Awesome

Mezzala

Models for estimating football (soccer) team-strength

Install

pip install mezzala

How to use

import mezzala

Fitting a Dixon-Coles team strength model:

First, we need to get some data

import itertools
import json
import urllib.request


# Use 2016/17 Premier League data from the openfootball repo
url = 'https://raw.githubusercontent.com/openfootball/football.json/master/2016-17/en.1.json'


response = urllib.request.urlopen(url)
data_raw = json.loads(response.read())

# Reshape the data to just get the matches
data = list(itertools.chain(*[d['matches'] for d in data_raw['rounds']]))

data[0:3]

[{'date': '2016-08-13',
  'team1': 'Hull City AFC',
  'team2': 'Leicester City FC',
  'score': {'ft': [2, 1]}},
 {'date': '2016-08-13',
  'team1': 'Everton FC',
  'team2': 'Tottenham Hotspur FC',
  'score': {'ft': [1, 1]}},
 {'date': '2016-08-13',
  'team1': 'Crystal Palace FC',
  'team2': 'West Bromwich Albion FC',
  'score': {'ft': [0, 1]}}]

Fitting a model

To fit a model with mezzala, you need to create an "adapter". Adapters are used to connect a model to a data source.

Because our data is a list of dicts, we are going to use a KeyAdapter.

adapter = mezzala.KeyAdapter(       # `KeyAdapter` = datum['...']
    home_team='team1',
    away_team='team2',
    home_goals=['score', 'ft', 0],  # Get nested fields with lists of fields
    away_goals=['score', 'ft', 1],  # i.e. datum['score']['ft'][1]
)

# You'll never need to call the methods on an 
# adapter directly, but just to show that it 
# works as expected:
adapter.home_team(data[0])

'Hull City AFC'

Once we have an adapter for our specific data source, we can fit the model:

model = mezzala.DixonColes(adapter=adapter)
model.fit(data)

DixonColes(adapter=KeyAdapter(home_goals=['score', 'ft', 0], away_goals=['score', 'ft', 1], home_team='team1', away_team='team2'), blocks=[TeamStrength(), BaseRate(), HomeAdvantage()]), weight=UniformWeight()

Making predictions

By default, you only need to supply the home and away team to get predictions. This should be supplied in the same format as the training data.

DixonColes has two methods for making predictions:

predict_one - for predicting a single match
predict - for predicting multiple matches

match_to_predict = {
    'team1': 'Manchester City FC',
    'team2': 'Swansea City FC',
}

scorelines = model.predict_one(match_to_predict)

scorelines[0:5]

[ScorelinePrediction(home_goals=0, away_goals=0, probability=0.023625049697587167),
 ScorelinePrediction(home_goals=0, away_goals=1, probability=0.012682094432376022),
 ScorelinePrediction(home_goals=0, away_goals=2, probability=0.00623268833779594),
 ScorelinePrediction(home_goals=0, away_goals=3, probability=0.0016251514235046444),
 ScorelinePrediction(home_goals=0, away_goals=4, probability=0.00031781436109636405)]

Each of these methods return predictions in the form of ScorelinePredictions.

predict_one returns a list of ScorelinePredictions
predict returns a list of ScorelinePredictions for each predicted match (i.e. a list of lists)

However, it can sometimes be more useful to have predictions in the form of match outcomes. Mezzala exposes the scorelines_to_outcomes function for this purpose:

mezzala.scorelines_to_outcomes(scorelines)

{Outcomes('Home win'): OutcomePrediction(outcome=Outcomes('Home win'), probability=0.8255103334702835),
 Outcomes('Draw'): OutcomePrediction(outcome=Outcomes('Draw'), probability=0.11615659853961693),
 Outcomes('Away win'): OutcomePrediction(outcome=Outcomes('Away win'), probability=0.058333067990098304)}

Extending the model

It's possible to fit more sophisticated models with mezzala, using weights and model blocks

Weights

You can weight individual data points by supplying a function (or callable) to the weight argument to DixonColes:

mezzala.DixonColes(
    adapter=adapter,
    # By default, all data points are weighted equally,
    # which is equivalent to:
    weight=lambda x: 1
)

DixonColes(adapter=KeyAdapter(home_goals=['score', 'ft', 0], away_goals=['score', 'ft', 1], home_team='team1', away_team='team2'), blocks=[TeamStrength(), BaseRate(), HomeAdvantage()]), weight=<function <lambda> at 0x123067488>

Mezzala also provides an ExponentialWeight for the purpose of time-discounting:

mezzala.DixonColes(
    adapter=adapter,
    weight=mezzala.ExponentialWeight(
        epsilon=-0.0065,               # Decay rate
        key=lambda x: x['days_ago']
    )
)

DixonColes(adapter=KeyAdapter(home_goals=['score', 'ft', 0], away_goals=['score', 'ft', 1], home_team='team1', away_team='team2'), blocks=[TeamStrength(), BaseRate(), HomeAdvantage()]), weight=ExponentialWeight(epsilon=-0.0065, key=<function <lambda> at 0x122f938c8>)

Model blocks

Model "blocks" define the calculation and estimation of home and away goalscoring rates.

mezzala.DixonColes(
    adapter=adapter,
    # By default, only team strength and home advantage,
    # is estimated:
    blocks=[
        mezzala.blocks.HomeAdvantage(),
        mezzala.blocks.TeamStrength(),
        mezzala.blocks.BaseRate(),      # Adds "average goalscoring rate" as a distinct parameter
    ]
)

DixonColes(adapter=KeyAdapter(home_goals=['score', 'ft', 0], away_goals=['score', 'ft', 1], home_team='team1', away_team='team2'), blocks=[TeamStrength(), HomeAdvantage(), BaseRate()]), weight=UniformWeight()

To add custom parameters (e.g. per-league home advantage), you need to add additional model blocks.