Learning goals and lesson guide
**Estimated reading time:** 8 min read.
**Level:** Intermediate. **Study time:** Allow 35–45 minutes with a calculator. **Prerequisites:** Percentages, expected value, variance and a paper journal. All records are fictional. Statistical intervals below are teaching models whose assumptions must be stated.
One record shows eight wins from ten selections. Another shows 550 wins from 1,000. The first percentage is higher, but it is based on much less information. Neither figure alone tells you whether the record is profitable, complete, independent or likely to continue.
This lesson develops a method for interrogating a win-rate claim. You will inspect the denominator, prices, sampling process and uncertainty, then write a conclusion that matches the evidence. You will also see why collecting more rows does not automatically fix selection bias or repeated observations of the same event.
## 1. Ask what counts as a trial
A win rate needs a numerator and denominator. Are pushes excluded? Are void selections included? Are partial wins counted as full wins? Are unresolved rows missing? Different conventions can produce different percentages from the same underlying ledger.
State the convention before calculating. A record with six wins, three losses and one push has a decisive-outcome win rate of six out of nine, or 66.67%, if pushes are excluded. Its fraction of all selections that won is six out of ten, or 60%. Neither should be presented without its definition.
For quarter lines and cash-out records, a single win-rate number may be especially unhelpful. Retain returns and stakes alongside outcome categories. A half win is not financially equivalent to a full win, even when both are positive net results.
## 2. Prices determine what a win is worth
Consider ten equal one-unit selections at 1.20 with eight wins and two losses. Winning profit is eight × 0.20 = 1.60 units; losses total two units. Net result is minus 0.40 despite an 80% win rate.
Now consider ten equal one-unit selections at 3.00 with four wins and six losses. Winning profit is eight units and losses are six, leaving plus two units despite a 40% win rate. These invented examples show why win rate cannot be evaluated independently of prices.
If odds vary, calculate each row's actual return. Multiplying total wins by one convenient average price can be wrong when winning and losing selections have different price distributions. The ledger is the primary record; summary statistics should be derived from it transparently.
## 3. Small samples can swing sharply
At eight wins from ten, one additional loss changes the fraction to eight from 11, about 72.73%. One additional win changes it to nine from 11, about 81.82%. A single observation has a large effect because the denominator is small.
At 800 wins from 1,000, one additional loss changes the fraction only slightly, to about 79.92%. A larger sample stabilises the observed fraction under comparable conditions. It does not prove that the underlying process stayed constant or that the observations were collected fairly.
Avoid treating a round number such as 100 or 1,000 as a universal certificate of reliability. The necessary information depends on the effect size, variability, dependence and question. A tiny claimed improvement may require far more evidence than a large descriptive difference.
## 4. Use an uncertainty interval as a model-based summary
For an independent Bernoulli model with a fixed success probability, a Wilson interval is one way to describe uncertainty around a sample proportion. With observed proportion h = wins/n and z = 1.96 for a conventional 95% interval, calculate the adjusted centre as (h + z²/(2n)) ÷ (1 + z²/n).
The half-width is z × sqrt[h(1 − h)/n + z²/(4n²)] ÷ (1 + z²/n). Subtract and add that half-width to the centre. A spreadsheet or statistical library can implement the arithmetic, but the model assumptions remain your responsibility.
For eight wins from ten, the approximate Wilson interval is 49.0%–94.3%. For 550 from 1,000, it is approximately 51.9%–58.1%. The first observed rate is more dramatic, yet its interval is much wider.
These intervals do not mean that 95% of future individual results will fall between the endpoints. An individual binary result is a win or loss. Nor do they guarantee the next period has a constant probability in that range. The confidence level describes repeated coverage of the interval method under its assumptions.
## 5. Know when the simple interval is inadequate
Sports selections can be correlated, have different probabilities and arise from a changing process. Multiple props on the same player or several markets from one match may share important sources of uncertainty. Treating them as independent identical trials can make an interval look more precise than justified.
