xgEdge

Betting

Model performance

Three questions, answered separately: whether the model's probabilities meant what they said, across every selection it priced on a played match; which of the model and the bookmaker forecast the same matches better; and how the signals the published rule would have flagged have done at an equal stake of one unit. Each is read over its own population, so no one of them answers another.

Model vs market Market ahead 1X2 · Brier 0.597 vs 0.571
Signal record −17.7% ROI 1,697 settled 1,023 matches −300.49u
Recorded since 20 August 2026 1,735 matches

What the evidence tells us

  1. 1X2 forecaster

    Market leads

    0.026 Brier difference, in the market's favour

    On the same matches the market's probabilities scored better on both measures, and neither interval reaches zero.

    Model vs market
  2. Retrospective rule result

    Below zero

    −17.7% ROI −25.3% to −10.2%

    Staking every signal today's rule finds, at the price quoted then, lost money, and the whole interval lies below zero.

    Signal record
  3. Signal calibration

    Overconfident

    −10.7 pp observed less predicted −12.9 pp to −8.5 pp

    Signals won less often than the model's probabilities said, by more than chance explains.

    Calibration on the record
  4. Prospective validation

    Collecting

    190 / 202 settled / recorded

    Signals are being recorded live, but too few have settled to read them yet.

    Prospective validation

Question 1

Model vs market

On the same matches, which forecaster is more accurate: the model or the bookmaker with its margin removed? Then, whether the model forecasts better than knowing how often each outcome happens. Neither says whether disagreeing with the market pays.

Live Each forecast as written before kickoff, against the bookmaker's price at the time · model v1 · comparison v1.

Model vs market

Model vs market: 1X2

The model's home, draw and away probabilities beside the bookmaker's, with the margin removed, on the same matches, each against what happened. Both are read as they stood before kickoff — the model's last published probability and the last pre-kickoff market price stored beside it — so nothing is re-priced with hindsight, and the same matches, buckets and floors apply to both.

Who leads

Market leads

Brier difference 0.026 in the market's favour 95% interval 0.019 to 0.032
Log-loss difference 0.038 in the market's favour 95% interval 0.028 to 0.047

A side leads only when both measures favour it and neither interval includes nought; each interval resamples matches, 2,000 times from seed 20261102.

Brier score
Model
0.597
Market, margin removed
✓ Better 0.571
Base rate
0.645
Log loss
Model
1.000
Market, margin removed
✓ Better 0.963
Base rate
1.066

2,088 matches 6,264 forecasts 19 August 2026 – 14 November 2026

On these matches the market's probabilities, margin removed, scored better than the model's on both measures.

Lower is better for the Brier score, the log loss and the average calibration gap under each chart. Both scores are taken per match over the three outcomes together and averaged over the matches, whichever view is showing. The base rate is a forecast that gave every match this sample's own home, draw and away frequencies, which it could only have known with hindsight.

Model

Average calibration gap: 3.6 pp. It depends on the buckets and summarises the diagram; it is not a score of accuracy.

Reliability of the model: its probability against how often the outcome happened Each point is a band of the model's forecasts: across, the model's average probability in the band; up, how often those outcomes happened. A point on the diagonal is perfectly calibrated. The table below holds the same figures beside the market's. 0% 0% 25% 25% 50% 50% 75% 75% 100% 100% Too cautious Too confident

Model probability → · ↑ Happened

Above the diagonal the forecasts were too cautious, below it too confident. The size of a point is its number of forecasts. Bands with fewer than 30 forecasts are not drawn.

Market, margin removed

Average calibration gap: 1.4 pp. It depends on the buckets and summarises the diagram; it is not a score of accuracy.

Reliability of the market, margin removed: its probability against how often the outcome happened Each point is a band of the market's forecasts, the bookmaker's margin removed: across, the market's average probability in the band; up, how often those outcomes happened. A point on the diagonal is perfectly calibrated. The table below holds the same figures beside the model's. 0% 0% 25% 25% 50% 50% 75% 75% 100% 100% Too cautious Too confident

