The BTP model is a machine. It does not have feelings, a Bet365 account, or opinions about Wigan. It produces probability estimates. This post is data analysis — it is not a tipster service, a recommendations column, or a suggestion that you do anything at all with any of this. The model has never watched a match in its life. It doesn’t even have a life. It has features.
A note before you scroll: this isn’t the standard weekend Heinz. It comes from the midweek catch-up window — ten rescheduled League One and Championship fixtures crammed into Tuesday and Wednesday. The weekend GW44 round gets its own Heinz on Friday. This is a separate, midweek-only set, and for reasons explained below it takes a Yankee shape rather than a Heinz — four selections instead of six.
⚡ THIS WEEK’S HEADLINE
Four selections — three of them are the model’s top call, three of them are draws, and one is the sort of value the market tends to miss.
The standout is Luton v Northampton at EV 2.28 — the biggest model-vs-market gap we have recorded this season. Inside the Yankee, the three selections where the model’s primary prediction IS the selection also form a Patent (three-selection structure, seven combinations). That subset carries the strongest signal.
We ran every one of the ten midweek fixtures through the same EV pipeline as our weekend Heinz: model probability divided by Bet365’s normalised implied probability, ranked by ratio. Six fixtures technically cleared the bar (EV ≥ 1.10, model probability ≥ 25%). On review we stripped that back to four. The two dropped — Wigan–Rotherham and AFC Wimbledon–Stockport — passed the raw filters, but only because Bet365 underpriced the draw relative to our model’s draw estimate. In both fixtures the model’s primary prediction is a different outcome (Wigan to win; Stockport away). Keeping them in a Heinz would have meant backing outcomes the model itself rates as second or third most likely. Better to be honest about that.
The idea is the same as the weekend version. We are looking for fixtures where Bet365’s implied probability is significantly lower than what our model estimates. When the gap is wide enough, that is where the model’s interpretation of recent form diverges most sharply from the market price. That gap is the expected value. All four selections are from League One — both Championship fixtures in this window failed the filters (details below).
Championship predictions use our Platt-calibrated Logistic Regression (goals_logreg_v1_cal). League One uses the original Random Forest (leagueone_goals_v1). Why different models? Read this.
HOW EV WORKS — PLAIN ENGLISH
Bookmakers express their probability as implied odds. Divide our model’s probability by that implied probability and you get the EV ratio. EV of 1.0 = the model and the market agree exactly. EV of 2.28 = the model thinks this outcome is more than twice as likely as the market does.
EV doesn’t guarantee a correct call on any single fixture — the bookmaker might be right and we might be wrong. What it measures is where our model’s estimate diverges most from the market price. Over a large enough sample, positive-EV selections should return more BTP tokens than they cost. One midweek’s results prove nothing either way. We only know if the model has edge by tracking it across hundreds of picks. We are not there yet.
Why a Yankee — and What a Patent Tells Us
The weekend Heinz takes six selections from a pool of around 22 qualifying fixtures. This midweek round only offered nine fixtures with Bet365 odds in the database. Six of those cleared our raw EV filter — but when we examined each pick against the model’s own primary prediction, only four stood up as genuine model-backed value calls. Two of the six were draws priced as value purely because Bet365 underpriced the draw; the model itself was calling a decisive result in both fixtures. We have dropped those two.
That leaves four selections, which fits a Yankee structure: 11 combinations (6 doubles, 4 trebles, 1 four-fold). No singles — the Yankee requires at least two selections to return anything. Inside the four, three of them align directly with the model’s top prediction. Those three sit as a Patent (7 combinations: 3 singles, 3 doubles, 1 treble). The Patent is the tightest, most model-aligned core of this set. The Yankee adds a fourth value pick — Stevenage at Bolton — where the model’s primary prediction is actually Bolton at home, but the market’s 4.00 on Stevenage is wide enough that the EV still clears the bar.
