,
12 tweets,
5 min read
Read on Twitter
Using Player-Level Stats to Predict Goals - a {THREAD}
As #FPL managers, we are interested in predicting which players are going to score goals.
Whilst the particular stats we choose to use are largely a matter of taste, this thread explores which are the most predictive.
1/n
As #FPL managers, we are interested in predicting which players are going to score goals.
Whilst the particular stats we choose to use are largely a matter of taste, this thread explores which are the most predictive.
1/n
Advanced football stats (chiefly Expected Goals, xG) are never far from being a hot topic, not least in our FPL Community.
I aim to take an objective look at which of the commonly-used stats best predict future goals, much in the style of my xA study.
2/n
I aim to take an objective look at which of the commonly-used stats best predict future goals, much in the style of my xA study.
2/n
This involves splitting last season into two halves & observing how well individual metrics from the 1st half (H1):
1. Describe goal rates in H1
2. Predict themselves in H2
3. Predict goal rates in H2
by looking at the strength of the relationship & the variance explained.
3/n
1. Describe goal rates in H1
2. Predict themselves in H2
3. Predict goal rates in H2
by looking at the strength of the relationship & the variance explained.
3/n
The metrics examined are:
~ Shots in the Box (SiB)
~ Touches in the Box (TiB)
~ Big Chances (BC)
~ Shots on Target (SoT)
~ Expected Goals (xG)
For all of which we look at the rate of occurrence per 90 mins, at a player level, for players with at least 450 mins in H1 & H2.
4/n
~ Shots in the Box (SiB)
~ Touches in the Box (TiB)
~ Big Chances (BC)
~ Shots on Target (SoT)
~ Expected Goals (xG)
For all of which we look at the rate of occurrence per 90 mins, at a player level, for players with at least 450 mins in H1 & H2.
4/n
Shots in the Box (SiB90)
1. Description - regressing goals per 90m in H1 of the season vs SiB90, we see a nice positive correlation.
~ The x-coefficient of 0.3452 means that roughly 35% of SiB resulted in goals.
~ The R-Squared, 0.72, is the % of variance explained.
5/n
1. Description - regressing goals per 90m in H1 of the season vs SiB90, we see a nice positive correlation.
~ The x-coefficient of 0.3452 means that roughly 35% of SiB resulted in goals.
~ The R-Squared, 0.72, is the % of variance explained.
5/n
2. Self-prediction - H1 vs H2
~ the ideal is a 1:1 relationship (y=x) and full variance explained (R-Sq = 1);
~ we will see that the x-coefficient (0.87) & the R-Sq (0.74) for SiB are relatively weak - you can't count on it predicting itself particularly well over time.
6/n
~ the ideal is a 1:1 relationship (y=x) and full variance explained (R-Sq = 1);
~ we will see that the x-coefficient (0.87) & the R-Sq (0.74) for SiB are relatively weak - you can't count on it predicting itself particularly well over time.
6/n
3. Predicting Goals
Here we can see a positive relationship between H1 SiB & H2 Goal rates, but the variance explained (0.5) turns out to be the lowest of all the measures we examine.
In other words, this measure is the least capable of nuanced future predictions.
7/n
Here we can see a positive relationship between H1 SiB & H2 Goal rates, but the variance explained (0.5) turns out to be the lowest of all the measures we examine.
In other words, this measure is the least capable of nuanced future predictions.
7/n
Touches in Box
1. Relatively poor description of goals (1st image).
2. Excellent self-prediction: a very reliable measure for projecting forwards.
3. Overall an average Goals predictor, due to the combination of effects in 1 & 2.
8/n
1. Relatively poor description of goals (1st image).
2. Excellent self-prediction: a very reliable measure for projecting forwards.
3. Overall an average Goals predictor, due to the combination of effects in 1 & 2.
8/n
Big Chances
1. Reasonable descriptive quality
2. Poor predictor of itself - perhaps due to the paucity of events and space for subjectivity in the labeling.
3. Hence a somewhat compromised predictor of future goals.
9/n
1. Reasonable descriptive quality
2. Poor predictor of itself - perhaps due to the paucity of events and space for subjectivity in the labeling.
3. Hence a somewhat compromised predictor of future goals.
9/n
Shots on Target
1. Rather a poor describer; less than 20% of SoT's result in goals.
2. Excellent self-predictor; this can be extrapolated confidently.
3. Decent predictor of future goals.
10/n
1. Rather a poor describer; less than 20% of SoT's result in goals.
2. Excellent self-predictor; this can be extrapolated confidently.
3. Decent predictor of future goals.
10/n
(OPTA) Expected Goals
1. Excellent description; relationship ~1:1 & largest variance explained.
2. Very good self-prediction; high x-coef & R-Sq. Rates can be projected with confidence.
3. The most amount of variance explained of all the measures, i.e. the best predictor.
11/n
1. Excellent description; relationship ~1:1 & largest variance explained.
2. Very good self-prediction; high x-coef & R-Sq. Rates can be projected with confidence.
3. The most amount of variance explained of all the measures, i.e. the best predictor.
11/n
Conclusions
~ All of these measures will to some degree help predict future goal rates.
~ By every measure xG is shown to be the most reliable predictor.
~ xG also benefits from being in the same units as goals - you don't need to convert it!
Hope this helps; leave me alone.
~ All of these measures will to some degree help predict future goal rates.
~ By every measure xG is shown to be the most reliable predictor.
~ xG also benefits from being in the same units as goals - you don't need to convert it!
Hope this helps; leave me alone.