Market probability → · ↑ Happened

Above the diagonal the forecasts were too cautious, below it too confident. The size of a point is its number of forecasts. Bands with fewer than 30 forecasts are not drawn.
Calibration by band of probability, the model beside the market
Model Market, margin removed
Probability Forecasts Average Happened Forecasts Average Happened
0–10% 41 8.2% 7.3% 163 7.8% 3.7%
10–20% 527 16.4% 13.5% 911 16.0% 14.7%
20–30% 2,314 26.3% 23.3% 2,492 25.4% 24.4%
30–40% 1,885 33.8% 32.2% 948 34.8% 34.7%
40–50% 834 44.4% 49.9% 705 44.8% 46.7%
50–60% 463 54.4% 64.8% 514 54.9% 55.6%
60–70% 167 63.9% 76.0% 315 64.6% 70.8%
70–80% 31 73.7% 71.0% 171 74.2% 78.4%
80–90% 2 Too few 44 83.1% 84.1%
90–100% 0 Too few 1 Too few
  • Within 2 pp
  • 2–4 pp off
  • More than 4 pp off

What was measured

Period
Matches kicking off from 19 August 2026 to 14 November 2026.
Sample
2,088 matches and 6,264 forecasts: three per match, one for each of home, draw and away. Three forecasts on one match are not three independent matches.
Bands
Ten equal-width bands of the probability scale, closed on the left. A band is drawn from 30 forecasts, counted per source, so a band can be drawn for one and not the other.
Bookmaker
One book per match, the first with every 1X2 price in a fixed order of preference: Bet365 → Paddy Power → Pinnacle. In this sample: Bet365 1,956 · Paddy Power 31 · Pinnacle 101. One book's price, not a consensus.
Market price
The last pre-kickoff market price: the price stored with the model's last evaluation before kickoff, a median 4.8 h before the match (quartiles 2.5 h to 8.9 h). It is the last price captured, not the price at the moment the match started. Hours before kickoff the price was captured:
  • 0–1 h 280
  • 1–3 h 508
  • 3–6 h 561
  • 6–12 h 473
  • 12–24 h 46
  • 24–48 h 156
  • 48–72 h 63
  • 72 h and more 1
Margin
The bookmaker's margin, 7.3% on average, was removed by scaling the three implied probabilities so that they sum to 100%.
Model probability
The last probability the site published before kickoff, stored beside the price. Nothing is re-priced with hindsight.
Model revision
The model's inputs changed on 26 August 2026, in the sample's first week; 148 of these matches were priced before the change. The comparison measures what the site published, not one frozen model.
Warnings
369 matches carry a model warning and 25 have a side with no sample. All are included and scored.
Last updated

Download the data behind both charts: CSV JSON Every band of every view and source with its counts and frequencies, and the method's metadata. No odds and no per-match rows.

Beating the base rate is not beating the market. The tables below measure the model against how often each outcome happens, a far easier opponent than the bookmaker above.

Skill against the base rate, by market

Retrospective Model v1 replayed over past matches, measured 17 August 2026.

How much better than the base rate each market's probabilities scored when the model was replayed over 7 competitions' past seasons, measured on 17 August 2026. Zero is the base rate; above it the model adds information.

Forecast skill against the base rate per market, highest first
Market Skill vs base rate
Away win +7.20%
Home win +7.19%
Home team to score +3.77%
Home team 2+ goals +3.62%
Away team 2+ goals +2.61%
Away team to score +2.56%
Over 1.5 goals -0.01%
Both teams to score -0.43%
Over 2.5 goals -0.89%
Draw -1.04%

Over every market at once: +2.50% against the base rate.

Beating the base rate is about forecasting, not betting: a market can beat its base rate and still lose money against the bookmaker's price.

Do the model's goals match the goals scored?

Every played match the model priced before kickoff while version 1 was in force, since 26 August 2026. Both figures are 1.0 for an honest model.

Goal-difference slope
1.544
95% interval 1.40 to 1.69 · 1,973 matches
Above 1.0, real goal differences are wider than predicted: the model rates clubs as closer to each other than they are. Below 1.0, it rates them as further apart.
Goals scored per goal predicted
1.073
95% interval 1.05 to 1.10 · 1,973 matches
Above 1.0, more goals are scored than predicted; below 1.0, fewer.

Computed weekly; this reading as of 16 November 2026.

Live Each forecast as model v1 wrote it before kickoff.

Every forecast

Market-specific calibration

When the model gives an outcome 60%, does it happen about 60% of the time? Every selection it priced on a played match, grouped by the probability it gave, against how often that selection happened.

This section inspects the model on its own across the individual betting markets; the section above benchmarks it against the bookmaker on 1X2.

1,735 matches 29,495 forecasts

Scored from the last projection made before kickoff. A projection written after a match started is never counted.