Both structures are shown below. The Patent is the strongest signal; the Yankee is the broader value view. Neither is any kind of suggestion to stake anything. This is data analysis. Where you take it from there is none of the model’s business — but if you do, BeGambleAware is there and 18+ rules apply.
The Full EV Landscape — All 10 Midweek Fixtures
All ten midweek catch-up fixtures sorted by best EV outcome. ★ = Yankee selection. ◆ = Patent subset (the three model-aligned picks inside the Yankee).
◆ = Patent subset (three model-aligned picks). ★ = Yankee selection (all four picks). Faded rows = dropped, failed filter, or no odds available. B365 implied = normalised Bet365 probability (overround removed).
WHY ZERO CHAMPIONSHIP PICKS THIS WEEK
Only two Championship fixtures land in this midweek window. Portsmouth v Ipswich’s best EV was 1.05 on the home win — below the 1.10 minimum, with the market and model essentially in agreement. Southampton v Blackburn’s only above-1.0 EV was 1.02 on the away win, but the model gave Blackburn just 17.4% — well below the 25% probability floor.
When the model and the market agree, there is nothing left in the gap. When they disagree on a heavy outsider, the EV looks tempting but the underlying probability is too low to be interesting. Both Championship fixtures fell into one of those two camps. Hence the all-League-One Yankee this week — that’s what the data returned, not an editorial preference.
★ The Yankee — Four Where the Model Diverges From the Market
Four selections from nine priced midweek fixtures. Three are the model’s top call (and form the Patent subset). One is a broader value pick where the model’s top call is different but the EV is strong.
LEAGUE ONE · WED 15 APR
Luton v Northampton
SELECTED: DRAW
✓ Model’s top call
EV RATIO
2.28
odds: 5.00
42.4%
18.6%
The standout. Bet365 has Luton at 1.33 — strong favourites against bottom-of-the-table Northampton — implying just 18.6% on the draw. The model gives the draw 42.4% and rates it as the most likely single outcome. The market is pricing a Luton stroll. The model is pricing a grind.
LEAGUE ONE · TUE 14 APR
Bolton v Stevenage
SELECTED: STEVENAGE AWAY
◇ Value pick — model’s #2 outcome
EV RATIO
1.54
odds: 4.00
36.4%
23.6%
The one genuine value pick outside the Patent. The model’s top call here is actually Bolton at home (40.8%), but only by a sliver. Stevenage at 36.4% is effectively co-favourite in the model’s view, while the market prices them at 4.00 — a 23.6% implied. That gap is why this survives the EV bar even though it is the model’s #2 outcome rather than its #1.
LEAGUE ONE · TUE 14 APR
Leyton Orient v Mansfield
SELECTED: DRAW
✓ Model’s top call
EV RATIO
1.54
odds: 3.30
43.0%
27.2%
Highest single-outcome model probability in the whole Yankee — the model gives the draw 43.0%, more than home and away wins combined. Bet365 has the draw at 3.30 (27.2% implied). The model sees this fixture as a pick’em with neither side projecting daylight on the rolling form.
LEAGUE ONE · TUE 14 APR
Huddersfield v Cardiff
SELECTED: HUDDERSFIELD HOME
✓ Model’s top call
EV RATIO
1.22
odds: 3.00
37.8%
31.0%
The only home win in the Yankee, and the tightest edge of the four. Cardiff are travelling as away favourites at 2.15, but Huddersfield’s recent rolling form rates higher than their league position implies. The model gives Huddersfield 37.8%; Bet365 implies 31.0%. A modest edge on a fixture where the market is leaning the other way.
★ The Yankee — All Four Selections
A Yankee consists of 11 combinations from four selections: six doubles, four trebles, and one four-fold. No singles — the Yankee needs at least two selections to return anything. The product of individual EV ratios gives the four-fold EV: 2.28 × 1.54 × 1.54 × 1.22 = 6.6× what the market implies.
The combined model probability of all four landing is 2.51%. That sounds low. The market prices it as 0.505% (implied by combined odds of 198×). The gap between those two numbers is what the EV analysis is actually saying.
