Why a 66% Win Rate Can Still Lose Money: ROI vs Win Rate in AI Football Predictions
The one-sentence version: Win rate tells you how *often* a predictor is right. ROI tells you whether being right that often actually *makes money* at the odds on offer. They are not the same thing, and a predictor can have an excellent win rate while quietly losing money on every 100 bets. This article shows exactly how, using one predictor's own published numbers.
If you are trying to choose between AI football prediction sites, the number you see first is almost always a win rate — "68% accuracy," "73% banker success," "2 in 3 correct." It is the headline because it sounds impressive and it is easy to understand.
It is also the wrong number to judge a predictor by.
The metric that determines whether following predictions makes or loses money is return on investment (ROI) — the profit or loss per unit staked, measured at the actual odds you would have bet. A predictor can be right most of the time and still hand back money on every cycle, because *how often you win* and *how much you win when you win* are two different things. Win rate measures the first. It says nothing about the second.
To make this concrete rather than theoretical, we work through a real, public example below. The point is not to attack any company. The point is to show you a metric trap that almost the entire prediction industry — us included, before we learned better — has fallen into at some point.
The mechanic: why a coin that lands heads 66% of the time can be a losing bet
Start with the arithmetic, because it is the whole argument.
Imagine a predictor that wins 66 out of every 100 bets. Sounds like free money. Now add the missing piece: the average odd on those bets is 1.46 (decimal). That is a heavy favourite — the kind of pick where the "obvious" team is priced short because everyone can see they should win.
Run the 100 bets at a flat 1-unit stake:
- 66 winners × (1.46 − 1) profit each = +30.4 units
- 34 losers × 1 unit each = −34.0 units
- Net result: −3.6 units
You were right two times out of three — and you still lost roughly 3.6% of everything you staked. Not because the predictions were bad, but because the odds were too short to pay for the times you were wrong. At odds of 1.46, you need to win about 68.5% of the time just to break even. A 66% win rate falls short of that line, so it bleeds money slowly and invisibly.
This is the core idea, and it is worth stating plainly:
Every set of odds has a break-even win rate. Beat it and you profit; miss it and you lose — no matter how high the raw win rate looks. At odds of 2.00 you need 50%. At 1.46 you need ~68.5%. At 1.25 you need 80%. A win rate is only meaningful next to the odds it was won at.
Win rate on its own is a vanity metric: true, verifiable, and completely uninformative about profit.
A worked example: NerdyTips's real, auditable 66.6%
The cleanest way to show this is with a predictor that publishes real data. NerdyTips is a well-known AI football tips site, and to its genuine credit it does something most tipsters never do: it publishes its raw match-by-match data openly, on GitHub, for anyone to check.
Let us be completely clear before going further, because fairness matters here:
- NerdyTips's headline win rate is real. Their homepage states that 66.6% of 271,154 matches predicted since May 2021 finished as forecast, plus a higher-confidence "Banker" tier quoted around a 73% win rate.
- It is auditable, and it checks out. Their public dataset (github.com/Dani-Ionescu/nerdytips-data) includes
allMatches.xlsx(roughly 168,295 rows from 2024–2026, each with the predicted market, the odds, and the result) and aprogress.csvaggregate. Recompute the win rate from those rows and you land on the same ~66.6%. The number is not inflated, cherry-picked, or fake. They deserve credit for publishing enough data to be checked at all — most sites publish nothing.
So this is not a "gotcha." The win rate is honest. The problem is what the win rate *doesn't* tell you — and NerdyTips, like almost everyone, publishes the win rate but not the ROI.
Deriving the ROI from their own odds
Because the dataset carries the odds on every single tip, we can do the one calculation the site doesn't display: settle every tip at a flat 1-unit stake using *their own published odds*, and total the result.
The average winning odd across the dataset is about 1.46 — these are predominantly short-priced favourites. Applying each bet's own odds, row by row, across all ~168,295 settled tips:
| NerdyTips segment | Win rate | Flat-stake ROI (from their odds) |
|---|
| Best Tip (overall) | ~66.6% | −4.2% |
|---|---|---|
| 1X2 (match result) | — | −5.1% |
| Over/Under (goals) | — | −4.4% |
| BTTS (both teams to score) | — | −6.1% |
| "Banker" high-confidence tier | ~70.7% | −2.2% |
First, no confidence tier is profitable. The highest-conviction "Banker" tier still returns −2.2%. Higher confidence buys a higher win rate, but the market prices those safer picks even shorter, so the ROI stays negative. There is no filter, tier, or market inside this dataset that turns a profit at flat stakes.
Second, the overall −4.2% is slightly worse than the round-number "−3.6%" from our earlier 100-bet illustration. That is not an error — it is a real and instructive effect. When you settle bet-by-bet instead of applying one blended average, the wins concentrate at the *very shortest* odds (the easiest picks win most often, and they also pay the least). So the profit you collect on winners is smaller exactly where winners are most common, and the true per-bet return comes out a little worse than the naive average suggests. The more short-priced favourites dominate a portfolio, the more this bites.
None of this makes 66.6% a lie. It makes 66.6% irrelevant to the question you actually care about: over a year of following these tips at these odds, does your bankroll go up or down? On this data, it goes down by about 4% of everything you stake — a genuine, published win rate quietly attached to a negative return.