Brier score Beats base rate
0.210
Base rate
0.217
Against base rate
3.0% better
Log loss Beats base rate
0.609
Base rate
0.624
Against base rate
2.4% better

Lower is better for both. The base rate is a forecast that gave every selection how often it happened over this same sample, which it could only have known with hindsight.

Reliability: the model's probability against how often it happened Each point is a band of forecasts: across, the model's average probability in the band; up, how often those selections happened. A point on the diagonal is perfectly calibrated. The table below holds the same figures. 0% 0% 25% 25% 50% 50% 75% 75% 100% 100% Too cautious Too confident

Model probability → · ↑ Happened

Above the diagonal the model was too cautious, below it too confident. The size of a point is its number of forecasts. Bands with fewer than 30 forecasts are not drawn.
Calibration by band of model probability
Model probability Forecasts Model avg Happened Gap
0–10% 256 8.2% 10.2% +2.0 pp
10–20% 2550 15.9% 17.5% +1.6 pp
20–30% 4976 25.5% 25.6% +0.1 pp
30–40% 4692 34.7% 36.5% +1.8 pp
40–50% 4324 44.9% 47.0% +2.0 pp
50–60% 4011 54.9% 54.8% −0.1 pp
60–70% 3291 64.8% 62.2% −2.6 pp
70–80% 3059 75.0% 72.6% −2.4 pp
80–90% 2111 84.2% 81.7% −2.5 pp
90–100% 225 91.8% 89.3% −2.5 pp
  • Within 2 pp
  • 2–4 pp off
  • More than 4 pp off

Retrospective Each forecast as written before kickoff; the signals by today's rule v2 applied to history.

Every forecast, the signals and the market

Reliability: model, signals and market

When the model says 40%, does it happen 40% of the time? Every settled forecast, the signals today's rule keeps among them, and the bookmaker's price with its margin removed on the same forecasts, each grouped by the probability it gave and set against how often it happened.

A series on the diagonal forecasts honestly. That is not a return: a forecast can be honest and still be priced better by the market.

Reliability: model, signals and market against how often it happened Three series over the diagonal: across, the probability a band of forecasts gave on average; up, how often those selections happened. The figures, every band of every series with its count, are in the table beside the chart. Model: 9 of 10 tenths drawn, 1 under the floor, on 11,010 forecasts. Signals: 6 of 10 tenths drawn, 3 under the floor, on 1,686 forecasts. Market: 9 of 10 tenths drawn, 1 under the floor, on 11,010 forecasts. 0% 0% 25% 25% 50% 50% 75% 75% 100% 100% Too cautious Too confident

Probability given → · ↑ Happened

A band is drawn from 100 forecasts on 50 matches. The table lists every band, drawn or not, with its count.
  • Model, every settled forecast 11,010 forecasts won or lost
  • Model, the signals 1,686 forecasts won or lost, kept by today's rule applied to history
  • Market, margin removed 11,010 of the same forecasts with the bookmaker's full book
  • Model: 9 of 10 tenths drawn, 1 under the floor, on 11,010 forecasts.
  • Signals: 6 of 10 tenths drawn, 3 under the floor, on 1,686 forecasts.
  • Market: 9 of 10 tenths drawn, 1 under the floor, on 11,010 forecasts.
Reliability by band of probability: model, signals and market
Probability Model Signals Market
0–10% Too few 60 forecasts Too few 7 forecasts 7.8% , happened 3.7% 352 forecasts
10–20% 16.0% , happened 10.3% 834 forecasts 16.0% , happened 9.6% 178 forecasts 15.6% , happened 14.5% 1,610 forecasts
20–30% 25.9% , happened 18.3% 2,219 forecasts 25.2% , happened 17.3% 352 forecasts 25.1% , happened 23.7% 2,756 forecasts
30–40% 33.9% , happened 25.8% 1,815 forecasts 34.7% , happened 23.8% 382 forecasts 35.1% , happened 32.3% 1,308 forecasts
40–50% 45.1% , happened 34.5% 1,251 forecasts 44.7% , happened 30.7% 300 forecasts 44.7% , happened 41.9% 1,411 forecasts
50–60% 55.0% , happened 42.9% 1,537 forecasts 54.5% , happened 42.5% 214 forecasts 54.3% , happened 52.0% 1,790 forecasts
60–70% 64.9% , happened 53.5% 1,500 forecasts 64.7% , happened 52.0% 173 forecasts 64.3% , happened 65.4% 852 forecasts
70–80% 74.5% , happened 64.4% 1,007 forecasts Too few 60 forecasts 74.7% , happened 73.4% 779 forecasts
80–90% 84.4% , happened 73.8% 686 forecasts Too few 20 forecasts 82.9% , happened 80.8% 151 forecasts
90–100% 91.7% , happened 83.2% 101 forecasts — Too few 1 forecast

Question 2

Signal record

Does disagreeing with the market pay, at the prices that were there and after the uncertainty? This is today's rule applied to history, not a live track record. A model can beat its base rate, and even the market, and still lose money here.

