20% 勝率，不是五場內必勝一場 A 20% win probability does not promise one win in five races 20% 胜率，不是五场内必胜一场

One race produces one realised result

In a 12-runner race on 2 November 2025, the normalized market probabilities totalled 100%. Runner 11 won with a probability of 19.99%; runner 4, the highest-probability runner at 23.42%, did not. That does not make 19.99% more accurate than 23.42%. It shows that one of several possible outcomes was realised on the day.

A 20% probability must not be read as a guaranteed win within five races. If the five forecasts involve different fields, conditions and data cut-offs, they may not even belong to one directly comparable group.

Three runners are shown for illustration; the full field contained 12.

The meaning of 20% appears only over time

If a method repeatedly assigns probabilities near 20%, and roughly 20% of those runners win across a sufficiently large comparable sample, that band is broadly calibrated. One winner does not turn 20% into 100%, and one loser does not turn it into zero.

Comparable means that the prediction target and information boundary stay consistent. Win probabilities cannot be mixed with place probabilities, and a pre-race forecast cannot be mixed with an estimate that already uses post-race sectionals. A complete sample matters more than a handful of memorable races.

Across 21,123 runners, the direction is useful but imperfect

The historical market dataset contains 21,123 runners across 1,712 races. Each runner's win odds were converted to market probability and normalized so the field totalled 100%. Actual win rates generally rose with the probability bands, but mean market probability and observed win rate did not match perfectly. This checks historical market probabilities; it is not a report of Sigma model performance.

The 15–20% band averaged 17.22% market probability and won 18.69% of the time, a 1.47-point gap. The 30%+ band was 40.23% versus 42.71%, a 2.48-point gap. Sampling variation, band construction and odds timing may all affect the differences, so they should not be presented as fixed mispricing.

Market probabilities are normalized within each race; these gaps are not betting returns.

Align the target, timestamp and full field

A win probability answers whether the runner wins. A place probability describes a different event because several runners can place. Normalized win probabilities should total about 100% across the field. In a race where the first three finishers qualify for place dividends, the field's place probabilities should total about 300%. A model ranking is not necessarily a probability either.

Pre-race information changes with withdrawals, going, weights and market updates. Before comparing two probabilities, check the race, prediction target, data cut-off and model version. An early model estimate and a later market price can use different information; their gap is not automatically an extra signal.

Three checks before reading any win rate

First, identify the event: win, place or something else. Second, identify the number: normalized probability, unnormalized score or a market probability derived from odds. Third, record the data cut-off. Forecasts become comparable only when their definitions and timing match.

Probability is useful because it makes uncertainty recordable and testable, not because it makes one race feel certain. Preserving every forecast before the result is known provides stronger evidence than publishing only the successful examples afterwards.