How the model finds value the market has mispriced

Ask most people what a sports prediction model does and they will say it predicts who wins. That is not what the Axia Model is built to do, and the difference is the entire reason it has an edge.

Two different questions

There are two questions you can ask about a football match, and they are not the same question.

The first is: who will win? The second is: is the price on this outcome higher or lower than it should be? A model built to answer the first question is competing with everyone else trying to call results. Most matches are genuinely uncertain, so such a model will be right a little more than half the time and feel clever doing it, and it will still have no idea whether the bets it implies are worth making.

A model built to answer the second question is doing something different. It is not trying to be right more often than anyone else. It is trying to be paid more than the true odds when it is right. The term for that is value: a bet has value when the true probability of the outcome is higher than the probability implied by the price. A model that finds value does not need to win most of the time. It needs to be paid properly for the times it does win.

The Axia Model answers the second question. It is a pricing engine first and a recommendation engine second, and almost everything distinctive about it follows from that.

How the model prices a match

Before the model looks at a single bookmaker price, it builds its own.

It takes the inputs that genuinely move match outcomes. Underlying performance data, the kind that measures how well a team actually played rather than what the scoreline happened to say. Recent form, weighted properly by recency and by the quality of the opposition rather than by reputation. Squad availability and the specific context of the fixture. From those inputs it produces its own probability for each market on the match: how likely the match is to stay under 2.5 goals, how likely the draw is, how likely each side is to win.

That internal probability is the model's honest opinion, formed without reference to the odds. Forming it independently is the whole point, and it is worth being clear about why. If the model looked at the bookmaker's price first, it would anchor to it. Anchoring is a well-documented effect: once you have seen a number, your own estimate drifts toward it. A model that anchors to the market cannot tell you the market is wrong, because it has quietly agreed with the market before it finished thinking. Independence is what preserves the model's ability to disagree, and the ability to disagree, accurately, is the only thing that can produce an edge.

The edge is the gap, not the pick

Only once the model has its own price does it look at the market.

Now it has two numbers for the same outcome: its own probability, and the probability implied by the bookmaker's odds. Converting odds into an implied probability is straightforward: decimal odds of 2.00 imply a 50 percent chance, odds of 4.00 imply 25 percent, and so on, by taking one divided by the price. The model does this for every outcome it has priced.

Most of the time, the two numbers are close. Bookmakers are good at their job, and an efficient price is one that closely reflects the true probability. When the model's number and the market's number agree, the model does nothing. There is no edge in agreeing with the market, and acting anyway just pays the bookmaker's margin for nothing.

A recommendation is generated only when the gap is large enough to matter: when the model's probability is far enough above the implied probability of the price that, repeated across many similar situations, acting on the gap should pay. The size of that gap, expressed as the advantage the model believes it has, is the edge, and the long-run profit it should produce is the expected value, or EV. The recommendation is the gap. The team that happens to be named in it is almost incidental.

This is why a fair description of the model is not that it predicts results. It finds prices the market has not fully sharpened, and ignores everything else.

A worked illustration

A simple, hypothetical example makes the mechanism concrete. The numbers here are illustrative, chosen to show the logic rather than drawn from a specific match.

Suppose the model prices a particular match and concludes the chance of it staying under 2.5 goals is 55 percent. That is the model's independent opinion, formed before it looks at any odds.

Now it looks at the market. One bookmaker is offering a price of 2.10 on under 2.5 goals. One divided by 2.10 is about 0.476, so that price implies a 48 percent chance. The model thinks the true chance is 55 percent. There is a gap of roughly seven percentage points between what the model believes and what the price assumes. That gap clears the bar, so the match becomes a recommendation, at that price.

Now change one thing. Suppose instead the best available price is 1.80. One divided by 1.80 is about 0.556, an implied chance of 56 percent. The model still thinks the true chance is 55 percent. Now the price is, if anything, slightly worse than fair. There is no gap to act on, so there is no recommendation, even though the model's view of the match has not changed at all.

That is the entire method in one example. The same opinion about the same match produces a recommendation at one price and silence at another. The model is not betting on the under. It is betting on the gap, and when the gap is not there, neither is the bet.

Why the biggest gaps are not the best bets

Here is where a full season of real data corrects a tempting assumption.

You might expect that the bigger the gap between the model and the market, the better the bet. Sort the 2025-26 record by the size of the modelled edge and that assumption does not survive.

Set aside the 20 to 35 percent row: at only seven recommendations it is far too small a sample to mean anything, and it is shown only for completeness. The rest of the table tells a clear story. The recommendations where the model saw a small-to-moderate edge did the work. The 5 to 10 percent band, the largest single group at 294 selections, returned close to 25 percent. The 0 to 5 percent band returned a healthy 16.5 percent. But the band with the very largest modelled edges, above 35 percent, lost money across the season.