Retrospective Today's rule v2 applied to history.

Retrospective record · Rule v2

Lost money on this record (interval below zero)

ROI
−17.7%
95% interval
−25.3% to −10.2%
Signals
1,697
Matches
1,023
Units
−300.49u
Rule
v2
How this is computed

Today's rule, version 2, applied to each historical selection at its last evaluation before kickoff. These were not necessarily signals published live at the time, and a quoted historical price may not have been available to you. This record can hold the verdict down; it can never raise it. Only signals published live, in the forward ledger, can.

The 95% interval resamples matches rather than selections, 2,000 times from seed 20261102, so selections on one match move together.

11 voids are staked and counted. 929 settled rows are left out of every figure: their price was captured after the evaluation it was paired with.

Retrospective: today's rule asked about each selection at its last evaluation before kickoff, not a list published at the time. A quoted historical price may not have been available to you. Computed .

How this record is calculated

Computed .

Signal record (retrospective · rule v2)

All competitions, all seasons

Retrospective Today's rule v2 applied to history.

Filters
Dates
More filters
Competitions

Bands of the model-implied edge at each signal's own evaluation. The rule's threshold (5%) is set by the rule and is never tuned from this view: a band only narrows which signals are shown.

Odds
Settled signals
1697
across 1023 matches
Won–lost–void
494 won, 1192 lost, 11 void
29.3% of decided signals won
Units · ROI
−300.49u
−17.7% ROI · equal stake of one unit
Calibration gap
−10.7 pp
95% interval −12.9 pp to −8.5 pp, matches resampled
Avg model probability 40.0% · observed frequency 29.3%
Avg model-implied edge · avg price
+18.2%
at evaluation · average price 3.69

Cumulative units

Cumulative units of the selected signals Running units over 1697 settled signals, week by week, from Monday, by kickoff from 20 Aug 2026 to 9 Nov 2026, ending at −300.49u. −400.00u −300.00u −200.00u −100.00u 0.00u 20 Aug 2026 9 Nov 2026
One unit staked per selection, added up week by week, from Monday by kickoff day; the line ends at −300.49u, the units in the cards above. A void adds nothing.
Show as table
Cumulative units of the selected signals
Week from Units Running total
17 Aug 2026 −8.00u −8.00u
24 Aug 2026 −16.61u −24.61u
31 Aug 2026 −16.11u −40.72u
7 Sep 2026 −23.60u −64.32u
14 Sep 2026 −3.85u −68.17u
21 Sep 2026 −2.00u −70.17u
28 Sep 2026 −1.00u −71.17u
5 Oct 2026 −51.17u −122.34u
12 Oct 2026 −47.23u −169.58u
19 Oct 2026 −14.79u −184.36u
26 Oct 2026 −59.17u −243.54u
2 Nov 2026 −54.95u −298.49u
9 Nov 2026 −2.00u −300.49u

By market family

One row per market family, in the registry's order and never sorted by return. A family the rule qualified nothing in is kept, with dashes. Every selection is in one row, so the rows add up to the cards above.

The signal record in this period by market family: settled signals, won–lost–void, units, ROI, the share of decided signals won and the average price.
Market family Settled W–L–V Units ROI Won Avg price
1X2 573 130–438–5 −104.70u −18.3% 22.9% 4.48
Goals 0 0–0–0 — — — —
BTTS 0 0–0–0 — — — —
Team goals 1124 364–754–6 −195.79u −17.4% 32.6% 3.30

A dash in the ROI column means fewer than 30 settled selections.

By band

The same signals cut four ways, in bands fixed before any figure was read: by model-implied edge, by odds, by how long before kickoff each was evaluated and by how old its price was. Each table adds up to the cards above, follows the filters, and is never sorted by ROI. Retrospective: today's rule applied to history.

