How Our Predictions Are Made

1. Team Form Analysis

Every prediction starts with recent performance. We analyse each team's last 5–10 matches, looking at goals scored and conceded, shot data, and results. Home and away form are evaluated separately — a team's home record is a poor predictor of their away performances, and mixing the two introduces noise that reduces model accuracy.

These figures feed directly into the attack strength and defensive weakness parameters used by the statistical model in step three. A team averaging 2.1 goals per game at home over their last six outings enters the Poisson calculation differently from one scoring 0.8 — that gap is quantified, not guessed at.

2. Historical & Contextual Data

Single-match form can mislead. A team on a three-game losing streak might have faced three top-table sides; their underlying numbers may be stronger than the results suggest. We factor in head-to-head records, league position, and season-long performance data to distinguish short-term variance from genuine decline.

This contextual layer adjusts the raw form figures before they enter the model. A team whose recent form looks average but who consistently overperform against a particular opponent will see their expected goal rate nudged upward for that specific fixture.

3. Poisson-Based Statistical Model

This is the core of the system. At its heart, the model treats the number of goals each team scores in a match as a Poisson-distributed random variable — a statistical approach with roots in football analytics research going back to Maher (1982) and refined by Dixon & Coles (1997), still widely used in sports modelling today.

The premise is straightforward: given a team's attacking strength (how many goals they typically score against an average defence) and their opponent's defensive weakness (how many goals they typically concede against an average attack), we can calculate the probability of that team scoring 0, 1, 2, or more goals in a specific match. The same calculation runs in parallel for the opponent, giving us a joint probability matrix — every possible scoreline, each with a quantifiable likelihood.

Concretely, if the model estimates Team A has an expected goal rate of 1.8 and Team B has an expected rate of 1.2, the Poisson distribution tells us Team A has roughly a 30% chance of scoring exactly 1 goal, a 27% chance of scoring 2, and a 17% chance of scoring 0. Multiply the two distributions together and you get the probability of every scoreline — 1–1, 2–0, 0–0, and everything else. These probabilities are then summed to produce the match outcome odds: home win, draw, or away win.

The attack and defence parameters themselves are not static. They are recalculated each match day using a rolling window of the most recent data, so the model reflects current form rather than season-long averages that may no longer be relevant after a managerial change, a key injury, or a shift in tactical approach.

4. Match Simulations

A single Poisson calculation gives us the most probable scorelines, but football is inherently noisy. To account for that randomness, we run thousands of simulated match outcomes using the expected goal rates from step three as the foundation — each simulation draws a result consistent with the underlying probabilities rather than simply predicting the most likely single outcome.

Across 10,000 simulations of the same fixture, the model might see Team A winning 48% of the time, a draw 28%, and Team B winning 24%. Those percentages form the raw prediction for the match. This Monte Carlo approach also gives us derived probabilities — the likelihood of both teams scoring, of the total goals exceeding 2.5, and of specific half-time/full-time combinations.

5. Trend & Momentum Analysis

The Poisson model is only as good as its input parameters. To keep those parameters current, we track short and long-term trends — goal-scoring patterns, defensive trajectories, and consistency ratings — on a per-team basis. A team that has tightened up defensively over their last four matches sees their defensive weakness parameter adjusted downward before the next model run.

This is separate from the raw form data in step one. Trend analysis looks at the direction of change, not just the current value. A team might have mediocre form over ten matches but clear upward momentum over the last three, which a purely static model would miss.

6. Confidence Scoring

Every prediction receives a confidence score between 0 and 100, reflecting how strongly all factors agree. The score is derived from the spread of simulation outcomes — matches where the model consistently produces the same result across thousands of runs receive higher confidence scores than those where the results are evenly split.

A score above 75 means the simulation spread heavily favours one outcome: the data is telling a clear story. A score between 50 and 75 suggests a moderate edge but legitimate uncertainty. Below 50 means the model sees the match as genuinely hard to call — the probabilities are too close to justify a strong prediction, and you will see fewer tips from us on those fixtures. High confidence is a reflection of strong agreement in the data, never a guarantee.

A Worked Example

Take a fictional fixture between a strong home side and a mid-table away side. The home team's recent form shows 2.0 expected goals per game at home; the away side concedes 1.4 on the road. The Poisson model estimates a 62% probability of a home win, 22% draw, 16% away win. Over 10,000 simulations, the home win holds up in roughly 6,200 of them — a confidence score of 72. Not a guarantee, but the data points clearly in one direction.

We are building a system to track every published prediction against the actual result, which will eventually feed into a public accuracy page. Until then, each day's tips — with their confidence scores — are available on our Telegram bot and homepage.

Why This Approach

The Poisson framework was chosen because it is well-documented, transparent, and produces interpretable results. Unlike black-box machine learning models that offer no explanation for their outputs, a Poisson-based system lets us trace every prediction back to its inputs — the expected goal rates, the form data, the historical adjustments. If a prediction is wrong, we can usually see why.

That said, no model is perfect. The Poisson assumption treats each team's goal count as independent, which is a simplification — in reality, match events influence each other. Extensions like the bivariate Poisson model or the Dixon-Coles adjustment for low-scoring matches address some of these limitations, and we evaluate these refinements as the dataset grows.