The 2025-26 season in numbers: what 719 recommendations tell us

Most performance write-ups lead with one number and stop there. This one starts with the number, then spends the rest of its time underneath it, because the headline figure is the least interesting thing about the season.

The headline, and why it is not the whole story

Across the 2025-26 football season the Axia Model produced 719 recommendations. Each one was published before kick-off and settled against the final result. To a flat, level stake the season returned 15.3 percent on turnover. The strike rate was 47.3 percent. The average price taken was 2.50.

Those four numbers describe the season. They do not explain it. A 15.3 percent return on turnover is a strong year for a sports model, and it is also the kind of figure that invites the wrong question. People see it and ask whether it will be 15.3 percent again. That is the wrong question, because no single season's return is a forecast.

The better question is how the number was produced, and whether the process that produced it is stable. A useful term here is return on stakes, sometimes shortened to yield. It is the season's profit divided by the total amount placed, and across 2025-26 that figure was 15.3 percent. It is not the same as a return on a bankroll, and it is not an annual interest rate. It is simply this: for every 100 units the model put to work, 15.3 came back as profit on top.

On the test of whether the process is stable, the season holds up. The return did not come from a single hot fortnight. It came from nine profitable months in ten, spread across 136 match days, with the worst day costing 6.2 units and the best returning 11.9. A model that makes money in nine months out of ten, at level stakes, across more than 700 selections, is showing something closer to a method than to luck. The rest of this article is the evidence for that claim.

The season, month by month

Read down that column and a pattern appears that the headline hides. There is no smooth 15 percent every month. The season has phases, and each phase teaches something.

The opening, August and September, was a steady start: 98 recommendations returning a shade over 14 percent in each month. Nothing dramatic, just the model doing its job at a modest volume while the new season's form data was still thin.

Then October, the only month in red, which gets its own section below.

The winter, November through January, was the engine room. Those three months carried 261 recommendations, well over a third of the entire season, and returned 16.1, 8.0 and 20.2 percent. This is the period when the European leagues are in full, dense fixture rhythm, and it is when the model has the most matches to price and the most opportunities to act on. Volume matters for a model with a modest per-bet edge, and the winter is where the volume lives.

February is the most quietly instructive month in the table. It was the busiest month of the season, 97 recommendations, and it returned 0.4 percent. A whole month of work, the largest single batch of selections, and it landed almost exactly on breakeven. February is the month that tells you what a real edge feels like from the inside. It is not a steady drip of profit. It is long flat stretches punctuated by months that do the heavy lifting, and the discipline is in treating the flat month and the strong month exactly the same way.

The spring, March and April, was the season's standout: 157 recommendations returning 36.7 and 33.6 percent. Two months like that will always draw the eye, and they should be read with the same caution as October but in the opposite direction. They are not the model's true rate any more than October was. They are the upside half of the same variance whose downside half was October.

May, to the 18th, returned 7.0 percent across 81 recommendations, a solid partial month to close the reporting period.

October, and what a losing month is for

October was the only month in the red: 25 recommendations, down 15.9 percent, a loss of just under 400 units. It is the most important month in the whole record, precisely because it lost.

A losing month is a test of two things. The first is whether the model changes its behaviour under pressure. It did not. The selection criteria in November were the same as in September. The model did not chase the loss with longer prices, and it did not quietly raise the implied stake on the next batch of selections, because it has no emotional state to manage. The second test is whether the edge survives the drawdown. It did. November returned 16.1 percent, and the season went on to its strongest two months in March and April.

There is a sample-size point worth making plainly. October carried only 25 recommendations, the smallest batch of any month. In any activity where the edge per bet is modest, a few dozen outcomes can land badly without anything being wrong with the method. October was not a model failure. It was the part of a real distribution that nobody puts in a brochure. We examine that month in full in what a 16 percent October taught us about a real edge.

Where the edge came from: by league

The season's profit was not spread evenly across competitions. Three leagues did the heavy lifting.

The pattern is consistent with how the model is designed to think. It looks hardest for prices the wider market has not fully sharpened. The most heavily traded league in the dataset, the Premier League, returned the least, at 6.7 percent, because that is where bookmaker prices are tightest and genuine mispricing is rarest. The Eredivisie and Ligue 1, less saturated with money and attention, returned more than three times that.

This is not a claim that the model understands Dutch football better than English football. It is the opposite. The model applies the same method everywhere. The difference in returns is a difference in the markets, not in the model. A heavily traded market has been corrected by a great deal of money and leaves fewer gaps. A lighter market leaves more. The Bundesliga figure, 31.3 percent, should be read with care: it sits on only 38 recommendations, a small enough sample that a handful of results swing it, and it should not be treated as a durable rate.