By model-implied edge

Whether a larger disagreement with the market goes with better results. The model-implied edge is the model's claim, not an observed return; the share won beside the mean model probability shows how far the results fell short of, or beat, what was claimed.

The signal record by model-implied edge band at evaluation: settled signals, matches, mean model probability, share won, calibration gap, units, ROI with its interval, mean model-implied edge and forward CLV.
Model-implied edge Settled Matches Model probability Won Calibration gap Units ROI Model-implied edge Forward CLV
5–10% 449 397 46.7% 38.3% −8.4 pp −65.18u −14.5% −25.2% to −2.4% +7.3% —
10–15% 330 300 40.1% 33.6% −6.5 pp −27.41u −8.3% −23.8% to +7.8% +12.4% —
15–20% 253 229 39.3% 26.8% −12.5 pp −46.50u −18.4% −36.5% to +1.2% +17.2% —
20–30% 396 351 36.7% 21.9% −14.8 pp −117.35u −29.6% −44.0% to −14.8% +24.8% —
30–40% 269 237 34.0% 22.0% −12.0 pp −44.06u −16.4% −39.3% to +6.8% +34.6% —

By odds

Whether the model is overconfident on longer prices. A band includes its lower price and excludes its upper.

The signal record by the price each signal was paired with: settled signals, matches, mean model probability, share won, calibration gap, units, ROI with its interval, mean model-implied edge and forward CLV.
Odds Settled Matches Model probability Won Calibration gap Units ROI Model-implied edge Forward CLV
Under 1.20 0 0 — — — — — — —
1.20–1.50 34 34 — — — — — — —
1.50–2.00 240 226 65.3% 53.3% −12.0 pp −21.32u −8.9% −19.7% to +2.3% +11.7% —
2.00–3.00 470 395 48.8% 35.5% −13.4 pp −78.37u −16.7% −27.1% to −6.5% +16.1% —
3.00–5.00 584 513 33.1% 23.0% −10.1 pp −95.77u −16.4% −29.8% to −2.3% +19.7% —
5.00 and over 369 324 19.3% 11.7% −7.5 pp −104.14u −28.2% −49.3% to −5.8% +23.4% —

By lead time

Whether the time between the evaluation and kickoff matters.

The signal record by how long before kickoff each signal was evaluated: settled signals, matches, mean model probability, share won, calibration gap, units, ROI with its interval, mean model-implied edge and forward CLV.
Kickoff − evaluation Settled Matches Model probability Won Calibration gap Units ROI Model-implied edge Forward CLV
<1h 468 271 40.9% 27.9% −13.0 pp −126.72u −27.1% −40.3% to −12.9% +18.1% —
1–3h 667 391 40.1% 29.2% −11.0 pp −124.85u −18.7% −30.5% to −6.8% +18.4% —
3–6h 112 75 39.0% 38.4% −0.6 pp +15.99u +14.3% −15.9% to +46.7% +18.0% —
6–12h 332 207 39.6% 27.7% −11.8 pp −60.07u −18.1% −34.8% to −0.6% +18.2% —
12–24h 30 21 — — — — — — —
>24h 88 58 — — — — — — —

By price age

Whether an older price goes with different results. The age is the evaluation's time less the time the price was fetched.

The signal record by how old each signal's price was at its evaluation: settled signals, matches, mean model probability, share won, calibration gap, units, ROI with its interval, mean model-implied edge and forward CLV.
Price age Settled Matches Model probability Won Calibration gap Units ROI Model-implied edge Forward CLV
0–1h 1229 748 39.8% 29.6% −10.3 pp −190.94u −15.5% −24.4% to −6.2% +18.3% +1.5% +0.0% to +2.8%
1–3h 69 44 — — — — — — —
3–6h 388 226 40.9% 28.0% −12.9 pp −104.57u −27.0% −40.5% to −12.8% +17.9% —
6–24h 2 2 — — — — — — —
>24h 9 7 — — — — — — —

A band under 100 settled signals or 50 matches shows its count only: no rate, no ROI, no tone. Forward CLV is the mean CLV of the public ledger's signals written in the same band, measured against Bet365 about an hour before kickoff; a dash until that band holds enough measured signals.

