Put an interval around every rate
standard error ≈ √( p(1 − p) / n ) 95% interval ≈ p ± 1.96 × SE
| Record | Hit rate | ≈95% interval |
|---|---|---|
| 11–9 | 55.0% | 33.2% – 76.8% |
| 55–45 | 55.0% | 45.2% – 64.8% |
| 550–450 | 55.0% | 51.9% – 58.1% |
At 100 results, a 55% hit rate is still consistent with a coin flip. At 1,000 it starts to separate. Records presented without their sample size — or without prices — should be read as anecdotes.
How PropBetEdge surfaces handle this
- Minimum samples before claims. Calibration views wait for a minimum number of graded calls; publication gates wait for a minimum sample before a model's picks become official.
- Sample-size tiers on inputs. Fight DNA labels confidence by bouts and observed seconds, and withholds a stat rather than showing a misleading zero — see Fight DNA.
- Thin samples capped. UFC confidence is capped at Medium when either fighter has a thin UFC history, however large the probability.
Regression to the mean
Extreme early results — a 14-for-18 start, a reliever's 0.00 ERA — are partly skill and partly luck. The luck part does not repeat, so later results drift back toward the underlying level. Models handle this by shrinking small samples toward a prior; readers should do the same with any short record.
Limitations
Intervals here assume independent, identically distributed results. Real records mix prices, sports and seasons, which usually makes the true uncertainty larger, not smaller.