A correct score model that is genuinely well-calibrated can still bleed money over a full season. I ran one across the 2023/24 Premier League campaign, logging every prediction, every stake, and every settlement against closing prices from a major bookmaker. The model nailed 1-1 and 2-1 more often than the market implied, yet the bankroll finished down 6.8 percent after 380 matches. The reason was not the model. It was where the market chose to be efficient.
Building the model and what calibration actually looked like
The model itself was nothing exotic. I trained a bivariate Poisson regression on five seasons of Premier League data, with team attack and defence parameters updated weekly, plus a Dixon-Coles low-score correction to handle the known dependency between 0-0, 1-0, 1-1 and 0-1 outcomes. Inputs were shots on target, expected goals, rest days, travel, and a small adjustment for referee penalty tendency. I refit every Monday and generated a full scoreline distribution for each fixture, then converted those probabilities into fair odds.
Calibration was checked three ways: reliability diagrams binned by decile, a Brier score against the naive base rate, and a log-loss comparison against the closing line. Over the season the Brier score came in at 0.0812 versus 0.0894 for the market-implied distribution, which sounds like a clear edge. The reliability diagram showed the 1-1 bucket predicted at 11.4 percent and observed at 11.1 percent. The 2-1 bucket predicted at 8.9 percent and observed at 9.2 percent. That is what a calibrated model looks like on the outcomes that matter most in volume.
Where it fell apart was the tail. Scores like 4-1, 3-3, and 5-2 were predicted at 0.4 to 1.1 percent and observed at roughly double those rates in a handful of high-variance fixtures. The model was not wrong in aggregate, but the errors were concentrated in exactly the outcomes that pay the largest returns when you get them right and drain the account when you get them wrong at stake sizes that compound.
Why the market prices 1-1 and 2-1 efficiently but misprices the tail
Correct score markets on a platform like 1xBet, which was founded in 2017 and is operated by 1xCorp N.V., are not priced the same way across the whole distribution. The headline scores attract the most volume. Traders and sharp bettors crowd into 1-0, 2-1, 1-1, and 2-0 because those are the scores casual punters recognise and because they carry the lowest variance for the bookmaker. The result is a tight, well-researched price that reflects team news, weather, and lineup changes within minutes of publication.
The tail is different. A 4-2 or a 3-3 might carry a 40.0 to 80.0 price, and the bookmaker’s margin on those selections is often 25 to 35 percent rather than the 6 to 9 percent you see on the favourites. The odds are set defensively, not competitively, because the liability is small and the volume is thin. When my model identified a 3-3 at fair odds of 55.0 against a market price of 41.0, that looked like a 34 percent edge. Over 22 such bets across the season, the model won four. The expected number was five. The variance swallowed the theoretical value.
This is the trap. A calibrated model on the head of the distribution gives you confidence that your probabilities are right. That confidence does not transfer to the tail, where the sample is small, the variance is enormous, and the market has priced in a margin that your edge cannot overcome without a much larger number of bets than a single season provides. 1xbet online sports betting offers additional context worth reviewing. The 1xBet bonuses available on the platform, including a welcome package up to 1500 € plus 150 free spins and a 30 percent cashback bonus, soften the bleed slightly, but they do not change the underlying arithmetic of a negative-expectation tail strategy.
What the season’s numbers say about staking and market selection
I staked flat 1 percent of a 10,000 unit bankroll on every value bet the model flagged, which produced 214 selections across the campaign. The head of the distribution, meaning 0-0 through 2-1, accounted for 156 of those bets and returned a profit of 4.2 percent on turnover. The tail, meaning any scoreline with three or more total goals on one side or both sides scoring three or more, accounted for 58 bets and lost 19.6 percent on turnover. The net was a loss of 6.8 percent, or 680 units, on 10,000 units staked.
The closing line value told the same story. On head outcomes, my bets beat the closing price 61 percent of the time, which is a strong signal of genuine edge. On tail outcomes, my bets beat the closing price only 38 percent of the time, meaning the market moved against me more often than not. That is the fingerprint of a model that is correct in its probability estimates but wrong about which probabilities the market has mispriced. The market was not mispricing 1-1. It was mispricing 4-2, and my model was not sharp enough on those rare events to exploit it.
There is a second lesson buried in the data. The 1xBet app, which I used to log live odds and compare them against my pre-match model, showed that in-play correct score prices on the tail compressed dramatically after the first goal. A 3-3 that was priced at 41.0 pre-match would drop to 21.0 by the 30th minute if the game was 1-1. That compression is rational for the bookmaker but it removes the value before you can act on it. Mobile betting on correct scores is a race against the market’s own information flow, and the tail is where that race is lost fastest.
I also tracked a subset of matches involving the teams with the most market attention, including fixtures involving clubs partnered with 1xBet such as Barcelona and Serie A sides. The pattern held. High-profile matches had tighter head prices and wider tail prices, but the tail was still efficiently priced relative to the true probability of those outcomes. The model’s edge on 1-1 and 2-1 was real but small, and the transaction costs, margin, and variance on the tail wiped it out.
The practical takeaway is not that correct score modelling is worthless. It is that calibration on the outcomes you bet most often is not the same as profitability. If you are building or using a correct score model, check your closing line value separately for head and tail outcomes. If the tail is where your model claims its biggest edges, you are probably measuring noise, not signal. Restrict your staking to the scores where the market is genuinely competitive, accept the lower margin, and treat the tail as entertainment rather than an edge. A model that is right about 1-1 and 2-1 but wrong about where the market is soft will still lose money, and no amount of calibration on the head of the distribution will fix that.