Neil-Warnock AI

Working with modern statistical learning techniques to "solve" the game of Fantasy Premier League.

Methodology

The approach used can be resumed as a two-layer prediction system feeding an constrained optimiser.

USE ONLY NORMALISED PERFORMANCE METRICS!

This will allow us to treat edge cases where there is not enough data about a team's, player's or manager's past performance. This is problem that is especially poignant in football; players can be transfered from abroad, teams can be promoted from lower leagues, posing a challenge to our data-hungry classifier algorithms. However; if we treat every performance metric as a distribution, we are able to fill in the blanks and guess that an entity's missing metric will satisfy the average, this plugs the hole in our predictive raft (and there will be lots of holes to plug in future also).

Prediction

See predictor.jl for a thorough functionality description.

Layer 1:

scoreline_predictor( gameweek fixtures ) -> predcited scores (club level)

Layer 2:

player_points_predictor( historical player data ) | predicted scores -> predicted points (player level)

Layer 1 detailed description

We will begin by attempting to predict high level outcomes of a football match, in other words, the expected scoreline given a home and away team pair. For this, we shall train, validate and compare the accuracy of several different statistical models on Premier League data ranging over the 2016-2017 to 2020-2021 seasons.

The outcome of this process will be a 'universal approximator' function which maps an input:

fixture = (<home team>, <away team>)

to an output:

predicted_scoreline = (<home goals>, <away goals>, <uncertainty>)

However, it is clear that simply passing the fixture data, most likely in the form of team-identifying symbols (e.g :arsenal and :tottenham), will not be enough features to train the model (in the example it should be plain to see that the correct outcome will be (5, 1, 0%)). We therefore require more features, engineered and raw, to qualify the likelihood of a scoreline outcome.

Using data from the football-data database should provide us with a good foundation for this type of event prediction.

Layer 2 detailed description

Optimisation

Having predicted the gameweek points haul for the entire cohort of players (or those within a reasonable expected accuracy window)*, we now turn to choosing which of the players to use to fill our 15 available spaces.

Investigate potential α

• use fpl-twitter and sentiment analysis to take inspiration from 'experts'
  • include this as a weighted parameter
• use betting odds data to tune scoreline prediction
  • build api to fetch data and update weekly?
• were in lundun aren't weh?