When we said one midweek proves nothing either way — this is what we meant. The Yankee lost. Specifically, it was cooked. This post is the accountability pass that was promised in advance, not a shift in tone about the model’s edge. If we only publish post-mortems when picks land, the whole exercise is fraudulent.
Cooked! The Midweek Yankee Post-Mortem
Follow-up to Monday’s Yankee post. Four picks, one winner, zero combinations settled. And yet — the part of the methodology that actually did a job was the part nobody reads: the filters. Same rule as the weekend Heinz write-up: if the picks land we publish, and if they cook we still publish.
THE SHORT VERSION
Yankee = 1 of 4 picks landed. All 11 combinations lost. Patent subset = 1 single returned, doubles and treble lost.
But of the six fixtures the process rejected — dropped from the final set or skipped for failing a threshold — every single one resolved the way the model or the filter said it would. The picks missed. The filter didn’t.
🔥 The Yankee — Four Picks, One Winner
Kickoff odds from Bet365 as recorded at publication. ✅ = won, ❌ = lost.
YANKEE ROI
−100%
11 bets, 0 returned
PATENT ROI
−53%
7 bets, 3.30 back on 7
FILTERS
5 / 5
all drops and skips vindicated
✅ The Part That Actually Worked — The Filters
Monday’s post listed five fixtures that made it through the funnel but were rejected at a later stage — two dropped because the model’s top call disagreed with the high-EV outcome, two Championship fixtures skipped for failing EV or probability thresholds, one near-miss that fell just below the EV bar. Here’s how those rejections resolved:
Five fixtures rejected at some stage of the funnel. Five resolutions that matched what the rejection logic implied. Nothing from the discard pile would have rescued the Yankee either.
⚠️ The Lesson Inside the Losses
BOLTON 5–1 STEVENAGE — THE ONLY NON-MODEL-ALIGNED PICK, THE WORST BLOW-OUT
Monday’s post was explicit: three of the four Yankee picks were the model’s top call (the Patent subset), and one was a “value” pick where the model’s top call was different but the EV was strong. That fourth one was Stevenage away at Bolton. The model’s primary prediction was Bolton home at 40.8%. Stevenage were the #2 at 36.4%. We took the #2 because the EV was 1.54.
Bolton won 5–1. The model’s top call was correct. The Yankee selection was the losing side of the same fixture.
This is one data point — not evidence that the “value-pick exception” is broken. But it is the loudest argument in the dataset for tightening the rule: a Patent (three selections, all model-aligned) is a cleaner structure than a Yankee that stretches to include a #2-outcome pick. Worth re-examining the filter at the next review.
THREE DRAWS PICKED. ONE DRAW LANDED.
The Yankee leaned heavily into draws — Luton, Orient, and the dropped Wigan pick would have made four. Of the three that survived the filter, only Orient–Mansfield actually drew. Luton beat Northampton 2–1; Huddersfield–Cardiff finished 1–1 (the draw landed, but our pick was home, so it still lost).
Across all nine played midweek fixtures, three were draws — roughly the long-run base rate. The model found the right number of draws in aggregate; it just put them in the wrong fixtures. That’s the thing models are worst at. It is why the draw EV gets so big in the first place — the market isn’t mispricing draws out of ignorance, it is pricing them for a game where one side usually breaks the stalemate. Longer read on the draw problem here.
📊 What This Changes — and What It Doesn’t
What it doesn’t change:
- The edge claim. One midweek of four picks is too small to move any honest estimate of the model’s long-run performance. This was said upfront, and it is still true after the fact.
- The filter logic. Every rejection this week was correct. If anything, the evidence is that the filters are doing more work than the picks are.
What it flags for the next review:
- The “value-pick exception” — picks where the model’s top call and the selection disagree. Our one example this week produced the biggest single loss. Worth examining whether Patent-only (strict model-aligned) outperforms Yankee-with-a-value-pick across a wider sample.
- Draw concentration. Three of four Yankee picks being draws creates correlated exposure — if draws are under-represented that midweek (as they were), the Yankee structurally cannot survive. Possibly worth a rule limiting draw selections to N-of-4.
THE HONEST FRAMING
The model doesn’t know it just had a bad midweek. It has no feelings about Wigan. It will produce the same kind of probability estimates on Friday’s weekend slate as it did on Monday’s. Whether those estimates have long-run edge will be answered by hundreds of picks, not four.
What we can say after this week: the picks missed, the filter worked, and the only non-model-aligned selection was the ugliest of the four losses. That’s a dataset of one, but it’s a dataset of one worth remembering.
Accountability post for the midweek Yankee. Next check-in: weekend Heinz results, Monday.
