Excel LADZ
Introduction
I believe in absolute transparency, especially when it comes to probability - the foundation of how risk is assessed and taken. Here, I’ll recount exactly how I built my model to forecast each football season and the reasoning behind it.
Hopefully, this helps you understand both the advantages and disadvantages of the model.
ATT & DEF Ratings
In the Excel LADZ European Football model, ATT (Attack Rating) and DEF (Defense Rating) are numerical measures of a team’s scoring strength and defensive strength. Essentially, the ATT rating quantifies how effective a team is at creating and scoring goals relative to other teams, while the DEF rating shows how well a team limits goals conceded compared to the league average.
For example, if a team has an ATT rating of 1.25, it means they score about 25% more goals than the league average — this reflects strong offensive performance. On the other hand, a DEF rating of 0.80 implies the team concedes 20% fewer goals than average, indicating strong defensive performance.
ATT and DEF ratings are used to calculate the expected goals for each team in a matchup, so it’s important they’re measured accurately.
I’ve chosen to compile the ATT and DEF ratings using a blend of on-field performance this season and relative squad market value per minute played. Specifically, the model applies an arbitrary weighting of 75% to current-season performance and 25% to squad value.
The idea is to strike a balance between what teams have actually produced on the pitch and the underlying quality of the players available to them. Season performance captures form, tactics, and execution, while squad value acts as a proxy for long-run ability and depth.
By combining both, the model is more likely to reduce the impact of short-term luck—such as teams overperforming from limited chances or underperforming despite strong underlying quality. This helps ensure teams aren’t unfairly rewarded for fortunate results or overly punished for bad luck, leading to ratings that better reflect true attacking and defensive strength.
The following formulas show how each team’s current-season performance ATT and DEF ratings are calculated. Put simply, the ratings are normalising the goals scored and conceded per game relative to the league average.
​
​
​
​​
​
​
​
​
​
Next, to assess the underlying strength of a team’s lineup, I’ve chosen to measure each team’s market value per minute played. This provides a clear, observable measure of talent and depth, avoiding the much harder task of precisely defining each individual player’s expected on-field performance.
Using historical data, the relative market values for each team are then mapped into equivalent ATT & DEF ratings to ensure they are directly comparable with the current-season team performance measure.
Specifically, five years of historical market value per minute played data is collected for each league independently using Transfermarkt. For each season, team values are normalised by dividing through by the league average, producing a relative measure of squad quality.
These relative values are then plotted against observed attacking and defensive outputs, and a power-law trendline is fitted to the data. This naturally anchors a squad with no market value to an ATT ratings of zero and, conversely, implies maximal defensive vulnerability for squads with no defensive quality.
The relationships between relative squad market value and both the ATT and DEF ratings for the EPL are plotted below as an example.
Finally, the overall ATT and DEF ratings are calculated by combining the team performance-based and lineup value-based measures using the previously defined 75% and 25% weightings. For example, below are the resulting ratings for the EPL after Matchday 25 of the 2025-26 season.
To finish, it’s important to acknowledge several limitations and modelling choices within the ATT and DEF rationing framework. First, the 75% / 25% weighting between current-season performance and squad market value is arbitrary by design. While it provides a sensible balance in most cases, it can be adjusted within the spreadsheet to reflect the amount of available information — particularly as the sample size of matches grows.
Second, at the start of a season, there is no current performance data from which to calculate ATT and DEF ratings. To address this in previous seasons, I’ve incorporated a trailing performance window that gradually phases out historical information.
Specifically, using a 20-game threshold, ratings are blended with prior-season performance. After 10 games, ratings are a 50/50 mix of last season and current season data; after 15 games, this shifts to a 25/75 split; and once 20 matches are played, ratings are derived entirely from current-season performance. This approach ensures stability early while allowing form to dominate as data accumulates.
Lastly, in combination with this, ratings can also be explicitly reverted toward the league mean at the start of the season to account for heightened uncertainty when sample sizes are small.
Expected Goals
Expected goals (xG) represent the average number of goals a team is expected to score in a match given the relative attacking and defensive strengths of the two teams involved. Rather than predicting a single scoreline, expected goals define the mean of a probability distribution for goals scored, which will be used in the next section to simulate full match outcomes.
