A presentation at Bet Better open-data talks (online) by Edward Glush
Well calibrated, unevenly priced What 18,238 settled sports markets say about market forecasts and where the margin sits Edward Glush · Bet Better, Melbourne · 2026 Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 1/9
A price is a forecast Decimal odds of 1.54 imply 1 / 1.54 = 65% once the margin is removed Forecasts can be scored: group by stated chance, count how often it happened This is calibration, the same test used for weather, medicine and ML models Question: how well calibrated is the aggregated fixed-odds market? Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 2/9
Data and method 18,238 settled two-outcome markets, November 2024 to July 2026 Five sports, match-level markets and player-proposition markets Margin removed proportionally; both sides of every market kept 36,476 rows: net 50/50 by construction, no one’s selections inside Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 3/9
Calibration by probability band All markets, both sides of every market included (36,476 observations) Probability band 10 20% 20 30% 30 40% 40 50% 50 60% 60 70% 70 80% 80 90% Observations 2,144 5,263 4,909 5,515 6,077 4,945 5,251 2,157 Mean implied 17.5% 25.2% 34.7% 45.7% 53.8% 65.2% 74.7% 82.5% Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 Realised frequency 18.8% 26.7% 37.0% 46.4% 53.2% 63.0% 73.3% 80.9% Gap (points) +1.4 +1.4 +2.2 +0.7 -0.6 -2.2 -1.4 -1.6 4/9
Reliability diagram Realised frequency against implied probability. Points sit near the diagonal: slightly above it below 50%, slightly b 90 Perfect calibration Observed (all markets) 80 80 90% 70 80% Realised frequency (%) 70 60 70% 60 50 60% 50 40 50% 40 30 40% 30 20 30% 20 10 10 20% 10 20 30 40 50 60 70 Mean implied probability, margin removed (%) Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 80 90 3 2 1 0 1 2 Gap: realised minus implied (points) 5/9 3
Finding 1 · the market is a very good forecas Every band lands within about two points of its implied probability A 65% chance happened about 63% of the time; a 25% chance about 27% A demanding benchmark for any forecasting model Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 6/9
Finding 2 · the residual tilt is systematic Every band below 50% overperforms (+0.7 to +2.2 points) Every band above 50% underperforms ( 0.6 to 2.2 points) Player-proposition markets show roughly twice the tilt of match markets The same markets carry the highest quoted margins (7.93% vs 4.77% two-way) Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 7/9
Interpretation · the margin lives on the longs Margin is not spread evenly: more of it sits on the outsider’s price Proportional de-vigging then leaves exactly this residual signature Caveat: the de-vig method choice affects measured tail behaviour The open data let anyone re-run it under another allocation Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 8/9
Try it yourself Study and teaching data set: betbetter.world/studies/market-calibration Daily margin index, 38 competitions: betbetter.world/studies/margin-index How margins are measured: betbetter.world/studies/bookmaker-margins Free model API, no key, clients in 13 languages: betbetter.world/api/ Everything above is CC BY 4.0 · statistics for education, not betting advice Edward Glush · Bet Better · betbetter.world/studies/market-calibration · CC BY 4.0 9/9
A price is a forecast, and forecasts can be scored. This deck walks through a calibration check of margin-free sports prices on both sides of 18,238 settled two-outcome markets (36,476 observations, five sports, November 2024 to July 2026). The market turns out to be a very good forecaster: every probability band lands within about two points of its stated chance. The small residual tilt is systematic, with outsiders overperforming and favourites underperforming, and is roughly twice as large in player-proposition markets, the same markets that carry the highest quoted margins. The reading is that the margin is loaded onto the longshot side.
The study, its teaching data set and the figures are open under CC BY 4.0: market calibration study, daily bookmaker margin index, how margins are measured and the free model API with no key. Statistics for education and analysis, not betting advice.
The following resources were mentioned during the presentation or are useful additional information.
The full write-up, Table 1 and the CC BY 4.0 data set (36,476 rows: month, sport, market family, margin-free probability, outcome).
Mean and median margin by market type, updated daily, with a per-competition file for 38 competitions. CC BY 4.0.
Method notes: measured against estimated margins, the de-vig allocation caveat and the two-way versus three-way comparison.
Model win probabilities, predicted scores and power ratings as plain JSON, with MIT client libraries in 13 languages.
