6 Winners, 13 Losers, and Still Green for the Week

19 value picks settled between 11 and 18 July. 6 winners, 0 places, 13 losers. That's a 32% win rate, which is actually bang in line with where the model's sat for a while now, all-time it's 32% across 168 picks too. What's different is the money: flat 1pt stakes came out at +£2.13 for the week, which works out to +11% ROI. Small number, real number, I'll take it.

The biggest win was The Balearic Sun at Epsom on 16 July, 4/1, +£4.00. That one basically carried the week, take it out and the other 18 picks are roughly break-even. The worst loss was The Lost Sock at Salisbury on 11 July, 5/2, £1.00, which is just a normal losing bet at normal odds, nothing dramatic there.

All-time the picture is still -7% ROI over 168 picks. One decent week doesn't undo that, and I'm not going to pretend it does. What I do notice is the win rate holding steady at 32% both this week and across the full sample, if that number stays consistent while the odds on the winners drift up, the ROI line should start climbing out of the hole. That's the theory, anyway. It's one data point.

The bigger job this week was the audit

I spent a chunk of the week going through settled data line by line and closed about twenty bugs across settlement, stats and the results tweet. Non-runners were sometimes being settled as losses instead of voids, including through the cron job. Festival meetings were resolving to a generic key instead of their actual course. Win-only picks that placed but didn't win were getting counted as "place" results when they shouldn't have been. Jockey and trainer ROI was using the wrong stake basis. None of this changes the headline P&L numbers above, those were already right, but it means some of the secondary stats on the site (jockey/trainer records, the results tweet, a few edge cases in settlement) weren't as clean as they should've been until now. I'd rather say that plainly than pretend everything was always correct.

On the product side, the two things I actually think matter: I backfilled real Closing Line Value onto every already-settled pick, not just what's settled since CLV tracking went in, so the CLV record now covers the site's full history. And the homepage now leads with each course's actual historical draw bias,computed from 200 days of settled results, never model output, rather than leading with a pitch. There's also a new /tips/findings page listing every course racing today sorted by how meaningful its draw bias gap is, and a Goodwood Festival guide is up ahead of the 28 July start, since that's the largest draw bias gap I've measured on any course so far.

I also shipped a Yankee option on top of the existing double, a free-standing matched betting calculator, and an evening edition that publishes provisional tips for tomorrow's card the night before, that last one's manual-dispatch only for now while I watch whether the provisional picks actually match what the morning run settles on.

Next week I want to see if the win rate holds at 32% with a full second week of clean settlement data behind it, and whether the CLV backfill lines up with what the ROI has actually done.