A win rate needs a sample count and an uncertainty estimate. A short run can differ widely from the underlying win probability. The useful sample size depends on the question and the assumptions.
Planning a sample: State the comparison rate (about 52.38% break-even at -110), the confidence method and level, whether the hypothesis is one-sided or two-sided, and the effect size you want to detect. A planned hypothesis test also needs a target power: the probability of detecting that effect if it exists. Correlated bets reduce the information in a raw bet count. There is no universal count for 95% confidence.
What this means: A 10-game winning streak is noise, not signal. A 50-game streak with 30 wins (60%) is suggestive but not conclusive. Real edge assessment requires hundreds of bets in similar market conditions.
Running hot vs. edge: The human brain pattern-matches aggressively. A bettor running at 58% over 100 bets may be experiencing normal variance — the 95% confidence interval around 100 bets at true 52.4% includes 58% wins. Before claiming edge, calculate confidence intervals on your results.
The Shark Snip model training backtests show calibration curves and Brier scores because win rate alone is insufficient evidence of model quality. CLV distribution is a better early indicator of edge than win rate in small samples.