Where the edge came from: by market type

The season also leaned in a clear direction by the kind of market it bet.

The shape here is deliberate. Selections on whether a match would stay under 2.5 goals made up the largest single block of the record, 405 of the 719, and returned a steady 10.2 percent. The goals markets are where the model finds the most frequent, most reliable small edges, and they carried the volume of the season.

The higher-percentage lines, the draw at 43.5 percent and the home win at 26.7 percent, sat on far smaller samples and at longer prices. They are higher variance by nature: the draw in particular is an outcome the casual market consistently underbacks, which is why it can be profitable, but it lands less than two times in five. The season's design is visible in this table. The steady goals markets provide the base, and a smaller set of longer-priced match-result selections adds the upside. We explain why the model is built that way in how the model finds value the market has mispriced.

What the prices tell you

Sort the season not by league or market but by the odds taken, and one of its clearest lessons appears.

Read the strike-rate column and the return column together. As the prices get longer, the model wins less often, falling from 54 percent at short odds to 32 percent at long odds. And yet the return rises the whole way down the table. Short-priced selections, the ones that win most often and feel most comfortable, 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, so the market sharpens them hardest and leaves the least value there. Longer prices get less attention and stay looser. The model is built to be comfortable being wrong more often in exchange for being paid properly when it is right. That is the opposite instinct to a tipping service selling the warm feeling of frequent winners, and the season's data is the argument for it.

How accurate were the probabilities

A model that recommends value needs its probabilities to be honest, so it is worth asking how well the model's stated likelihoods matched what actually happened. The technical word for this is calibration: a well-calibrated model that says an outcome is 60 percent likely should see that outcome happen close to 60 percent of the time across many such cases.

On that test the season was sound but not flawless, and it is worth being plain about both halves. The recommendations the model rated as more likely did, broadly, win more often than the ones it rated as less likely, so the probabilities are correctly ordered. In the lower and middle ranges the model tracked reality closely. In the upper-middle range, around 60 to 75 percent, the model was somewhat overconfident: a number of those selections landed closer to a coin flip than the rating implied. The small group of highest-confidence picks, rated above 80 percent, performed strongly.

That upper-middle softness is a known area of model work, and it matters less to the bottom line than it might first appear. Profitability comes from the price taken relative to the true probability, not from raw strike rate, which is why the season returned 15.3 percent even with a sub-48 percent win rate. Reporting the calibration honestly, rather than only the return, is part of how we hold the record to a standard a serious reader can trust.

The shape of the risk

A headline return tells you the destination. It does not tell you the road. For that you need the drawdown.

The largest peak-to-trough fall in the season's running profit was 14.1 units, measured against a one-unit level stake. The longest unbroken run of losing recommendations was nine; the longest winning run was eight. The best single day returned 11.9 units and the worst cost 6.2.

Those numbers describe what holding the model through the season actually felt like from the inside. A 14.1-unit drawdown is not alarming for a strategy of this kind, but it is real, and anyone evaluating the model should size their involvement so that a fall of that order is something they can sit through without changing their behaviour. A return figure shown without a drawdown figure is half a picture, and usually the more flattering half. We publish both.

What the season says about the model

Three conclusions, none of which depends on the headline being repeated.

First, the edge is modest per recommendation and only becomes a return through volume and discipline. The average selection carried a small mathematical advantage. Compounded across 719 of them at a level stake, that advantage produced the season. There is no version of this where a handful of selections carries the year, and the month-by-month table proves it.

Second, the model is stable under stress. It did not flinch in October and it did not over-extend in March. Consistency of behaviour is the trait that lets a record be trusted, and it is the trait a human-led tipping service tends to lose first, because a human after a bad month is under real pressure to do something visible about it.

Third, the record is the product, not a marketing asset bolted onto one. Every figure in this article reconciles to a line in a recommendation log that was written before the matches were played. That is the standard we hold the season to, and we make the wider case for it in tracked, not claimed.

What we are carrying into 2026-27

The model continues, the league coverage widens, and the published record keeps running in the open. The 2026 World Cup falls inside the next reporting period and brings a wave of casual money into the market, which historically creates the kind of mispricing the model is built to find.

The most useful thing we carry forward is the habit shown in the table above: publish the poor months as plainly as the strong ones. A season is not 15.3 percent. It is a steady autumn, a poor October, a winter that did the volume, a flat February, a strong spring, and the discipline to treat all of them the same way. That discipline, more than any single number, is what the 2025-26 season is evidence of.