Nine months up, one down: what a 16% October taught us about a real edge
The 2025-26 season returned 15.3 percent. Nine of its ten months were profitable. This article is about the tenth, because the tenth month is the one that proves the other nine.
What October actually was
For 25 recommendations across October 2025, the Axia Model lost money. The month finished down 15.9 percent on stakes, a fall of just under 400 units. The strike rate for the month dropped to 40 percent, below the season's average of 47.3 percent. After two solid opening months in August and September, each returning a little over 14 percent, the running profit turned down for the first time.
It is worth being precise about what that was and what it was not. It was not a model failure. It was not the edge disappearing. It was a normal run of variance in a small sample, and to see why, it helps to look closely at the mathematics underneath a betting model, because that mathematics is the whole reason a month like October is expected rather than alarming.
Variance, and the mathematics of a small edge
A betting model like Axia's does not win most of its bets. Across the full 2025-26 season it won 47.3 percent of them. It is profitable not because it wins often but because, when it wins, it is paid more than the true odds. The term for that is a modest edge per bet: each individual recommendation carries only a small mathematical advantage, and the season's return is that small advantage compounded across a very large number of selections.
The consequence of a modest edge is that the result of any short run of bets is dominated by chance, not by the edge. Think of it like a weighted coin that lands in your favour 52 times in 100 over the very long run. Across ten flips, that weighting is almost invisible: you might see three heads, you might see seven, and neither tells you anything. Across ten thousand flips, the weighting is almost all you see. The edge is real at every scale. It is only legible at large ones.
October carried 25 recommendations. That is the ten-flips end of the spectrum, not the ten-thousand. A modest edge applied to 25 outcomes can easily produce a losing month, a flat month or a strong month, and which one you get is mostly noise. This is not a special pleading invented after a bad result. It is a basic and well-understood property of any strategy with a small edge and a finite sample, and it is the reason the only honest way to read a model is over a season, not over a month. October was the part of a real distribution that nobody puts in a brochure, and a method that could never produce an October would be a method to distrust, because no genuine edge in an efficient market is large enough to rule a bad month out.
What the model did next
This is the part that matters most, and it is the part that is easy to skip past.
The model did nothing different.
It did not lengthen its prices to win the loss back faster. It did not quietly raise the implied stake on the next few selections to chase the deficit. It did not narrow into the one league that had been kinder, or widen into markets it does not normally touch. The selection criteria it applied in November were precisely the criteria it had applied in September. The bar a recommendation had to clear did not move.
That sounds simple. It is the single hardest thing to do in this field, and it is the clearest advantage a model holds over a human. A person who has just lost for a month is under real, visible pressure to do something about it. Change the approach, take a bigger swing, find the bet that wins it all back. Almost all of those instincts are wrong, and almost all of them make the next month worse. A model has no such instincts. It has no emotional state to manage, no loss to feel, no need to be seen to react. It simply keeps applying the same method, which is exactly what the mathematics of the previous section says it should do. After a run of variance, the correct response is to keep the process identical and let more results arrive. The model does that automatically. A human almost never does.
The recovery
November returned 16.1 percent. December returned 8.0 percent. January returned 20.2 percent. And the season went on to its two strongest months of all, March at 36.7 percent and April at 33.6 percent.
The edge was not broken by October. It was interrupted by it, and those are very different things. The only way to tell an interruption from a breakage is to hold the process completely steady and let the sample grow. If the edge is real, the results revert toward it. If the edge was never there, they do not. The 2025-26 season ran that test in real time, without anyone choosing to run it, and the edge reasserted itself clearly across the months that followed. The full month-by-month picture is set out in the 2025-26 season in numbers.
It is worth noting the discipline cut both ways. The model did not panic in October, and it also did not over-extend in March and April when it was winning heavily. The same unchanged criteria produced the poor month and the strong months. That symmetry is the point. A process is only trustworthy if it behaves identically whether it is winning or losing, and the season is the evidence that this one does.
The number that matters more than the headline
If you read only one figure from the whole season, we would rather it was not the 15.3 percent return. We would rather it was the drawdown.
A drawdown is the fall from a previous high point in the running profit to the lowest point that follows it, before a new high is set. It measures the worst stretch, the deepest hole, the point at which someone following the model would have felt the most discomfort. Across the 2025-26 season the largest peak-to-trough drawdown 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.