◆ The Patent — The Three Model-Aligned Picks
Strip the Yankee back to only the picks where the model’s primary prediction IS the selection and you are left with a Patent: seven combinations from three selections — three singles, three doubles, and one treble. The key feature: unlike the Yankee, the Patent includes singles, so any one correct result returns something. The Patent is the tightest subset inside the Yankee — the three picks where the model’s top call and the chosen outcome line up exactly.
The Patent carries the strongest signal of anything in this post. All three picks are outcomes the model rates as the single most likely result in that fixture. The broader Yankee adds one more value pick — Stevenage away — where the model’s primary prediction is Bolton at home, but the market’s 4.00 on Stevenage is wide enough that the EV still clears. The Patent is cleaner; the Yankee is wider.
EV Ratio — The Four Selections Visualised
Bar length = EV ratio. The dotted line marks EV = 1.0 (break-even — model and market agree). Anything to the right of the line represents model disagreement with the market price.
1.00
2.28
1.54
1.54
1.22
◆ = Patent subset. Bar scale: EV 1.0 = 38% width. Green = EV ≥ 1.60. Teal = 1.30–1.59. Amber = 1.20–1.29. Red line = break-even.
Season Accountability — The Running Total
Last weekend’s six Heinz picks all lost. We wrote about it openly: zero from six on the headline picks, four from four on the EV ≥ 1.20 picks that ranked just below the cut-off. No new picks have settled since that post, so the running total below is unchanged. The midweek selections in this post will be added to the tally once results are in, and reported in next Monday’s accountability update.
A 28.6% correct call rate sounds low until you factor in the average odds: 4.76. Four correct calls from fourteen, at those odds, produces a positive token return — which is what +35.9% reflects. That number is a measure of model calibration against market price, not a financial result. By league:
Championship picks are averaging odds of 6.27 — long-shot territory — and two correct calls produce +108.9% on a small sample. League One sits at −9.4% over eight picks at shorter average odds of 3.62. Both are far too small a sample to draw conclusions from. We will update this table after the midweek fixtures complete.
THE HONEST CAVEAT
Fourteen picks is not a sample size. It is an anecdote with a spreadsheet attached. A good run or a bad run across 14 data points tells you almost nothing about whether the model has genuine edge over the market. Adding four midweek picks to the pool — three of them draws — will not change that.
We need hundreds of tracked picks before the token ROI figure becomes meaningful. We are publishing these results every gameweek because building that record — in public, without editing — is the only honest way to find out. We will keep going.
The BTP model was built to produce honest probability estimates and write interesting data posts — not to tell anyone what to do on a Tuesday evening. It has identified four fixtures where its probability estimate diverges from Bet365’s market price. It finds that analytically interesting and makes no promises about it. The model has no financial interest in any outcome. It simply disagrees with the market on four occasions this midweek, and it stands by every decimal point right up until kick-off, at which point it becomes everyone else’s problem. If the decimal points in this post look like something to act on — think twice, set limits, and visit BeGambleAware.
Data notes: Championship predictions use the Platt-calibrated LR model (goals_logreg_v1_cal). League One uses the original Random Forest (leagueone_goals_v1). Bet365 decimal odds fetched 14 April 2026 via API-Football. Implied probabilities normalised to remove overround. EV = model probability ÷ normalised implied probability. Peterborough v Port Vale (Thursday 16 Apr) excluded — Bet365 had not priced the fixture at time of publication. Six fixtures cleared raw EV + probability filters (EV ≥ 1.10, model ≥ 25%); two were dropped on model-alignment review, leaving four in the final Yankee.
Full model methodology here.
Not betting advice. BTP tokens are an imaginary unit used solely to track model performance over time. No real money is implied or recommended. Football probability models are for analytical and entertainment purposes only. If you do bet, please do so responsibly, stick to limits you set in advance, and remember that probabilities describe the long run — not the next match. 18+ · BeGambleAware.org · Gambling Commission.