Every BTP match preview includes a section called “The Model’s View”. It shows numbers like “49% home / 22% draw / 29% away” and a predicted scoreline like “1-0”. This page explains what’s actually behind those numbers in plain English — no maths required.

What the model does

For every Championship and League One match, our model predicts two things:

• The probability of each outcome — how likely it is that the home team wins, the away team wins, or the game ends in a draw. These three numbers add up to 100%.
• The most likely scoreline — a specific final score (e.g. “Stockport 1-0 Stevenage”) that the model thinks is the single most probable result, given that the favoured side wins.

That’s it. The model doesn’t tell you what to do with the numbers — it just gives you a quantitative read on the match, against which our editorial commentary (the “Editor’s Take” section in each post) sometimes agrees and sometimes disagrees.

How to read the probability numbers

The most important number is the home team’s win probability. We use a simple plain-English label for each band so you don’t have to think in percentages:

What’s actually inside the model

If you don’t care about the methodology, skip this section — the headline numbers are all that matter for reading the previews. If you do, here’s the plain-English version of the technical bits.

1. The 1X2 model (who wins?)

To predict the probability of each outcome, we use a technique called logistic regression — a standard statistics method that looks at all the matches that have already been played this season and learns what factors correlate with home wins, draws and away wins. The inputs include:

• Recent form for both teams
• Goals scored and conceded
• Underlying chance creation (expected goals — see below)
• Home/away splits
• Key player availability (when we have it)

The raw output is then run through a second step called Platt calibration — a tweak that adjusts the probabilities so they actually match how often things happen in real games. Without calibration, models tend to be over-confident. Calibration brings the numbers back to honest.

2. The goals model (what’s the score?)

To predict scorelines, we use a separate model based on something called the Poisson distribution with a tweak known as the Dixon-Coles adjustment. In plain English:

• The Poisson part predicts how many goals each side is likely to score on average — what statisticians call “scoring rates” (we use the Greek letter λ, lambda, for these — when you see “λ home 1.55” in a preview, that just means “the home team is expected to score about 1.55 goals on average”).
• The Dixon-Coles part is a small adjustment that makes the model better at low-scoring games like 0-0 and 1-1, which a plain Poisson model under-predicts.

3. The “most likely scoreline” (1-0 / 2-1 / etc.)

When you see a predicted scoreline like “1-0 Stockport”, it’s worked out like this:

1. The 1X2 model says who is the most likely winner — say, Stockport.
2. The goals model gives a probability for every possible scoreline (1-0, 1-1, 2-0, 2-1, etc.).
3. We then pick the single most likely scoreline among the scores where Stockport win. That’s the “conditional modal scoreline” — “conditional” because it’s restricted to the favoured side’s wins, “modal” because it’s the most common.

Why do it this way? Because if you just take the most likely scoreline overall, you often get 1-1 even when the model strongly favours one side to win — which feels self-contradictory to a reader. The conditional version keeps the scoreline consistent with the headline probability.

What the model doesn’t do

• Extra time and penalties. Our probabilities cover the regulation 90 minutes only. If a play-off second leg goes to extra time, no public model prices that — the tie’s resolution becomes a coin-toss as far as the model is concerned.
• Mid-game red cards / serious injuries. The pre-match probability assumes both teams complete the game with their starting line-up.
• Refereeing variance. Penalties given/not given, sending-offs, VAR decisions — the model treats these as random.
• Things a human sees. Off-pitch drama, managerial sackings, transfer rumours mid-match, tactical surprises — the model can’t read the news.
• Mid-season transfers. A player’s cumulative season stats stay in the data even after they leave; we manually check this for each preview but it’s a known fragility.

This is why every preview also has an “Editor’s Take” section. The model gives you the math; the editorial gives you the things the math can’t see. Sometimes they agree (like the recent Saints-Boro semi-final). Sometimes they don’t (like Hull-Millwall the night before, where the editorial backed Hull at 24% and Hull won 2-0).

What we will and won’t claim

BTP is an analytical site, not a tipping service. We publish probability predictions because we find them interesting and because they give a frame against which to discuss matches. We do not give betting advice; we have retired all “EV / value / contrarian pick” framing across the site after a backtest in April 2026 showed no profitable variant. Probability numbers are presented for context, not for action.

One sentence to take away

When you see “Clear favourite at 55% / draw 19% / underdog 26%” in a BTP preview, you can read it as: “The model thinks Stockport are the most likely winner by a meaningful margin, with the draw being the least likely outcome.” Everything else in the model section is supporting evidence for that headline.

Model versions in production at the time of writing: goals_logreg_v1_cal (Championship 1X2), goals_poisson_v1 (Championship goals), leagueone_lr_v1_cal (League One 1X2), leagueone_poisson_v1 (League One goals), wsl_lr_v1_cal (WSL 1X2). All probabilities cover regulation 90 minutes; extra time and penalties are not modelled.