How Our Predictions Are Built

A transparent walk-through of the data, the model, and the human review that produces every pick we publish.

Why We Wrote This Page

Most tipster sites are a black box: they show you a pick, claim a track record, and ask you to trust them. We think readers deserve more than that. This page lays out exactly what data goes into a SafeBet Football prediction, what the model does with it, and where a human analyst is allowed to override it. If you understand how a recommendation is produced, you can decide for yourself whether to follow it.

Step 1: Data Collection

Every morning, our infrastructure pulls fresh data for every match in our coverage. The core inputs are:

  • Match results and scorelines — the last 5 seasons for every team in the model.
  • Shot-level data — every shot on goal, with location, body part, and assist type. This is what underpins our expected goals (xG) calculations.
  • Match context — home/away venue, kick-off time, weather forecast, surface, days of rest since the previous fixture, distance traveled.
  • League context — current standings, points-to-target (relegation, promotion, European places), and stage of the season.
  • Squad availability — confirmed and expected line-ups, suspensions, injuries, international call-ups.
  • Live market odds — opening and current prices from a basket of major bookmakers, used as a sanity check on our own probabilities.

We deliberately do not buy or trade any private "insider" information about lineups, transfers, or referee decisions. Everything we use is publicly observable.

Step 2: The Statistical Model

Our core engine is a Poisson-based bivariate goal model with team-strength parameters that decay over time. In plain English, that means:

  • Each team has a separate offensive strength and defensive strength, calibrated from how many goals they score and concede compared to league average.
  • Recent matches are weighted more heavily than matches from a year ago. This is critical for catching teams in form changes — a team that won 10 in a row five months ago but has lost six straight should not be modeled as elite.
  • A separate home advantage parameter is applied per league (the home edge in Argentina is different from the home edge in the Premier League).
  • The model outputs a full probability distribution over every possible scoreline (0-0, 1-0, 0-1, 1-1, and so on up to 6-6+). From that distribution, we derive every market we cover: 1X2, double chance, BTTS, and over/under at every line.

We retrain the model continuously as new results land and recalibrate league-level parameters every two weeks.

Step 3: Identifying Value

A correct prediction is not the same as a profitable bet. If our model thinks the home team has a 70% chance to win, and the bookmaker prices that outcome at 1.30 (implied probability ~77%), there is no value — the price is shorter than our model suggests it should be. We only label something a "recommended pick" when our modeled probability is at least a few percentage points above the implied probability of the available odds. This is what professional bettors call a positive expected value (+EV) edge.

The discipline of skipping fixtures is just as important as picking ones to bet. On a typical matchday we will model 100+ fixtures and only publish predictions on a fraction of them.

Step 4: Human Review

Models are powerful but blind. They cannot see that a key playmaker is suspended, that a club has just sacked its manager, that a "dead rubber" fixture between two safe mid-table teams has different incentives than a relegation six-pointer. Before any pick is published:

  • A human analyst reviews the model output against the day's news, lineup announcements, and weather.
  • Picks where the analyst sees a meaningful contextual reason the model is wrong are removed from the slate, even if the math looks favorable.
  • Picks where the analyst is confident in the model's edge get published with a confidence rating (low / medium / high).

Step 5: Tracking and Auditing

Every published pick is logged the moment it is issued, with the odds available at the time. After the match finishes, the result is recorded — win, loss, or void. We do not delete losing picks. We do not re-grade picks after the fact. The cumulative log of wins, losses, and yield is what determines whether the service is worth subscribing to, and it is visible to all members.

This is the single most important thing we do. A prediction service that quietly hides its losses is essentially worthless, because there is no way to verify whether the strategy works.

What This Means For You

Even with all of the above, individual picks lose. That is a feature of how probability works, not a bug. A 60% favorite still loses 40 out of every 100 times you bet it. To benefit from any value-based service, you need to:

  • Use consistent stakes (we recommend a flat 1% of your bankroll per pick).
  • Track your own results across at least 100 picks before drawing any conclusions.
  • Be prepared for losing weeks — they happen even in profitable months.
  • Never chase a losing day with bigger stakes. If our edge is real, it shows up over hundreds of bets, not three.

Where To Go Next