A rough week, and why the 5/2+ picks got a haircut

Eight winners from 25 this week. 32% win rate, which is exactly in line with where the model's been running all year, but the losses this time were expensive enough that flat-stake P&L came out at -£6.74, or -27% ROI for the week.

All-time it's now 193 picks settled, 32% win rate, -8% ROI. So this week didn't wreck the average, it just sat on the wrong side of it. I don't have a tidy explanation for why this particular seven days went worse than the baseline — the picks weren't obviously worse quality on paper, they just didn't hit. That's horse racing. I'd rather say that plainly than invent a reason.

What I did spend the week on is something that's been bugging me for a while: the confidence number on longer-priced picks was lying, or at least not telling the truth clearly enough.

Splitting the edge calculation by price

The AI-estimate percentage is supposed to measure real edge over the market price. Up to now it used one flat conviction number per tier regardless of whether the pick was 6/4 or 9/1. When I ran the calibration numbers properly this week, it was obvious that number was being propped up by short-priced winners. Picks at 5/2 or bigger in the same conviction tier were winning at barely a third of what the tier implied — 2 from 16. That's not a rounding error, that's a tier that doesn't mean what it says for half the price range it covers.

So from 24 July, the edge is measured separately above and below 5/2. The honest result is that some longer-priced picks now show a lower number than before, and a few of them no longer clear the bar to be a Value Pick or get real auto-bet money. That's going to shrink the pool of picks at bigger prices for a while. I think that's the right trade — a Value Pick badge that's wrong on 5/2+ shots isn't worth having, and this doesn't touch anything published before the 24th, so the historical record stays as it was.

In the same batch of changes I also tightened up what counts as a Value Pick at all: a strategic review found picks that only scraped through on a bare majority of the model's self-consistency check were running at -31% ROI (n=62) against +35% for picks with full agreement (n=66). Bare-majority picks still publish and settle as ordinary tips, they just don't count as Value Picks anymore — no real money, not a Top Pick, not in the CLV numbers.

Also shipped: void races now show a proper "Void · stake back" badge instead of quietly vanishing off the card, free bets can be logged in My Bets with correct £0-outlay ROI math, and pick cards now show a live "now +X%" value readout instead of freezing at the morning price.

Next week I want to see what the 5/2+ pool looks like once the new edge calculation has had a full week to run, and whether the full-agreement-only rule changes the shape of the picks that get published day to day. I don't know yet if it'll be a smaller list or just a different one.