One of the most powerful aspects of data science is its ability to quantify real world phenomena. To use a popular example in the realm of football, fans have always recognised that some chances are of a higher quality than others - a six-yard tap-in is clearly more likely to result in a goal than a speculative shot from outside of the box - but the advent of expected goals (xG) and other metrics has changed the way we analyse and understand the game.
But what about the influence of weather? While we might instinctively feel that a rain-slicked pitch and strong crosswinds can affect styles of play, these effects are rarely quantified. Can the elements actually influence styles of play or match outcomes? Using Met Office weather observation data in conjunction with StatsBomb data for the 23/24 season, I’ll explore the impacts of adverse weather conditions on football playing methods.
This project grew in scope quite quickly, so I broke it up into two articles. The first will begin by breaking down the data collection process, before explaining the analysis approach and finally the observations and correlation between rain and directness and pace towards goal, as well as secondary metrics such as ball in play time. The second article will focus on the impacts of wind and temperature.
Data Collection and Cleansing
Tools used:
- Snowflake (data collection and manipulation)
- Python (football data collection and manipulation)
- Streamlit (data bisualisation web app)
This analysis is based on Met Office weather observations for UK postcodes, and therefore includes match data for the entire English Premier League, Championship, League One and League Two. I selected the 23/24 season as the most recent full season of data, but if there's interest I'll aim to update this analysis later with past and current season data to increase the sample size.
Stadiums
I gathered a list of Stadiums in the UK with the help of Google and ChatGPT, then tweaked the team names and stadium names to match those used by StatsBomb. I then ran a simple function to translate the postcode to the Met Office postcode sector format (e.g. ‘PL2 3DQ’ -> ‘PL2_3’). This would allow me to look up the weather observations for a particular stadium postcode sector.
Met Office Weather Observations
Met Office “best estimate” observations via the Snowflake Marketplace. Sourced from Nowcast Analysis - a technique for very short-range forecasting.This combines their station, satellite and radar observations onto a 2km grid, with terrain adjustment to interpolate between station locations, giving a best estimate of actual weather conditions in population centres. A new observation is produced every hour for each postcode in the UK, and includes the parameters listed below as well as several others such as solar elevation angle and visibility screen level.
I selected the following columns as the most likely to have an impact on playing conditions:
| Column | Definition |
|---|---|
| Total_rainfall_amount_previous_hour | Amount of rainfall in the previous hour measured in millimetres. |
| Wind_speed_10m | Average wind speed over the previous 15 minutes at 10m above the ground. |
| Wind_gust_10m | Extreme value of wind speed over the previous 15 minutes at 10m above the ground. |
| Feels_like_temperature | Approximate temperature in celsius over the previous hour. |
See full datasheets here:
StatsBomb Season Matches
A basic overview of each match in the four leagues for the 23/24 season, gathered using the `matches` method of the StatsBombPy library. (I later realised that I could have saved time by inserting the postcode and sector reference into the match metrics instead of creating my separate stadium table)
Using stadium postcode sectors and match timestamps in a join, I created a new table containing the match overviews and observations within three hours of kick off timestamps. For example, a 3PM kick-off would return the 3PM, 4PM and 5PM observations to hopefully give a full picture of the weather conditions just before and during the match.
I also performed some minor datetime manipulation after discovering that StatsBomb kick off times are published in Eastern European Time (UTC+2), a fact that was quite well hidden in the StatsBomb documentation.
StatsBomb Team Match Metrics
Using the match IDs from the Season Matches dataset I gathered the detailed team match metrics for all matches across the four leagues. It includes over 150 different metrics, from “Goals scored from throw-ins” to “Shots generated from high press” to “Percentage of opponent dribbles completed”.
This article focuses mainly on team match directness and match pace towards goal. These are both metrics which measure a team’s directness, with team_match_directness measuring the pitch distance travelled in buildup a shot (e.g. more horizontal passes = greater distance = lower directness value), and team_match_pace_towards_goal measuring from the start of possession to the shot in metres per second.
I selected a few additional columns that I felt could potentially be affected by weather conditions. These are shown in the table below:
| Column | Definition |
|---|---|
| Team_match_directness | A ratio of the distance towards goal from the start of a possession that ended in a shot, divided by the total distance travelled in buildup to the shot. |
| Team_match_pace_towards_goal | The average speed of buildup for possessions that end in shots, from the start of possession to shot (m/s) |
| Team_match_gk_pass_distance | Average length of pass from a goalkeeper when playing out from the back. |
| Team_match_gk_long_pass_ratio | Percentage of goalkeeper long passes that are completed. |
| Team_match_ball_in_play_time | Actual clock time the ball spends in play |
| Team_match_dribble_ratio | Percentage of dribbles that were completed |
| Team_match_high_press_shots_conceded | Shots conceded from possessions that were lost within 5 seconds of a opposition defensive action in the team’s half |
| Team_match_passing_ratio | Passing completion rate. |
This is a sample of the available metrics that I selected as I felt they could be impacted by weather conditions. For example, a wet surface might cause a team to lose possession more frequently under a high press, or high wind speeds could result in fewer successful long range passes. I would welcome suggestions on other relevant StatsBomb metrics for analysis.
Team Match Stats + Weather Metrics
The final dataset was created by joining the combined weather observation and match overview data with the detailed match metrics, joining on match ID. This resulted in a table containing team metrics and weather conditions for every 23/34 match.
Approach
I aimed to find any correlations between match metrics and weather conditions using Pearson correlation. This method shows how two variables relate, with values ranging from -1 to 1: +1 signifies that both increase together, -1 signifies that one increases as the other decreases, and 0 suggests no linear relationship at all.
