Behind the picks

Not a technical document. Just what actually happens before a pick shows up, and what happens after it turns out to be wrong.

How a pick gets made

Every pick starts as a number built from actual results. Who beat whom, by how much, all season. Not from opinions about form, not from anyone's read on which team looks hungrier.

Then the same match is assessed a second time, from scratch. The two assessments are deliberately kept apart, so neither can see the other's answer and quietly talk itself into agreeing. If they disagree sharply, that is information, and usually it means we skip the match.

Before anything is published, there is one more check: late news. Lineups, injuries, a rested starter, a suspension. A pick built on yesterday's information is not a pick, it is a guess with better presentation. If the news makes it stale, it gets dropped.

The stake and the edge are worked out separately from all of that, in plain code. Nothing gets to grade its own homework on how confident it should sound.

Most matches produce nothing

This is the part people find strange. On a normal day we look at twenty or more matches and publish three. Sometimes none.

Picks get thrown out for boring reasons, and the boring reasons are the important ones:

• Not enough history. Early season, a new team, a national side that has played four matches together. The numbers exist but they do not mean much yet.
• We disagree with the market too much. If our number is wildly different from the price everyone else is offering, the boring explanation is usually right: we are wrong, not them. There is a hard limit on how far we are allowed to disagree before the pick is binned.
• The edge is too thin. Below a set threshold it is noise, not an opportunity.
• Late news killed it. See above.

"No bet" is a real answer here, not a failure to find one. A service that finds a great opportunity every single day is not being careful, it is being generous with your money.

We check our own work, and it stings

Everything gets logged. Every pick, timestamped before kickoff, settled automatically against the final score, kept permanently whether it won or lost. That log is public.

But the count of wins is the easy part. The harder question is whether our confidence means anything. So we compare what we claimed against what happened.

That is not a comfortable sentence to put on your own website. It is also the single most useful thing we know about ourselves, and we would rather publish it than quietly hope nobody measures.

What we did about it

Checking your work only counts if you act on the answer. Three times now, the answer was to remove something.

In July we switched off baseball's biggest league. It was our highest volume sport, and our own record said it was losing money. Not a redesign, not a "we will monitor it". Off. You can see the exact date it stops appearing in the record.

In September we switched off football over and under 2.5 goals. Same story, worse detail: out of forty two of those picks, forty one had been on the same side of the line. A model that only ever finds value in one direction is not finding value, it has a lean. We had already tried the gentle fix months earlier and it had not worked, so the market is paused until the underlying approach is rebuilt.

And we tightened the disagreement rule. Looking back over every settled pick, the ones where we disagreed most violently with the market were our worst performers by a wide margin. The threshold came down. Fewer picks, better ones.

What you will not find here

• Guarantees, or any claim about what next month will look like.
• Deleted losses. Nothing comes down after the fact.
• Screenshots of a good week presented as if it were the whole story.
• Bookmaker referral links inside a pick. Here is how this is funded instead.

That is really the whole idea. Publish it, measure it, and believe the numbers over the story we would like to tell about ourselves.