Our worst week, and what broke in it

Nine assessments, seven losses. At the prices involved that outcome has a probability of a few percent, which means it happens roughly once every couple of years and this was the once.

#Eight of the nine were fine

By fine we mean the reasoning holds up on re-reading, the sources were real, and the number would be the same if we produced it again today knowing only what we knew then. Seven of those eight lost. That is what variance does and there is nothing to fix in it.

Six of the eight were still taken at a better price than the market closed at, which is the reason we do not treat the week as evidence of anything beyond a bad run.

#One was not fine

A team sheet was published two hours before kick-off and then amended. The research pass read the first version. Nothing downstream checked the timestamp, so an assessment went out built on a lineup that had already been superseded, and the number was materially wrong rather than unluckily wrong.

That is a process failure and it has a process fix: every claim now carries the time it was published as well as the source, and a claim older than the most recent update from the same origin is refetched rather than trusted.

#Why the distinction is the whole point

If bad weeks and broken processes are counted the same way, then every bad week produces a change and the method drifts toward whatever the last few results suggested. That is the fastest way to destroy an edge that was real.

Separating them requires publishing both, which requires publishing the weeks you would rather not.

#What this does not tell you

Whether the timestamp fix is sufficient. It addresses the failure we found. It does not tell us how many similar failures happened before and settled in our favour, which by construction we would never have looked for.