Qualification definition updated on 19 October 2026. Historical figures were recalculated using the same rules as the live board. See methodology history

Sample
1697 settled signals on 1023 matches, winners and losers alike.
Correlation
Selections from the same match may be correlated. The selection count is not a count of independent experiments; the match count is the smaller figure to read it against.
Price and moment
Historical selections use the price paired with the last qualifying pre-match evaluation. Prices were current when evaluated and may have moved before kickoff.
Excluded pairs
929 settled rows whose price was captured after the evaluation that was meant to price it are left out and counted here.
Odds source
One bookmaker's price, the first of Bet365 → Paddy Power → Pinnacle to quote the market. Not the highest price on the market and not a consensus.
Availability
Prices as captured, not as matched. A price may have moved, been limited or been withdrawn before a bet could be placed.

Detailed breakdowns

Retrospective Today's rule v2 applied to history.

The selections above cut by market, by season, by month and by competition, every row with the same ROI floor and every table adding up to the cards. They are historical breakdowns in a fixed order: nothing here is sorted by return.

Performance by market Show

One row per selection of a market, in the registry's order. A historical breakdown, not a recommendation to prefer the markets with a positive return. Beats base rate means the model predicted that market better than a historical base-rate forecast when its skill was measured; it does not mean it beats bookmaker prices, and it is not a measure of profit.

The record by market
Market Forecast skill Settled W–L–V Units ROI
Home team to score Beats base rate 27 11–16–0 −8.64u —
Home team not to score Beats base rate 114 13–100–1 −37.67u −33.0%
Away team to score Beats base rate 62 38–24–0 +1.15u +1.9%
Away team not to score Beats base rate 126 19–105–2 −37.81u −30.0%
Home win Beats base rate 300 71–226–3 −80.77u −26.9%
Away win Beats base rate 273 59–212–2 −23.92u −8.8%
Home team 2+ goals Beats base rate 229 70–157–2 −49.56u −21.6%
Home team at most 1 goal Beats base rate 183 73–110–0 −21.32u −11.7%
Away team 2+ goals Beats base rate 223 60–162–1 −34.86u −15.6%
Away team at most 1 goal Beats base rate 160 80–80–0 −7.08u −4.4%
Performance by season Show

One row per season, newest first, with the same ROI floor. No season is ranked above another.

The record by season
Season Settled W–L–V Units ROI
2026/2027 1697 494–1192–11 −300.49u −17.7%
Performance by month Show

One row per month a selection kicked off in, oldest first, by the kickoff's day in Zagreb, with the same ROI floor. No month is ranked above another.

The record by month
Month Settled W–L–V Units ROI
August 2026 143 37–105–1 −28.57u −20.0%
September 2026 455 139–312–4 −41.60u −9.1%
October 2026 730 217–509–4 −143.35u −19.6%
November 2026 369 101–266–2 −86.96u −23.6%
Performance by competition Show

One row per competition holding a selection, by name, with the same ROI floor. Most competitions hold too few selections for a return to mean much, so none is ranked above another.

The record by competition
Competition Settled W–L–V Units ROI
Allsvenskan 130 44–86–0 −5.27u −4.1%
Austrian Bundesliga 43 9–34–0 −22.17u −51.6%
Bundesliga 46 12–34–0 −13.92u −30.3%
Championship 82 23–59–0 −24.70u −30.1%
Czech First League 47 10–37–0 −13.35u −28.4%
Ekstraklasa 40 21–19–0 +17.69u +44.2%
Eliteserien 75 18–57–0 −25.01u −33.4%
Eredivisie 66 21–45–0 +4.91u +7.4%
HNL 40 16–24–0 +11.01u +27.5%
Jupiler Pro League 51 12–39–0 −18.53u −36.3%
La Liga 94 31–60–3 −8.90u −9.5%
Liga Portugal 38 15–23–0 +5.03u +13.2%
Ligue 1 74 17–57–0 −20.74u −28.0%
Niké Liga (Slovakia) 42 12–30–0 −18.15u −43.2%
Parva Liga (Bulgaria) 67 19–47–1 −17.20u −25.7%
Premier Division (Ireland) 45 15–30–0 −1.07u −2.4%
Premier League 77 18–59–0 −27.11u −35.2%
Russian Premier League 23 7–16–0 +1.01u —
Scottish Premiership 53 16–37–0 −4.76u −9.0%
Serie A 78 22–56–0 −8.12u −10.4%
Super League Greece 49 16–33–0 −9.07u −18.5%
Superliga (Denmark) 56 13–43–0 −23.28u −41.6%
SuperLiga (Romania) 89 27–62–0 −25.69u −28.9%
SuperLiga (Serbia) 68 20–46–2 −18.88u −27.8%
Swiss Super League 61 15–42–4 −3.40u −5.6%
Süper Lig 59 19–40–0 −4.34u −7.4%
Ukrainian Premier League 39 8–30–1 −14.51u −37.2%
Veikkausliiga 65 18–47–0 −11.96u −18.4%

A dash in the ROI column means fewer than 30 settled selections.