In this model, a team’s expected goals are calculated by combining three core components. First, the team’s ATT rating captures its scoring strength, which is then adjusted by the opposition’s DEF rating to reflect defensive resistance. This product is scaled by the league-average goals per team per match, ensuring results remain consistent with the competition’s scoring environment.
Finally, a home advantage (or away disadvantage) multiplier is applied to account for systematic differences in performance by venue.
Using historical data from the past five seasons, the home advantage for each league is estimated and tested for significance using 95% confidence intervals. In each case, the estimated mean effect is statistically significant, and this league-specific average home advantage is applied uniformly within the model.
The home advantage is applied symmetrically to both teams’ expected goals. For the home team, expected goals are multiplied by the square root of 1 + Home Advantage, while the away team’s expected goals are scaled by the reciprocal of this factor.
This ensures that the combined effect across both teams equals the full home advantage, while preserving balance in the total expected goals of the match.
Simulating a Game
To simulate a football match, goals must be generated from a probability distribution rather than treated as fixed outcomes. In this model, the number of goals scored by each team is assumed to arise from a stochastic counting process, where goals occur randomly but around a known average rate.
​
This naturally leads to the Poisson distribution, which is widely used to model the number of events occurring in a fixed interval when those events are independent and relatively rare.
​
Here, each team’s expected goals (xG) acts as the mean parameter of the Poisson process. Rather than drawing directly from a Poisson random generator, goals are simulated using the BINOM.INV function in Excel.
​
This works because the Poisson distribution can be derived as the limiting case of a Binomial distribution as the number of trials becomes large and the probability of success becomes small, while the expected value remains fixed. Practically, this allows repeated simulations to be run efficiently in Excel.
​
However, empirical testing reveals limitations. When the Poisson distribution is assessed using a chi-squared goodness-of-fit test, the null hypothesis is rejected in the EPL, La Liga, and Bundesliga at the 5% significance level. This indicates that the goal counts in these leagues deviate significantly from the Poisson assumption.
​
This failure is driven by overdispersion in the goal data, where the observed variance in team goal counts exceeds the mean—directly violating a core assumption of the Poisson distribution.
​
In practice, this manifests as high-scoring teams recording more extreme goal totals than the Poisson model predicts. As shown in the EPL table below, the upper goal bins contain a greater frequency of matches than expected under a Poisson process. The number of zero-goal outcomes is also lower than predicted.
These excess counts inflate the chi-squared test statistic, particularly in the tail of the distribution, which in turn drives the p-value lower and leads to rejection of the null hypothesis.
One way to account for this overdispersion is to replace the Poisson assumption with a Negative Binomial or generalised Poisson distribution, both of which allow the variance to exceed the mean and therefore accommodate heavier tails in the goal distribution.
In fact, a Generalised Poisson distribution has been successfully used in my NFL model this season for precisely this reason.
​
An alternative approach is to manually reweight specific scorelines—for example, increasing the probability of outcomes like 0–0 that are empirically underrepresented—while proportionally scaling down the remaining probabilities to ensure they still sum to one.
​
This type of calibration has previously been employed by the 538 soccer model, offering a pragmatic fix that improves empirical fit without changing the underlying distributional assumption.
However, the current football model retains the standard Poisson framework. This prioritises simplicity, while remaining flexible to future upgrades as the model continues to be tested, refined and expanded.
Simulating a Season
To simulate an entire season, every remaining fixture is simulated using the match-level process described above, producing a full simulated league table for that run. This process is then repeated 10,000 times, generating a distribution of possible season outcomes rather than a single deterministic prediction.
​
Each simulation reflects the inherent randomness in match results, allowing probabilities to be assigned to final league positions, points totals, and outcomes such as titles or relegation.
Increasing the number of simulations improves the stability of these estimated probabilities by reducing Monte Carlo error; however, some variance remains even with 10,000 runs. As a result, small probability differences should be interpreted with caution, particularly for closely matched teams or low-probability events.
Important Note
After reading this page, it should be fairly obvious that the model has a few disadvantages.
​
Team-Performance ATT & DEF ratings don't consider the impact of red cards, weather and referees on goals accrued and conceded. Furthermore, the Lineup-Based ATT & DEF ratings are backwards-looking. If a team experiences an injury crisis, the projections aren't able to pick up the value of the starting lineup for the next immediate match.
Therefore, please recognise that this is purely a mathematical model. I accept no liability for any outcomes or decisions made using it!
​
ENJOY!