Premier League Predictions: A Non-xG Approach

There are many viable approaches to assessing games in the Premier League. In recent years, the relatively new Expected Goals (xG) has dominated the discussion. This innovative stat, which assigns a goal probability to every shot in a game, is beloved by some and bemoaned by others. Many managers and pundits remain on the fence regarding the value of xG.

In our opinion, xG is a useful stat to consider, but it does come with a few limitations – most notably a failure to consider the quality of the player taking the shot. A chance inside the box may have an xG of 0.6 if Alexander Isak is lining up to shoot but would almost certainly be lower if the chance fell to a Liverpool midfielder or defender.

With the above in mind, we now present an approach to punting that removes xG from the equation entirely – creating a four-step plan that combines home and away percentage, shots on target per game, and goals per game.

Step 1: Predicting Possession

The first step in our system is to predict the likely possession of the two sides. We will use the Match Day 7 clash between Chelsea and Liverpool at Stamford Bridge as an example to illustrate how this process works.

Many sides often adopt slightly different approaches when playing at home and away. As such, the stats we are looking for are Chelsea’s average possession percentage at home in the Premier League and Liverpool’s average possession percentage in their away fixtures. Ahead of this game, those percentages were as follows:

• Chelsea Average Home Possession - 61.30%
• Liverpool Average Away Possession - 72.40%

Clearly, Chelsea can’t enjoy 61.30% possession in a game where Liverpool has the ball 72.40% of the time. Our next step is to scale this figure so that the total possession equals 100%. The calculation for this is as follows:

• Chelsea Predicted Possession = ((Average Chelsea Home Possession / (Average Chelsea Home Possession + Average Liverpool Away Possession)) × 100
• Liverpool Predicted Possession = ((Average Liverpool Away Possession / (Average Chelsea Home Possession + Average Liverpool Away Possession)) × 100

Plugging our 61.30% and 72.40% figures into the above equations gives us predicted percentage figures for the game of:

• Chelsea: 45.85%
• Liverpool: 54.15%

Step 2: Shots on Target Per Possession Point

Next, we need to assess how many shots on target each side has, and concedes, per possession point.

Starting with the Blues, Chelsea had an average of 4 shots on target per game for that average possession percentage of 61.30%. To work out shots on target per percentage point, we simply divide 4 by 61.30%, giving a value of 0.065 shots on target per percentage point.

From a defensive standpoint, we consider the percentage of time Chelsea are out of possession i.e., 39.70% (100-61.30 = 39.70), and their average shots on target against per game, which stood at 3.33 before the Liverpool game. This gives a shot on target against per percentage point of 3.33/39.70 = 0.085.

Next, we repeat the process for Liverpool. The Reds headed into the game with an average of 4 shots on target per away fixture. Combining this with their average possession percentage of 72.40% gives a shot on target per percentage point of 4/72.40 = 0.055.

Turning to Liverpool’s defensive side, Arne Slot’s troops had conceded an average of 4.67 shots on target from an average opposition possession percentage of just 27.60% (100-72.40 = 27.60). This gives a shot on target against per percentage point of 4.67/27.60 = 0.169.

Finally, we use the above information to create a predicted shots on target per percentage point for each of the sides.

Chelsea Calculation

• Chelsea Predicted Shots on Target per Possession Percentage Point = (Chelsea Average Shots on Target per Possession Percentage Point + Liverpool Average Shots on Target Against per Possession Percentage Point) / 2
• Chelsea Predicted Shots on Target per Percentage Point: (0.065 + 0.169) / 2 = 0.117

Liverpool Calculation

• Liverpool Predicted Shots on Target per Possession Percentage Point = (Liverpool Average Shots on Target per Possession Percentage Point + Chelsea Average Shots on Target Against per Possession Percentage Point) / 2
• Liverpool Predicted Shots on Target per Possession Percentage Point: (0.055 + 0.085) / 2 = 0.07

Step 3: Combining Predicted Possession and Predicted Shots on Target

Having obtained the predicted possession and shots on target per possession percentage point, we can now predict how many shots on target Chelsea and Liverpool can be expected to have in the game. This is as simple as combining the two values, as follows.

• Chelsea Predicted Shots on Target: 45.85 × 0.117 = 5.36
• Liverpool Predicted Shots on Target: 54.15 × 0.07 = 3.79

Step 4: Conversion Ratio

For our final step, we must assess how efficient each side is at converting their shots on target into actual goals. To create a goal per shot on target figure, we simply divide goals scored by shots on target.

Before this game, Chelsea had scored 11 times from 26 shots on target in the Premier League, creating a conversion rate of 11/26 = 0.42.

Liverpool, meanwhile, had scored 12 times from 28 shots on target in the Premier League, creating a conversion rate of 12/28 = 0.43.

With this final piece of the jigsaw, we can now combine our predicted Shots on Target values with our conversion rates to predict the goals scored by each side.

• Chelsea: 5.36 × 0.42 = 2.25
• Liverpool: 3.79 × 0.43 = 1.63

This process presents a clear and logical route towards the number of goals we can expect a side to score in any given game.

These predicted goal figures can be used to estimate likely score distribution (e.g. via a Poisson model) and in turn, inform betting markets such as correct score, over/under, and match result.

It is worth noting that the power of this model increases as the season progresses and the dataset encompasses more games. However, when combined with situational factors - motivation, injuries, suspensions, etc. - this method is a useful tool to have in your punting armoury, without the need for xG.