Most football prediction tools give you a number and ask you to trust it. BetSignals works differently. Every signal you see has been produced by two separate models working independently, and the output of both is shown to you, every time. Nothing is smoothed over or hidden.
This guide explains what the models do, how they arrive at a signal, and how to read the output correctly.
---
Why two models?
A single model, however well-built, reflects the assumptions and blind spots of its designer. If those assumptions are wrong, or the model overfits to a particular type of match, the errors compound silently and you never know.
BetSignals runs two independent models on each fixture. The first is a Negative Binomial model that weights goals, shots, and defensive exposure. The second, referred to internally as V12, is also Negative Binomial but uses decay-weighted form (recent matches count more), league-specific dispersion and home advantage parameters, a low-scoring game correction, and a weaker-side lift to account for the tendency of underdogs to score more than raw averages suggest.
When both models arrive at the same conclusion via different routes, that agreement carries more weight than either could alone. When they disagree, that disagreement is shown to you directly.
---
What the models look at
The models draw on recent team performance data for each fixture. The core inputs are:
- Goals scored and conceded across the last eight matches for each team
- Shots for and against, used as a proxy for attacking intent and defensive exposure
- Recent form, expressed as a win/draw/loss ratio across those matches
- Home advantage, calibrated at a league level rather than applied as a flat value across all competitions
- Head-to-head history, which is fetched and shown alongside each signal as context, though it does not feed directly into the probability calculation
Neither model is a simple goals-per-game formula. Both are probabilistic: they generate a full score distribution, then derive market probabilities from that distribution rather than picking a predicted result.
The models do not use player availability, injury lists, or team news. Signals are based on team-level performance data only.
For player-level signals, the models incorporate SignalRates, which adjusts a player's raw stats for the opposition's tendency to concede that specific event type. A player's shots-per-game figure means more against a side that routinely allows attempts than against one that defends deep and restricts volume.
---
How a signal is produced
For each match and each market, both models produce an independent probability estimate.
Those estimates are then compared against the available bookmaker odds to determine whether a value bet exists. If the model probability is meaningfully higher than the implied probability in the odds, that market is a candidate for a signal.
The key step is the cross-check. If both models agree that the probability exceeds the bookmaker's implied odds, a signal is issued. If they disagree, a red X is shown instead.
---
Reading the output: star ratings and red X
The signal rating reflects how strongly the two models agree and how large the identified edge is.
- ★★★ Both models agree on the dominant outcome, and the model probability for that outcome exceeds 50%
- ★★ Both models agree on the outcome, with a model probability between 40% and 50%
- ★ Both models agree on the outcome, but the model probability is below 40%
- Red X The models disagree. No signal is issued. The outputs are shown so you can make your own judgement.
A ★★★ signal is not a guarantee. It means the data alignment is strong. Football always contains variance that no model can fully account for. What star ratings help you do is prioritise: which markets have the most consistent data behind them and which are borderline.
For a deeper breakdown of how to use ratings in practice, see Reading Model Confidence Levels.
---
What the model does not do
It does not predict scorelines. The models produce probabilities for specific markets, not a specific match outcome. The distinction matters. A model might show high confidence in BTTS landing in a fixture while being far less certain about the correct score.
It does not account for in-play events. Signals are generated pre-match, based on data available before kick-off. Red cards or weather conditions on the day are outside the model's scope.
It does not use player availability or team news. The models are built on team-level performance data. Injuries, suspensions, and late lineup changes are not factored in. If a key player is ruled out close to kick-off, the signal will not reflect that. This is worth bearing in mind when applying signals to your own decisions.
It is not infallible. Any model built on historical data carries the assumption that historical patterns are somewhat predictive of future results. That assumption holds well on average, but individual matches can and do deviate from it. The goal is to find expected value over a large sample, not to guarantee a return on every bet.
---
How the models are validated
Before any model change goes live, BetSignals runs backtesting against historical fixtures. This measures whether the model's probability estimates are well-calibrated: does a market the model rates at 70% probability actually land around 70% of the time? If not, the model is adjusted.
This process is called model calibration. It is the difference between a model that generates plausible-sounding numbers and one whose numbers are actually useful for betting decisions.
---
Using the model output well
The model gives you an edge in identifying where bookmaker odds may be mispriced. That edge is statistical and plays out over many bets, not on any single selection.
To get the most from it:
- Use signal ratings to guide which markets to prioritise, not to decide which to bet on blindly
- Take red X markets seriously as information: a split between models often reflects genuine uncertainty in the fixture
- Combine model output with your own reading of the match where you have context the model does not
- Manage your stake sizes consistently regardless of confidence level, and read the bankroll management guide if you have not already
The model is a tool. It improves the quality of your decisions. It does not make them for you.
---
Next reads
- Understanding Signal Ratings: the full breakdown of what ★, ★★, and ★★★ mean in practice
- Reading Model Confidence Levels: how to interpret the confidence bracket on each signal
- How SignalRates Works: the player stat methodology that feeds into the model
- What is a Value Bet?: why probability vs odds is the core of any data-led approach
---
18+ | Gambling should be enjoyable. If it stops being fun, take a break. For support and advice visit BeGambleAware.org. See our full responsible gambling guide.