Why Can a 64% Accurate Sports AI Still Lose Money?
Because accuracy ignores price. A bet at decimal odds d breaks even at a win rate of 1 divided by d, so a model that mostly picks favourites at 1.45 needs 69 percent just to break even, and 64 percent loses about 7 percent of stakes. Profit depends on calibrated probability against the market price, not hit rate.
Accuracy is not the number that pays
A YouTube short from the KV-Cache channel opens with a model that picked 64 percent of past games correctly and still lost money. Its explanation: bookmakers run models too, sharp money moves the line before you bet, and the payout is shaved, so even-money outcomes pay 1.91 instead of 2.00.
The core point is right, and it reduces to one comparison. A bet at decimal odds d breaks even when it wins 1 divided by d of the time. Accuracy only means something next to that break-even rate, and the break-even rate depends entirely on the prices of the games the model picked. A 64 percent hit rate is excellent at odds of 1.91 and a losing record at odds of 1.45.
Break-even rates by price
The break-even win rate is the implied probability of the price you took:
At 1.91 (American minus 110), break-even is 52.4 percent. At 1.50 (minus 200), it is 66.7 percent. At 1.45 (about minus 222), it is 69.0 percent. At 1.25 (minus 400), it is 80.0 percent.
Models that output a win probability and pick the more likely side will mostly pick favourites, because favourites are by definition the more likely side. Their accuracy comes largely from that base rate. Picking the market favourite in every game, with no model at all, would also produce a high hit rate, and it would lose roughly the bookmaker's margin over time.
What the 4.7 percent in the video means
Two sides at 1.91 imply 52.36 percent each, 104.7 percent in total. The extra 4.7 points is the overround. The bookmaker's theoretical margin on balanced action is slightly smaller, about 4.5 percent of stakes, because the overround is measured against probability rather than money. Either way, you need to win about 52.4 percent at 1.91 just to break even.
A worked example: two 64 percent models
These numbers are illustrative. Two models each make 100 one-unit picks and each gets 64 right.
Model A picks favourites. Its picks average decimal odds of 1.45. Sixty-four winners return 64 times 1.45, or 92.8 units, on 100 staked. That is a loss of 7.2 units, a return of minus 7.2 percent, because it needed 69 percent to break even.
Model B picks near-even games. Its picks average 1.91. Sixty-four winners return 122.2 units on 100 staked, a profit of 22.2 units. It needed only 52.4 percent.
Same accuracy, opposite results. Model B's record would also be extraordinary: a sustained 64 percent at minus 110 is far beyond what markets usually allow, which is why a claim like that deserves suspicion rather than celebration. In practice, a reported 64 percent almost always looks more like Model A.
The general formula replaces accuracy entirely. The expected return on one bet is the true win probability times the decimal odds, minus one. Summed over bets, that is what decides profit. Accuracy is a by-product of which games were picked.
Calibration matters more than hit rate
Expected return needs a probability, not a pick. If the model says 75 percent and the true chance is 64 percent, it will see value in prices that have none. A model is calibrated when outcomes it rates at 70 percent happen about 70 percent of the time, across every probability band.
Two models can share the same 64 percent accuracy and differ completely in calibration. One might rate its picks at 64 percent on average and be right; another might rate them at 80 percent and be badly overconfident. The second will bet too often and at prices that look attractive only because its probabilities are inflated. Accuracy cannot see this; a reliability plot, log loss or Brier score can.
The phrase "past games" in the video matters too. A hit rate measured on games the model trained on, or on features that include information not available before kickoff, will not hold on new games. Test on later seasons the model never saw.
What to measure instead
Compare the model's calibrated probability with the market's margin-free probability for each game. Bet only where the model's probability times the offered odds exceeds one by more than the error in the model's estimates. Then track closing line value, which shows whether the prices taken were better than the market's final view, alongside profit and a proper scoring rule such as log loss against the market's own probabilities.
That comparison needs prices, closing lines and results for every game the model scored. Parlay Ledger provides real-time and historical odds, game results and player statistics across bookmakers in a unified format, which covers the inputs for this kind of evaluation.
Frequently asked questions
- Why can a 64 percent accurate betting model lose money?
- Because profit depends on the prices of the games it picked, not on accuracy alone. Favourites win often but pay little. If a model's picks average odds of 1.45, it needs 69 percent to break even, so 64 percent loses about 7 percent of stakes. The same accuracy at odds of 1.91 would be very profitable.
- What win rate do you need to break even at minus 110?
- About 52.4 percent. Minus 110 is decimal 1.91, and the break-even rate is 1 divided by the decimal odds. Both sides at 1.91 sum to 104.7 percent implied probability, and the extra is the bookmaker's overround. Shorter prices need higher win rates: 66.7 percent at 1.50 and 80 percent at 1.25.
- Does picking favourites make a model look better than it is?
- Yes. Favourites are the likelier side by definition, so a model that mostly picks them earns a high hit rate from the base rate alone. Choosing the market favourite in every game, with no model, would also score well on accuracy while losing roughly the bookmaker's margin. Judge a model on returns and calibration against prices.
- What is calibration in a sports prediction model?
- Calibration means stated probabilities match observed frequencies: outcomes rated 70 percent happen about 70 percent of the time. It matters because expected value uses the probability, not the pick. An overconfident model sees value in prices that have none. Check it with reliability plots, log loss or Brier score on out-of-sample games.