Those numbers describe what holding the model actually felt like from the inside. A return figure tells you the destination. A drawdown figure tells you the roughest part of the road to get there. A 14.1-unit drawdown is not extreme for a strategy of this kind, but it is real, and a season that returned 15.3 percent still contained a stretch where the running total fell by 14 units and a run of nine recommendations in a row that lost. Anyone shown the return without the drawdown has been shown half the picture, and usually the more flattering half. We publish both, at the same size, for that reason.
What a drawdown means for staking
The drawdown figure is not just a disclosure. It is a practical instruction, and it is worth being explicit about how to read it.
The single most common way to turn a winning method into a losing experience is to stake too large for the drawdown the method carries. If a strategy can fall 14 units, and someone is involved at a stake size where a 14-unit fall is frightening or unaffordable, they will not still be there when the method recovers. They will change their behaviour, or stop, at the worst possible moment, which is precisely the bottom of the trough. The edge was real the whole time. They simply staked themselves out of being able to wait for it.
So the honest use of a drawdown figure is this: size involvement so that the worst stretch in the record, and very probably a worse one than that, since the future is not bound by the past, is something that can be sat through without flinching. A model's edge only reaches the person following it if that person is still following it when the results turn. Stake size, not selection quality, is what most often breaks that chain. This is also why the model is measured to a flat, level stake, a choice we explain in how the model finds value.
Why a losing month is evidence, not a flaw
There is a counterintuitive point here that sits at the centre of how we think about the company, and it is worth stating directly.
A track record with no losing months is not reassuring. It is suspicious.
Genuine edges in efficient markets are small. Small edges produce uneven results. Uneven results include losing stretches. That chain is not a weakness in the argument; it is the argument. A model that has never had a losing month has either not run for very long, or is not being shown to you in full, or is not measuring itself honestly. There is no fourth option, because the mathematics does not allow one. The presence of October in our record, stated as plainly and at the same size as March, is not damage we are managing. It is part of the evidence that the rest of the record is real.
This is why we publish the losing months at all, when it would be easy and tempting not to. A record assembled to show only the good months is not a stronger record. It is a weaker one, because the moment a reader knows a bad month would have been hidden, every good month is worth less. We make that argument in full in tracked, not claimed. October is doing real work in the credibility of the season, and it can only do that work if it is visible.
What a drawdown cannot tell you
There is one thing a drawdown figure does not settle, and honesty requires naming it: a drawdown, on its own, cannot tell you whether you are looking at normal variance or at an edge that has genuinely faded.
From inside a losing run, the two look identical. A model with a real edge having a bad month, and a model whose edge has quietly broken, both produce the same thing on the equity curve: a line going down. The reassurance in the sections above, that October was variance and not breakage, was confirmed by what came after it, the recovery across November to April. But that confirmation only arrived later. In the moment, it could not be read off the drawdown alone.
So the drawdown is not the thing we actually watch to judge whether the model is healthy. We watch the process underneath it. Is the model still finding the same kind of modest, well-evidenced edges it has always found, or has the shape of its opportunities changed? Are its probabilities still landing where they should across a large enough sample? Is the gap it is acting on still the gap that produced the record? Those questions are answered by the underlying behaviour of the model, accumulated over many selections, not by the depth of any single trough.
This is the honest position. A drawdown tells you how uncomfortable a stretch was. It does not, by itself, tell you why. Separating ordinary variance from genuine decay is a question of watching the process over a large sample, and it is one of the standing jobs of model evaluation rather than something a single month ever resolves.
What we took from it
Two things, and neither of them was a change to the model.
The first was confirmation that the discipline holds under stress. The model behaved identically before, during and after the drawdown, and that consistency is the trait the whole business is built on. A model that stays the same when it is losing is a model whose strong months can be believed, because they were produced by the same unchanged process.
The second was a reminder of how to talk about results. October became a permanent line in the record, presented at the same weight as every winning month, because the way a company handles its worst month is a fair preview of how it will handle everything else. A business that is straight about a bad October is more likely to be straight about everything that is harder to see.
A real edge does not mean never losing. It means losing without flinching, keeping the method identical while the sample is still small, and being there, unchanged, when the results turn back toward the edge. October 2025 is the month that let the 2025-26 season prove it could do that.