Why the whole industry leads with win rate
If ROI is the number that matters, why does almost every prediction site lead with win rate instead?
Because win rate is the number that is easiest to make look good, and hardest for a casual reader to challenge.
- You can lift a win rate simply by predicting shorter favourites. Tip only the heaviest favourites and your win rate climbs toward 70–75% automatically — while your ROI sinks, because you are paying a premium for "safe." Win rate rewards exactly the behaviour that destroys ROI.
- Win rate has no downside axis. It cannot show a loss. ROI can be negative and print it in red; a win rate is always a cheerful percentage between 0 and 100.
- Win rate needs no verification to sound credible. "68% accurate" requires no odds, no stake log, no settled bets — just a claim. ROI is only meaningful if you can see the actual bets and prices behind it, which forces a site to publish the very data most would rather not.
So win rate persists not because it is useful, but because it is *flattering and unfalsifiable*. That is precisely why you, as someone choosing where to spend attention or money, should treat a lone win rate as a yellow flag, not a green one — and ask the follow-up question the number is designed to distract you from: *at what odds, and with what ROI?*
How to evaluate any AI football prediction site (a checklist)
You do not need a maths degree to apply the lesson. Five questions separate a predictor worth following from a flattering headline:
- 1Do they publish ROI at all — not just win rate? If the only number on the page is a win rate, assume the ROI is unflattering. Sites publish their best number, and win rate being the *only* number is itself information.
- 2What are the average odds behind the win rate? A 65% win rate at average odds 1.40 is a loser; a 55% win rate at average odds 2.10 is a strong winner. Always pair the two. If they won't tell you the odds, the win rate is unjudgeable.
- 3**Is the ROI from *real bets placed*, or from a backtest? A model's theoretical ("signal") ROI and the ROI you actually capture after real prices, timing, and limits are different numbers. Honest sites show both and explain the gap.
- 4Can you download and recompute the underlying data? NerdyTips gets full marks here — the data is public and auditable. That is the standard. If you cannot check the numbers yourself, they are marketing, not evidence.
- 5Is the sample large enough to mean anything?** A 90% win rate over 20 bets is noise. Thousands of settled bets, published, is signal. Look for scale *and* transparency, not one or the other.
Run any prediction site through those five questions and the marketing collapses into either evidence or a blank space where the evidence should be.
How OddsFlow reports itself (and why our win rate is deliberately ordinary)
We hold ourselves to the checklist above, which means being blunt about our own numbers.
Our win rate is deliberately ordinary — around 55%. We are not chasing a 70% headline, and here is why: chasing win rate means tipping short favourites, and tipping short favourites is how you end up right most of the time and poorer every month. We would rather be right *less* often on lines that are actually mispriced.
That is the entire strategy. OddsFlow only signals positive expected value (+EV) bets — lines where our Dixon-Coles model estimates the true probability is higher than the bookmaker's price implies. Those bets often sit at longer odds, so we lose more individual bets than a favourite-tipping service does. But when we win, we win at prices that more than pay for the losses. That is what turns a 55% win rate into a positive real-money ROI of roughly +10.3%.
Crucially, that ROI comes from real money actually staked at sportsbooks — on the order of ~3,482 real bets, with one downloadable PDF slip per bet as proof — not from a backtest. We publish two numbers on purpose:
- a signal ROI (how the model's picks would settle at the flagged price), and
- a real-money ROI (what we actually captured after real execution).
The gap between them is honest friction — timing, limits, price movement — and we would rather show it than hide it. You can see the live figures, the methodology, and the sample size on our Accuracy page, the audit trail on our Verification page, watch signals form in real time in the Live Signal Room, and download the entire settled record from our open dataset at github.com/oddsflowai-team/oddsflow-transparency. Recompute any of it yourself — that is the point of publishing it.
We are not claiming to be uniquely virtuous, and we are not claiming a lower win rate is automatically better. We are claiming one specific thing: a ~55% win rate attached to a published, real-money, positive ROI is a more useful signal than a 66% win rate with no ROI at all — because the first tells you whether you would have made money, and the second cannot.
The takeaway
- Win rate answers "how often is it right?" ROI answers "does being right pay?" Only the second determines whether you make money.
- Every set of odds has a break-even win rate. At odds 1.46 it is ~68.5%, so a genuine 66.6% win rate — like NerdyTips's real, auditable number — still returns roughly −4% on turnover. The win rate is honest; it is simply not the number that decides profit.
- A lone win rate is a vanity metric. Judge any AI football predictor by ROI *at real odds*, on a large, downloadable, recomputable sample — ideally from real bets, not a backtest.
- Ask five questions of any site: Do they publish ROI? At what odds? Real bets or backtest? Can you download the data? Is the sample big enough?
Being right often feels like winning. Whether you are actually winning is a different measurement — and it is the only one that shows up in your bankroll.
*This article is for information and education only. It is not betting advice, and nothing here guarantees a profit. Past performance does not predict future results, all betting carries risk, and you should never stake money you cannot afford to lose. 18+ only. All figures cited for NerdyTips are drawn from their own public homepage and GitHub dataset; the ROI figures are derived from their published odds and are our calculation, not theirs. Have a question about how we measure ROI? Join the OddsFlow Community or message us on Telegram.*