A fixed-probability win-rate interval also ignores price differences. Even a well-estimated success fraction is not a full interval for financial performance when returns vary. The target statistic must match the claim you are evaluating.
At this stage, the practical response is to disclose the limitations and group the data sensibly. Later lessons introduce time-based validation and model scoring. Do not use a technically correct interval as decoration on a dataset that violates the assumed sampling process.
## 6. Selection bias is not repaired by a larger number
Suppose a record includes only selections that remained visible after the event, while deleted losses disappear. A large sample from that process is still biased. More observations can make the biased percentage look increasingly precise without making it truthful.
Other problems include beginning the record after a losing month, reporting only the strongest sport, excluding unavailable prices and counting hindsight selections. Ask how the sample was created, not only how large it is.
Predefine inclusion rules. For a paper study, record every eligible event or every published selection according to the stated scope. Keep no-selection decisions and missing-data reasons. A complete record may look less impressive, but it supports a more honest evaluation.
## 7. Repeated searching creates impressive-looking subsets
Imagine testing many combinations of league, weekday, price band and recent form, then reporting only the subset with the highest historical win rate. Even if no real predictive relationship exists, some subsets may look strong by chance.
The selected subset must be tested on later data that were not used to discover it. Record how many alternatives were tried and why the final rule was chosen. A historical pattern discovered after inspecting results is a hypothesis, not an independently validated finding.
Do not repeatedly redefine the sample after every disappointing result. That makes the method impossible to falsify. Keep a dated version of the rule and evaluate subsequent observations under that version.
## 8. A full audit exercise
You receive a fictional claim: “82% wins this month.” Ask for the count, prices, stakes, push treatment, date range, timestamps and complete record. Suppose it means nine wins from 11 selections at 1.15, with equal one-unit stakes.
The observed win rate is about 81.82%. Net result is nine × 0.15 − two = −0.65 units. The rounded headline concealed both a small denominator and a negative result. There may still be additional validity issues, but arithmetic alone already limits the claim.
Write a neutral audit conclusion: “The supplied 11-selection record contains nine wins, but at the stated prices and equal stakes it loses 0.65 units. The sample is small and does not establish future performance.” Avoid calling the author dishonest unless you have evidence of intentional deception; identify the measurable problem.
## 9. Practice and explained answers
1. Seven wins and three losses at 1.30 with one-unit stakes produce what net result? 2. Why should a percentage include its denominator? 3. Does an interval for win probability directly describe the next binary outcome? 4. Why can 100 correlated selections contain less information than 100 independent trials? 5. Can a large sample fix omitted losing records? 6. What should follow discovery of a strong historical subset? 7. Why preserve the rule version before collecting new results?
### Answer key
1. Seven × 0.30 − three = −0.90 units, despite a 70% win rate. 2. The count determines how much information supports the fraction and how much one new result can change it. 3. No. The interval concerns a model parameter; a single result remains binary. 4. Shared event conditions reduce the amount of distinct information and invalidate simple independence calculations. 5. No. More biased observations do not restore missing outcomes. 6. A prospective or held-out chronological test using the prewritten rule. 7. It prevents later outcomes from silently changing what the original method was claimed to be.
## 10. Completion check
Audit a fictional win-rate claim using four lenses: arithmetic, prices, sampling and uncertainty. Report what is established and what remains unsupported. A useful analyst can explain why an impressive percentage may be incomplete without assuming the opposite claim is automatically true.
Review [probability](https://betting.crazywingo.ph/article/implied-probability-bookmaker-margins-beginners), [returns](https://betting.crazywingo.ph/article/read-betting-odds-calculate-payouts), [market types](https://betting.crazywingo.ph/article/moneyline-spread-total-parlay-beginners), [market identification](https://betting.crazywingo.ph/article/sports-betting-basics-for-beginners) and [budget boundaries](https://betting.crazywingo.ph/article/betting-budget-limits-avoid-chasing-losses) as needed. Next comes the interpretation of opening, current and closing prices.