Retrospective Today's rule v2 applied to history.

Calibration of the signal record

The record's own selections under the filters above, scored as the calibration above scores every forecast: the model's average probability against how often the selections happened. Voids are counted and not scored; pending selections never enter.

1,697 settled signals · 1,686 decided · 11 void

Model average 40.0%
Observed 29.3%
Difference −10.7 pp Happened less often than priced
Brier score Worse than base
0.197 Base rate 0.193

Over the decided selections. The base rate is a forecast that gave each selection how often it happened over this same subset, which it could only have known with hindsight; lower is better.

Signals have occurred less often than their model probabilities implied over the recorded sample, which helps explain why a positive model-implied edge has not become a positive observed result.

This is a different population from the calibration above: only settled selections that met the qualification rule at their last pre-match evaluation. It describes the recorded sample, not a fixed property of the model.

How this population is defined

Market overview

Live Priced and current signals live over the next 72 hours; forecast skill from model v1 replayed over past matches; retrospective ROI by today's rule v2 applied to history.

For each selection the board prices in the next 72 hours: the matches that price it and how many of those prices are signals now, then whether the model forecasts it better than the base rate and what the retrospective record made betting it, since forecast skill and betting results are separate questions. Each record figure opens the selection's row in the breakdown above.

Matches priced, current signals, forecast skill, retrospective ROI and forward evidence for each selection
Selection Priced next 72 h Current signals next 72 h Forecast skill Retrospective ROI Evidence
Home win 12 2 −26.9% 300 signals · 95% interval −42.0% to −11.3% Retrospective record for Home win, in the breakdown above Not yet validated CLV collecting (32 measured)
Draw 12 0 No historical signals Retrospective record for Draw, in the breakdown above Not yet validated CLV collecting (0 measured)
Away win 12 2 −8.8% 273 signals · 95% interval −30.8% to +14.4% Retrospective record for Away win, in the breakdown above Not yet validated CLV collecting (19 measured)
Over 1.5 goals 11 0 No historical signals Retrospective record for Over 1.5 goals, in the breakdown above Not yet validated CLV collecting (0 measured)
Under 1.5 goals 11 0 No historical signals Retrospective record for Under 1.5 goals, in the breakdown above Not yet validated CLV collecting (0 measured)
Over 2.5 goals 12 0 No historical signals Retrospective record for Over 2.5 goals, in the breakdown above Not yet validated CLV collecting (0 measured)
Under 2.5 goals 12 0 No historical signals Retrospective record for Under 2.5 goals, in the breakdown above Not yet validated CLV collecting (0 measured)
Both teams to score 11 0 No historical signals Retrospective record for Both teams to score, in the breakdown above Not yet validated CLV collecting (0 measured)
Both teams not to score 11 0 No historical signals Retrospective record for Both teams not to score, in the breakdown above Not yet validated CLV collecting (0 measured)
Home team to score 9 0 Insufficient sample (27) Retrospective record for Home team to score, in the breakdown above Not yet validated CLV collecting (1 measured)
Home team not to score 9 2 −33.0% 114 signals · 95% interval −65.5% to +4.0% Retrospective record for Home team not to score, in the breakdown above Not yet validated CLV collecting (9 measured)
Away team to score 9 1 Insufficient sample (62) Retrospective record for Away team to score, in the breakdown above Not yet validated CLV collecting (9 measured)
Away team not to score 9 1 −30.0% 126 signals · 95% interval −57.5% to +1.2% Retrospective record for Away team not to score, in the breakdown above Not yet validated CLV collecting (15 measured)
Home team 2+ goals 9 1 −21.6% 229 signals · 95% interval −37.1% to −4.7% Retrospective record for Home team 2+ goals, in the breakdown above Not yet validated CLV collecting (40 measured)
Home team at most 1 goal 9 1 −11.7% 183 signals · 95% interval −28.6% to +5.8% Retrospective record for Home team at most 1 goal, in the breakdown above Not yet validated CLV collecting (16 measured)
Away team 2+ goals 9 2 −15.6% 223 signals · 95% interval −35.0% to +4.5% Retrospective record for Away team 2+ goals, in the breakdown above Not yet validated CLV collecting (25 measured)
Away team at most 1 goal 9 0 −4.4% 160 signals · 95% interval −20.5% to +11.9% Retrospective record for Away team at most 1 goal, in the breakdown above Not yet validated CLV collecting (22 measured)

Question 3

Prospective validation

Do the signals published live, as they were written and never recut, hold up once their matches are played? The record above is reconstructed; this is the forward test.

