“72% accuracy.” It is the most common number on a prediction service's landing page and, on its own, it is close to meaningless. A hit rate says how often picks won. It does not say at what price, over how many picks, over what period, on which markets, or whether the picks were the same ones paying customers received. Each of those can turn 72% into a losing record or 40% into a profitable one.
Break-even depends on the odds
Every decimal price implies the hit rate needed to break even at a flat stake: one divided by the odds. At 1.91, the standard price on a spread or total, it is 52.4%. At 1.50 it is 66.7%. At 1.25 it is 80.0%. At 3.00 it is 33.3%. A win rate is informative only when it is placed next to the break-even rate of the prices it was achieved at, and a service that publishes the first number without the second is asking you not to do that arithmetic.
- Picks
- 100
- Hit rate
- 70.0%
- Average odds used
- 1.30
- Break-even hit rate at 1.30
- 76.9%
- Return: 70 wins × 1.30
- 91.0 units
- Staked
- 100.0 units
- ROI
- −9.0%
Seventy wins in a hundred, and nine units lost. Short-priced favourites, double chance at 1.25, unders on low-scoring leagues: any market where the price is low produces a high hit rate and, without an edge, a loss.
- Picks
- 100
- Hit rate
- 40.0%
- Average odds used
- 2.80
- Break-even hit rate at 2.80
- 35.7%
- Return: 40 wins × 2.80
- 112.0 units
- Staked
- 100.0 units
- ROI
- +12.0%
Sixty losses in a hundred, and twelve units won. Nobody puts “40% accuracy” on a landing page, which is why the number on the landing page is the wrong one to read.
This is why every rate on PickAuditor is shown with ROI at a flat one-unit stake beside it, and why the leaderboard orders by ROI rather than hit rate by default.
Sample size
The second problem is n. Suppose a service posts 60% over 30 picks at 1.91, where break-even is 52.4%. That is 18 wins against an expected 15.7. The two-sided p-value against break-even — the chance of a gap at least that large if the service had no edge — is about 0.40. Two times in five, a service with nothing behind it produces a record that good. The same 60% over 300 picks is 180 wins against 157, and the p-value is about 0.008: once in roughly 125 tries. Same hit rate, entirely different evidence.
The p-value column on a profile is that calculation, using the average implied probability of the odds actually used. It is not a forecast, and a small value does not say the edge will continue; it says the record so far is unlikely to be chance. Below 30 graded picks PickAuditor does not rank a service at all and marks the row amber, because at that size almost any hit rate is consistent with no edge.
Which picks were counted
A hit rate is also a choice of denominator. Things that inflate it:
- Counting pushes as wins, or voiding losses for reasons never applied to wins.
- Mixing markets: a few dozen double-chance and heavy-favourite picks lift the overall rate while the picks a reader would follow sit below break-even.
- Choosing the period: “this month” when the month is good, “all-time” when it is not, or the reverse.
- Choosing the model: a service that runs many models and headlines the one that is hot this quarter will always have a hot model to headline.
- Grading after the fact: picks published after kickoff, or edited after the result.
- Premium against public: a record of picks no reader could see, with the free picks graded differently or not at all.
PickAuditor's rules address each. Pushes are excluded from hit rate and worth zero in ROI; results are shown per market and per sport with n; every period is shown; sub-models get their own record; a pick captured after kickoff is void; and only public-tier picks are counted, with that stated on every page that shows a record.
What to read instead
- ROI at a flat stake, with n and period. This is the number the hit rate was standing in for.
- Average CLV, which asks whether the market agreed with the picks by kickoff and settles faster than results do. See closing line value.
- The p-value, as a check on whether n is large enough to mean anything.
- The claim check on a service's profile, which puts the service's stated figure next to the audited one with the scope of each: “72% NBA accuracy” against the NBA match-winner hit rate over 365 days, n shown.
- Max drawdown and the longest losing streak, which describe what following the record would have felt like. See how to read a track record.
Hit rate is still shown on every row, because it is a fact about the record. It is never shown alone. The methodology defines each figure; the leaderboard lets you order by any of them.