −153.86 −12.6% −20.9% to −3.4% 0.2010 0.1911

By market selection exploratory

The record holds 1X2 home and away and the team-goals selections only. Away team at most 1 goal is the one selection with a positive ROI and the one where the model beats the market on Brier; its interval spans nought.

The record by market selection
Market selection Rows Matches Won Calibration gap Units ROI ROI 95% Brier, model Brier, market
Home win 221 221 24.3% −11.5% −56.82 −25.7% −44.0% to −7.0% 0.1860 0.1758
Draw under the floor 0 0 — — — — — — —
Away win 202 202 22.0% −4.3% −15.62 −7.7% −33.1% to +19.3% 0.1619 0.1593
Home team to score under the floor 18 18 — — — — — — —
Home team not to score under the floor 85 85 — — — — — — —
Away team to score under the floor 38 38 — — — — — — —
Away team not to score under the floor 91 91 — — — — — — —
Both teams to score under the floor 0 0 — — — — — — —
Both teams not to score under the floor 0 0 — — — — — — —
Home team 2+ goals 149 149 29.7% −16.1% −37.77 −25.3% −43.5% to −5.7% 0.2254 0.2035
Home team at most 1 goal 130 130 42.3% −9.1% −8.83 −6.8% −26.7% to +13.2% 0.2344 0.2259
Away team 2+ goals 166 166 29.1% −9.4% −11.59 −7.0% −30.6% to +17.1% 0.2142 0.2035
Away team at most 1 goal 121 121 54.5% −5.6% +4.84 +4.0% −14.0% to +22.9% 0.2391 0.2409
Over 1.5 goals under the floor 0 0 — — — — — — —
Under 1.5 goals under the floor 0 0 — — — — — — —
Over 2.5 goals under the floor 0 0 — — — — — — —
Under 2.5 goals under the floor 0 0 — — — — — — —
Total 1221 758 30.2% −9.5% −153.86 −12.6% −20.9% to −3.4% 0.2010 0.1911

By competition exploratory

Every competition is under the floor on its own, so no competition's ROI can be read.

The record by competition
Competition Rows Matches Won Calibration gap Units ROI ROI 95% Brier, model Brier, market
Allsvenskan under the floor 95 53 — — — — — — —
Austrian Bundesliga under the floor 25 17 — — — — — — —
Bundesliga under the floor 38 23 — — — — — — —
Championship under the floor 42 24 — — — — — — —
Czech First League under the floor 33 20 — — — — — — —
Ekstraklasa under the floor 32 21 — — — — — — —
Eliteserien under the floor 55 40 — — — — — — —
Eredivisie under the floor 46 31 — — — — — — —
HNL under the floor 27 23 — — — — — — —
Jupiler Pro League under the floor 37 25 — — — — — — —
La Liga under the floor 70 40 — — — — — — —
Liga Portugal under the floor 29 19 — — — — — — —
Ligue 1 under the floor 54 32 — — — — — — —
Niké Liga (Slovakia) under the floor 34 21 — — — — — — —
Parva Liga (Bulgaria) under the floor 46 27 — — — — — — —
Premier Division (Ireland) under the floor 40 25 — — — — — — —
Premier League under the floor 57 31 — — — — — — —
Russian Premier League under the floor 18 18 — — — — — — —
Scottish Premiership under the floor 45 33 — — — — — — —
Serie A under the floor 56 32 — — — — — — —
Super League Greece under the floor 36 20 — — — — — — —
SuperLiga (Romania) under the floor 57 34 — — — — — — —
SuperLiga (Serbia) under the floor 42 27 — — — — — — —
Superliga (Denmark) under the floor 41 24 — — — — — — —
Swiss Super League under the floor 53 29 — — — — — — —
Süper Lig under the floor 37 26 — — — — — — —
Ukrainian Premier League under the floor 23 14 — — — — — — —
Veikkausliiga under the floor 53 29 — — — — — — —
Total 1221 758 30.2% −9.5% −153.86 −12.6% −20.9% to −3.4% 0.2010 0.1911

By lead time exploratory

Under an hour is the worst lead time, with the market far ahead on log loss; three to six hours is the one band where the model beats the market on log loss. Above twelve hours the bands are under the floor.

The record by time from evaluation to kickoff
Lead time Rows Matches Won Calibration gap Units ROI ROI 95% Brier, model Brier, market
<1h 250 149 27.6% −14.0% −74.65 −29.9% −45.4% to −12.7% 0.1965 0.1773
1–3h 411 249 31.3% −8.4% −30.42 −7.4% −23.7% to +10.7% 0.2043 0.1964
3–6h 112 75 38.4% −0.6% +15.98 +14.3% −17.8% to +47.1% 0.2120 0.2186
6–12h 332 207 27.7% −11.8% −60.06 −18.1% −34.6% to +0.4% 0.1953 0.1828
12–24h under the floor 30 21 — — — — — — —
>24h under the floor 86 57 — — — — — — —
Total 1221 758 30.2% −9.5% −153.86 −12.6% −20.9% to −3.4% 0.2010 0.1911

By price age exploratory

Stale prices are not where the loss comes from: the rule already keeps them out, and the older bands are under the floor.

