How we predict

Last reviewed: 15 October 2026

Every probability on bet2odds comes from our own prediction engine. This page explains in plain English how a fixture becomes a set of probabilities, how we choose the headline pick, how matches are settled and how we measure whether the model is any good. The figures below are live and update whenever the engine settles new matches. Predictions are estimates, not guarantees.

Live model metrics

Active leagues: 10
Upcoming fixtures: 67
Predictions mirrored: 1,149
Settled predictions: 546

Period	Tips	Hit rate	ROI	Brier	Log loss	CLV
Last 7 days	140–62	69.3%	–	0.384	0.625	–
Last 30 days	140–62	69.3%	–	0.384	0.625	–
Last 90 days	140–62	69.3%	–	0.384	0.625	–
All time	140–62	69.3%	–	0.384	0.625	–

Tips only, all markets. Hit rate = won / (won + lost). ROI assumes a one-unit stake at our fair odds. Lower Brier score and log loss are better. Updated 15 Oct 2026, 11:50 WAT. Full breakdown on our results page.

From data to a pick

Each prediction passes through the same five steps. Results feed team ratings and a goal model; the goal model produces a probability for every possible scoreline; the scoreline grid is added up into markets; and one market is chosen as the headline pick.

What goes into the model

Fixtures, results and league tables reach the engine from licensed football data providers (match data is provided by football-data.org), supplemented by historical results. The website itself never calls a data provider: it reads a local copy of the engine's output. We want to be clear about which information the model actually uses and which we only show you as context.

Direct inputs

Results and goals. Goals scored and conceded in past matches set each team's attack and defence strength. Recent matches count for more than older ones, and early in a season a team's figures are blended with last season's.
Elo ratings. Every result moves both teams' ratings, by more when the result is a surprise or the margin is large.
Home advantage. Both the ratings and the goal model include a home advantage term.
Recent form. The last five matches give a small adjustment to expected goals, capped so one hot streak cannot dominate.
Market prices. When licensed bookmakers have priced a match, we remove their margin and use the result as one input.
League base rates. How often each outcome happens in that league, used when choosing the headline pick.

Shown as context only

Head-to-head records, the league table, separate home and away records, and goal, both-teams-to-score and over 2.5 rates appear on match and team pages to help you read a fixture. They are not separate model inputs: their information is already reflected through the results that build the ratings. Rest days and fixture congestion are not model inputs at present, so take a busy schedule into account yourself.

The probability model

Three components each estimate the match outcome, and the engine blends them into one figure:

Goal model. A Poisson model estimates the expected goals for each team from attack and defence strengths. The Dixon-Coles adjustment corrects the known tendency of simple goal models to underestimate low-scoring draws such as 0-0 and 1-1. This carries the most weight.
Elo. The rating gap between the teams gives a second, independent view of the home, draw and away chances.
Market. De-margined bookmaker prices, averaged across bookmakers, add the information the betting market holds. When no prices are available, the blend uses the other components.

Cup and continental matches are rated mainly from each team's domestic-league strength, scaled by the relative strength of that league; cup results only refine the estimate. Teams with too little data are marked Limited data, given low confidence and never shown as tips, top picks or accumulator legs.

The example below shows how a 1X2 prediction is displayed. Each probability converts to fair odds by dividing 1 by the probability.

From scorelines to markets

The goal model produces a grid with a probability for every scoreline from 0-0 up to 10-10. Every market is read from that same grid, so the markets always agree with each other:

1X2: add up all home-win scores, all draws and all away-win scores.
Both teams to score: every score where both sides have at least one goal.
Over/under 1.5, 2.5 and 3.5: every score whose total goals are above or below the line.
Double chance: two 1X2 outcomes combined, such as home or draw, taken from the final 1X2 figures.
Correct score: the three most likely individual scorelines.

If a licensed bookmaker offers a price above our fair odds, the model sees value; below it, the bookmaker's margin is working against you.

Confidence levels

Each pick carries a confidence band from 1 to 5. It combines how strong the pick is compared with a random guess, how closely the model's components agree, and how much data we hold on both teams. We show it as Low (1-2), Medium (3) or High (4-5). High confidence means the evidence points the same way; it does not mean the pick will win.

The headline pick

The headline pick for a fixture is not simply its most likely outcome. It is the market where our probability sits furthest above that league's normal rate, checked in a set order: match result and double chance first, then both teams to score, over/under 2.5, under 1.5 and correct score. A pick needs a meaningful gap and a minimum probability to qualify; over 1.5 goals is used only as a fallback. That is why a low-scoring league can still show an over 2.5 pick for an open fixture.

Locking and settlement

Predictions refresh as new data arrives until kick-off. At kick-off they are locked and never changed afterwards. Settlement uses the score after 90 minutes plus stoppage time, so extra time and penalties do not count. Matches that are cancelled, abandoned, awarded or postponed by more than three days are void, as are matches whose 90-minute score cannot be confirmed within three days. Every settled pick appears on our track record.

How we measure accuracy

Hit rate: the share of settled tips that won.
ROI: return on a one-unit stake per tip at our fair odds.
Brier score and log loss: how close our probabilities are to what happened across every prediction, not only tips. A model that says 70.0% should be right about seven times in ten.
Closing line value: whether our probabilities beat the final market price, where available.

What the model cannot know

The model works from results and prices. It does not see late injuries or confirmed line-ups, the weather on the day, a team's motivation in a dead rubber or a cup run, a red card in the tenth minute, or plain luck. All of these move results. No model removes that uncertainty, and over time bookmaker margins mean most bettors lose money. Use our numbers as one input, set limits and read our responsible gambling guidance. 18+ only.

Learn more

Frequently asked questions

Are bet2odds predictions written by people?

No. Every probability, confidence rating and headline pick comes from our own statistical model. People build and review the model, but nobody adjusts an individual prediction by hand.

Do predictions change after kick-off?

No. Predictions refresh until kick-off and are then locked permanently, so the track record shows exactly what was published.

Does the model use injuries and line-ups?

Not directly. The model works from results, ratings and market prices, so late team news is reflected only partly, when we have bookmaker prices that moved before kick-off.

What does High confidence mean?

High (bands 4-5) means the components of the model agree and the data is strong. It is still an estimate, and High confidence picks lose too.

How are extra time and penalties settled?

They are not counted. Every pick settles on the score after 90 minutes plus stoppage time.

Questions about this page? Contact: [email protected].