Where it stands

Recorded
202
Settled
190
Matches
84
Since
7 November 2026
CLV
mean +1.0% (−0.4% to +2.4%)

Settled so far: 190 of the 400 signals and 84 of the 250 matches validation needs.

Forward ledger

Live Rule version as written when the signal was recorded.

Every signal the public rule wrote, kept as it was written and settled after the match. Nothing in it is recalculated when the rule changes.

Signals recorded
202
since 7 November 2026
Settled
190
across 84 matches
Won–lost–void
56 won, 132 lost, 2 void
Units · ROI
−36.01u
ROI −19.0%

190 settled. The verdict can move above “not demonstrated” only from 400 settled signals on 250 matches.

Forward CLV

Live Each signal's close as the closer wrote it after kickoff.

Whether the market moved toward the model after a signal. Each signal's price is compared with the same bookmaker's price near kickoff; if the model saw something the market had not yet priced, the price should move its way.

The close is Bet365, about an hour before kickoff: the signal's own bookmaker, captured between 0 and 90 minutes before kickoff. It is a soft benchmark, not a sharp bookmaker's price at kickoff, so a positive CLV against it is a weaker finding than the same figure against a sharp book at kickoff.

How CLV is computed

CLV = margin-free probability at the close ÷ margin-free probability at the signal − 1

Above nought, the market moved toward the signal. Both probabilities come from the same bookmaker's complete book with the margin removed; where either book is short, there is no CLV.

Closes by status

Captured
188
Not yet closed
12
No capture
2
Outside the window
0
Another bookmaker
0
Short book
0

A signal closes once its match has kicked off. Only a captured close has a CLV; a missing one is counted by its reason and never filled in.

Collecting since 7 November 2026, 190 of 400 settled. No CLV rate is shown until 400 signals have settled on 250 matches; 188 have a CLV so far.

Retrospective CLV

Retrospective Today's rule v2 applied to history; each selection's last-evaluation price against the near-kickoff capture, computed, never stored.

The retrospective record's signals, each at its last evaluation before kickoff, against Bet365's price about an hour before kickoff. Near-kickoff captures began on 20 October 2026, so only matches from then have a close. This is history, not forward evidence.

471 of 738 (64%) are exactly nought. The record's price is its last evaluation before kickoff, often within an hour or two of the capture itself, so the price has had little time to move; the forward CLV above is measured from the moment a signal is first written, which leaves the market room to move.

Mean CLV
+0.25%
95% interval -0.14% to +0.64%
Median CLV
+0.00%
Above nought
17.9%
Signals with a CLV
738
across 415 matches

Distribution

Signals with a CLV, by CLV band; nought exactly is a band of its own
CLV band Signals
below −10% 17
−10% to −5% 43
−5% to −2% 62
−2% to 0 13
0 471
0 to 2% 15
2% to 5% 41
5% to 10% 46
10% and above 30

Closes: 953 kicked off before captures began · 738 captured · 4 outside the window · 2 never captured

How the record is kept

Retrospective Today's rule v2 applied to history.

Every settled selection that was a signal at its last evaluation before kickoff, in every competition the site prices, at the bookmaker's price paired with that evaluation. Nothing is re-priced, and nothing written after kickoff is counted.

An equal stake of one unit per signal: a win returns the price less the stake, a loss costs the stake, and a void returns it and still counts as staked. ROI is quoted only from 30 settled signals.

Pending selections, rows without a price, the excluded pairs and the daily combinations are left out. The full rules, how a selection settles and the history of the qualification rule are on the methodology page.

The 95% interval resamples matches rather than selections, 2,000 times from seed 20261102, so selections on one match move together.

11 voids are staked and counted. 929 settled rows are left out of every figure: their price was captured after the evaluation it was paired with.

Retrospective: today's rule asked about each selection at its last evaluation before kickoff, not a list published at the time. A quoted historical price may not have been available to you. Computed .

The full methodology

Read the research note