The record by age of the price at its evaluation
Price age Rows Matches Won Calibration gap Units ROI ROI 95% Brier, model Brier, market
0–1h 954 596 30.0% −9.6% −113.64 −11.9% −22.4% to −1.3% 0.2020 0.1917
1–3h under the floor 41 25 — — — — — — —
3–6h 215 132 28.9% −11.4% −49.31 −22.9% −40.8% to −4.0% 0.1916 0.1796
6–24h under the floor 2 2 — — — — — — —
>24h under the floor 9 7 — — — — — — —
Total 1221 758 30.2% −9.5% −153.86 −12.6% −20.9% to −3.4% 0.2010 0.1911

The interaction family exploratory

Seven pairs of coarse axes, with two tests a cell: calibration gap different from nought, and ROI different from the ROI the market's margin alone implies. 100 cells hold a selection and 29 are at the floor, a family of 58 tests.

22 have a Benjamini–Hochberg q of 0.05 or less: 22 calibration-gap tests and 0 ROI tests. The smallest ROI q is 0.0604, for 20–40% · 1X2: ROI −35.2% against −6.5% implied by the margin.

Overconfidence is significant in most cells. No cell shows a ROI different from what the market's margin alone would cost.

Data-quality census

Every candidate price on a kicked-off match, by the class that decides whether the rule may read it. Only the last class can hold a signal.

Candidate prices by data-quality class
Class Candidates Matches Settled In the record
incoherent pair (price captured after its evaluation) 1924 150 672 0
warned fixture 4972 340 1793 0
unsampled side 249 25 109 0
priced before the 26 August model revision 1972 150 697 29
short book (a selection of the market unpriced) 0 0 0 0
complete book, more than one bookmaker 0 0 0 0
complete book, one bookmaker 24587 1671 8647 1221
  • 1613 of 1671 matches are single-book: every candidate price on them comes from one bookmaker.
  • The margin removal was cross-checked against the site's own on 3189 complete books, with a largest share difference of 0.00e+00.
  • Of 8651 settled rows regraded from the final score, 209 disagree with the stored result, 20 of them in the record — mostly scores corrected after the settler graded them. The settler never regrades, and the record's ROI is quoted as stored.

What remains unknown

  • Forward CLV: whether the model's disagreement, at the moment a signal first appears, anticipates the market's move. The retrospective record cannot say; the forward ledger has been collecting it since its launch.
  • Whether the loss is the margin or more than it. The interaction family cannot separate an excess from noise in any cell.
  • Whether signals are worse forecasts than the model's average one. The selection effect is −2.0% (−5.0% to +1.1%); a larger forward sample is needed to tell a winner's curse from plain overconfidence.
  • The pending rows: 15940 candidate prices on kicked-off matches are pending. The census counts them and does not explain them; every forecast here means a settled one.
  • The 20 graded rows in the record whose stored result disagrees with the final score.
  • Any single competition: none reaches the floor.

Hypotheses to test prospectively

Each is tested only on the forward ledger, never on history, and none changes the public rule, its thresholds or the model.

  1. The model's disagreement carries information only at a long lead time, before the market has absorbed the news, and a disagreement found early anticipates the market's move.
  2. Part of the model's disagreement is real and most is noise: a probability shrunk halfway toward the market keeps the real part and drops most of the overconfidence.
  3. Overconfidence, not selection, is the main failure: a rule judging the shrunk probability should show a calibration gap near nought on its own signals.

The third is read from the same ledger rows as the second and needs no rule of its own.

Pre-registered shadow rules and their promotion test

Two shadow rules test the first two hypotheses. Both write to the ledger under their own names, are never shown on the board, and are judged only on signals written after they were registered. A third slot is left empty on purpose: no exploratory cut survived its adjustment on ROI, and choosing competitions or markets by their historical ROI is ruled out.

The promotion test

  • At least 400 settled forward signals on at least 250 matches.
  • The mean CLV's match-clustered 95% interval above nought.
  • Out-of-sample log loss on its signals no worse than the margin-free market's.
  • Persistence: the CLV's sign holds in each half when the signals are split by kickoff date.

A promotion is a methodology register entry and a new public rule version. The forward ledger is never recut.

The abandonment test

  • The mean CLV's interval lies wholly below nought at the first reading at the floor.
  • The log loss on its signals is worse than the market's, with an interval excluding nought, at that reading.

A rule is also abandoned if it has not reached the floor within 120 days of its registration: a sample that thin could never test it.

Early disagreement early

Hypothesis. a signal found at least 12 hours before kickoff moves the market toward the selection by the near-kickoff capture.

Definition. exactly the public rule's verdict — rule version 2, every criterion and threshold unchanged, one answer per question — asked at an evaluation at least 12 hours before kickoff. A price that becomes a signal only inside 12 hours is not an early one, and a signal is defined by the first such evaluation.

Half the disagreement shrunk_half

Hypothesis. half of the model's disagreement with the market is real.

Definition. over the public rule's prices, where the market's book is complete: the shrunk probability is the margin-free market plus half the model's disagreement with it, and the shrunk model-implied edge is that probability × the odds − 1. A price is a signal when it passes every public criterion but the threshold (not past kickoff, not an incoherent pair, not stale, full data, measured, under the 40% outlier line), its shrunk model-implied edge is 5% or more, and it is the largest among the sound answers to its question on its match.

Both rules were approved as written on 2 November 2026, with no parameter changed after a forward signal was read. The shrunk rule is kept although the walk-forward test predicts it fails: a pre-registered forward test of the calibrator that stays nearest the market is how that prediction is confirmed or overturned, and abandonment ends it cleanly if it fails.

Editions

Each edition keeps its own address and is never rewritten; a later reading is a new edition beside it.

  • 4 November 2026 this edition

The live record and the forward ledger are on Performance, the rule and every definition are in the Methodology, and today's signals are on the Overview.