That is not a contradiction. It is a signal, and it is one of the most useful things the season taught. When a model believes it has found an enormous edge, the most likely explanation is not that the bookmaker has made an enormous mistake. Bookmakers do not often leave 35-point gaps lying around. The more likely explanation is that something is wrong in the inputs: a price that has gone stale, a market that has not yet absorbed a team-news change, a fixture the model is misreading for a reason the model cannot see.

A naive system chases the biggest number on the screen. A mature model treats its own most extreme outputs with suspicion, because an extreme output is more often a warning than an opportunity. The Axia Model's profit comes from a high volume of modest, well-evidenced gaps, not from the outliers, and the data is the reason we are confident saying so.

Why value lives in longer prices

The season showed a second clear pattern, this time in the prices themselves rather than the modelled edge. Sort the record by the odds taken.

Read the strike-rate column and the return column together, because the relationship between them is the lesson. As the prices lengthen, the model wins less often, falling from 54 percent at short odds to 32 percent at the longest. And yet the return rises the whole way down the table. Short-priced selections, the ones that win most often and feel most comfortable to back, barely cleared breakeven and in fact lost 3.6 percent across the season. The longest-priced band, winning fewer than one time in three, returned more than 50 percent.

The reason is structural. Short-priced favourites attract the most money, because backing a likely winner feels safe, and that weight of money is exactly what the bookmaker sharpens hardest. The result is that the least value in the whole market sits on the outcomes that win most often. Longer prices attract less money and less scrutiny, so they stay looser, and a looser price is where a gap can survive.

The model is built to be comfortable with this. It will accept being wrong more often, in exchange for being paid properly on the occasions it is right. That is the opposite instinct to a tipping service selling the warm feeling of frequent winners, and the price-band table is the evidence for why the model is built the way it is rather than the comfortable way.

Why some leagues pay more than others

The same principle, that value lives where attention does not, shows up a third time when the season is sorted by competition.

The most heavily traded league in the dataset, the Premier League, returned the least. The lighter, less saturated leagues returned more. This is not a claim that the model understands Dutch or French football better than English football. The model applies exactly the same method to every league. The difference in returns is a difference in the markets, not in the model. A heavily traded market has been corrected by an enormous weight of money and leaves few gaps. A lighter market leaves more. The Bundesliga figure should be read with caution, as it sits on only 38 selections, but the overall direction is consistent with everything else here: the edge is largest where the crowd is smallest.

The discipline of declining

The hardest part of the method to see from the outside is the part where nothing happens.

On most matches, on most markets, the model finds no gap worth acting on, and it stays quiet. Across the 2025-26 season it published 719 recommendations. The number of matches it priced, examined and passed on without a recommendation was many times larger. A recommendation is not a quota to be filled on a schedule. It is what is left after everything that did not clear the bar has been discarded.

That restraint is the discipline. It is easy to have an opinion on every match. It is much harder, and much more valuable, to hold an opinion and then decline to act on it because the price does not reward it. A model that produces a recommendation for every fixture is not being thorough. It is being undisciplined, and it will spend most of its time paying the bookmaker's margin on bets with no edge.

Why level stakes

One more piece of discipline sits underneath the published record: every recommendation is measured to the same flat, level stake.

There are staking systems that vary the amount by confidence, and used carefully they have a logic. But a level stake does something important for a published record: it makes the record honest and easy to read. Every selection counts exactly as much as every other. No single result can be quietly inflated by having carried a larger stake, and no losing run can be hidden by having carried a smaller one. When the 2025-26 record shows 15.3 percent on turnover, that figure is the clean average of 719 equally weighted selections, and anyone can reconcile it. We make the wider argument for that kind of honesty in tracked, not claimed.

What this means for someone using the analysis

For a subscriber, the practical consequence of all of this is a particular shape of experience, and it is worth setting the expectation plainly.

The model will recommend selections that lose, often. It will recommend longer-priced outcomes that win less than half the time. It will be quiet on the marquee fixture everyone is discussing and active on a match nobody is talking about. None of that is the model misfiring. It is the model working exactly as designed, because value lives at longer prices, in quieter markets, and behind a strike rate below 50 percent.

The return comes from the gap between price and probability, compounded across a large number of disciplined selections and read over a season rather than a weekend. The full season that this method produced is set out in the 2025-26 season in numbers, and what it feels like to hold the method through a losing month is covered in what a 16 percent October taught us.