The month in full, 41 assessments settled
Every published assessment from the last month, settled, including the fourteen that lost and the two where our reasoning was plainly wrong.
Forty-one assessments were published last month. Twenty-seven won, fourteen lost, and the month finished ahead. None of that is the interesting part.
The number that carries information
Thirty-two of the forty-one were taken at a price better than the one the market settled on. That is the figure worth watching, for the reasons set out in why the closing line is the only honest scoreboard: it is measured on every bet immediately and it does not care whether the ball went in.
Nine closed against us. Six of those nine were fixtures where the research half found nothing and the number stayed close to the model, which is exactly where we would expect to be beaten by a market that had found something.
Two where the reasoning was wrong
Distinguishing these from ordinary losses matters, because they are the only ones that should change anything.
The first was a totals market where the model was reading a scoring rate from a run of fixtures against three promoted sides. That is a sample problem we should have caught and the adjustment is now in.
The second was a research failure of the kind described in why research runs before the model: three outlets carried the same injury claim, it was counted as three sources, and it originated from one. The deduplication was too loose and has been tightened.
What a good month is worth
Very little on its own. At the sample sizes an individual month provides, a result like this one is comfortably within the range a method with no edge would produce, and reading it as vindication would be exactly the error this piece exists to avoid.
What this does not tell you
Whether the two corrections above make the model better. They make it more correct on the two cases that produced them, which is a different and much weaker claim, and the honest test is whether they hold up out of sample.
Common questions
How many assessments do you publish a month?
It varies with the calendar. Last month it was forty-one, and all forty-one are settled here whether they won or lost.
Is a profitable month evidence the method works?
No. A month is a small sample, and a profitable one is roughly as likely under a small real edge as under no edge at all.
What do you do when the reasoning was wrong rather than unlucky?
Write it down separately. A losing bet with sound reasoning and a losing bet with a mistake in it are different events and should not be counted the same way.