As such, I was looking to see if there was a strong positive or negative correlation between any of the metrics and adverse weather conditions (evidenced by diagonal trendlines in scatter graphs, as shown below).
I created a small Streamlit web app to visualise the data and determine if there was any correlation between rainy/windy conditions and the match metrics. All the graphs seen from this point were generated in the Streamlit app.
Overall Observations
The values for `total_rainfall_previous_hour` ranged between 0 (for the absence of any rain) and a maximum value of 28.9 which was linked to a match between Arsenal and Manchester United on 12th May 2024. I initially thought this was a bug in the data - the next highest values in the dataset are 17.9 and 10.8 before dropping into single digits.
This match famously saw the roof of Old Trafford leaking huge amounts of water, proving to be a useful test case for the data collection.
Visualising the Pearson correlation in a bar chart we can see that coefficient for pace towards goal and rainfall is close to 0.
If we switch to a scatter graph we can see a small negative trend between rainfall and pace towards goal.
However, in all cases the trendline appears to be skewed negatively by
four weather data outliers highlighted below.
Notably, two of these teams are Manchester United and Arsenal, the other two are Watford and West Brom. Two sets of teams which participated in a match with extreme rainfall.
At this point it seemed appropriate to winsorise the weather metric outliers (“clamping” the values so that anything above a certain threshold is set to the threshold value) thereby reducing the effect of the outliers on our analysis. According to the Met Office, rain at a rate of 4mm or more per hour is termed “Heavy Rain”, so this seems like an appropriate value to use as our upper limit.
When we apply the 4mm rainfall clamp we no longer see the scatter chart trendline “dragged down” for directness and pace towards goal. The result is a nearly flat trendline, indicating no correlation between rainfall and directness.
The table below shows the correlation between Total Rainfall Previous Hour (upper values trimmed to 4mm/hr) and various match metrics:
| Match Metric | Pearson Value |
|---|---|
| team_match_pace_towards_goal | -0.0029 |
| team_match_directness | -0.0014 |
| team_match_gk_pass_distance | 0.001 |
| team_match_gk_long_pass_ratio | 0.0006 |
| team_match_ball_in_play_time | 0.0149 |
| team_match_dribble_ratio | 0.0122 |
| team_match_high_press_shots_conceded | -0.004 |
| team_match_passing_ratio | -0.0009 |
We can see from the above table that there is no significant correlation between rainfall and the selected match metrics. The most significant trends are Ball In Play Time and Successful Dribble Ratio increasing with rainfall, but these are still within a range that could be considered coincidental.
Team Observations
While there aren’t any obvious overarching trends, we can see some interesting data points if we take a granular, per-team view.
Plymouth Argyle
As a Plymouth Argyle fan, I was curious to note that they had a higher than average correlation between rainfall and various match metrics. Argyle employed a direct, transitional style of football and had the highest average pace to goal in the Championship for 23/24. For example, goalkeeper passing distance (0.132), directness (0.199) and pace towards goal (0.155) all increased with the levels of rainfall. The effect was slightly more pronounced for away games, reflecting an increased tendency to “go long”, particularly in wet conditions.
West Bromwich Albion
Another team whose goalkeeper passing distance increased drastically away from home, and under rain conditions (Pearson 0.294). Across the whole season Alex Palmer was 5th lowest for goalkeeper long passes in the Championship, attempting 4.48 per game. This tactic appeared to completely revert when playing in the rain, and particularly in away matches as shown in the scatter graph below.
Accrington Stanley
Accrington were significant data outliers when comparing passing completion percentage with rainfall. Their passing success decreased in correlation with increased rain, resulting in a Pearson value of -0.306. Their average passing ratio was one of the lowest in the division overall with 68%, an issue that was compounded for home games in wet conditions. This could point to poor drainage at the Wham Stadium contributing to incomplete passes.
Leeds United
Leeds saw a large decrease in ball in play time in relation to rainfall (Pearson -0.297). However, the Leeds matches which saw the most rainfall included opposition such as Millwall, Rotherham, Preston and Huddersfield - all teams with below average ball in play time so it’s likely that this correlation is purely coincidental and has nothing to do with the weather conditions.
Final Thoughts
From this analysis there appeared to be no obvious, overarching trends between increased rainfall and directness of play or any of the secondary metrics such as goalkeeper passing length, match in ball time or high press shots conceded. This is at least somewhat notable by itself, as many fans might assume rain a “leveller” of sorts.
Given that the majority of matches don’t experience heavy rain or wind, a single season of data isn't really enough to gain meaningful insights. While winsorising the data reduced the impacts of data outliers skewing the analysis, a larger sample size would further help to alleviate this issue.
There are many variables at play during a match such as team tactics, game state, opposition quality which impact playing patterns far more directly than the weather. This makes it difficult to find the signal in the noise - as we saw in the Leeds United example their correlation was more likely caused by the opposition than playing conditions.
Another factor to consider is the modernisation of stadiums and pitch technology which have lessened the impacts of rain and wind. Drainage systems prevent build up of water on grass, and a bowl-shaped stadium may be better protected from wind than open-cornered ones.
Finally, this study has only scratched the surface of analysing weather’s impact on football. I will update this blog with another article exploring the effect of wind and temperature based on the Met Office data, but my long term goal is to build a model that evaluates the impact over thousands of games and help to predict future match outcomes. To do this, I’ll need not only a single season of UK data but multiple seasons across different countries. The Open Meteo Historical Weather API looks like a great starting point for this.
If you’d like access to the Streamlit app, give me a shout on BlueSky, Twitter or LinkedIn and I’ll pass on the details.