Repository files navigation

Predicting results for theFIFA Men's 2022 World Cup andFIFA Women's 2023 World Cup.

Matches are predicted using a framework based on the team-level model inAIrsenal, a Python library for using Machine learning to pick a Fantasy Premier League team, which in turn usesbpl which is anumpyro implementation of our model.

This model is trained on international mens football results obtained fromhttps://github.com/martj42/international_results andhttps://github.com/martj42/womens-international-results.

The original model is a version ofDixon and Coles.

There is some background information and description of the resultshere.

TheNew Scientist also has an article, as doesThe Daily Mail, which prompted one reader to enthuse:"What an absolute load of rubbish, but typical of the anti-English establishment today!", and another to ponder:"I have NEVER KNOW a Prediction from a Scientist to be Right Ever, But on this Prediction I may be wrong.". High praise indeed!

OurAIrgentina model came6th in@Futbolmetrix1'sWorld Cup 2022 Sophisticated Prediction Contest outperforming models developed by FiveThirtyEight, Opta and Betfair!

The easiest way to use the code is viapoetry. If you have poetry installed, from this directory, you can do

poetry shellpoetry install

to first open a shell in a virtual environment, and then install the dependencies and thewcpredictor package.

There are a couple of command-line applications that can be run when thewcpredictor package is installed as described above.

In order to simulate the 2022 Men's World Cup$N$ times, you can do

wcpred_run_simulations --num_simulations N --tournament_year 2022 --training_data_start<YYYY-MM-DD> --training_data_end<YYYY-MM-DD> --resume_from=Group --output_csv<outputfilename>

For the 2023 Women's World Cup, you can run

wcpred_run_simulations --womens --num_simulations N --tournament_year 2023  --years_training_data 20 --resume_from=Group --output_csv<outputfilename>

Notice here, instead of passing in the training start and end dates, we can simply pass how many years of training data to look backwards in time for. You can use the--resume_from CLI argument to specify at what point to simulate the tournament from.

The results, in the form of a table of how many times each team got to each stage of the competition, will be saved in the specified csv file. At present, the allowed values fortournament_year are "2014", "2018", "2022" (the default) and "2023".

⚠️Warning:⚠️ The simulations can usea lot of memory. If you want to run more than around 250 simulationns we recommend checking the help (wcpred_run_simulations --help) and setting the--num_thread and--per_tournament arguments (start low andd then slowly increase them whilst monitoring CPU/memory usage on your system).

Once you have a csv file saved from running the simulation, you can plot the top ten most frequent winners by running:

wcpred_plot_winners --input_csv<inputfilename> --output_png<outputfilename>

and the results will be saved in the specified png.

You can also make a plot showing how far in the tournament a selection of teams got, by running e.g.:

wcpred_plot_progress --input_csv<inputfilename> --output_png<outputfilename> --team_list"England,Wales"

Note that both these commands can be run with--help to see the options.

In a python session, to simulate the Men's 2022 World Cup, you can do something like:

>>>fromwcpredictorimportTournament,get_and_train_model>>>t=Tournament("2022")# can also choose "2018" or "2014">>>model=get_and_train_model(start_date="2016-06-01",end_date="2022-11-20")# choose dates for training data>>>t.play_group_stage(model)>>># at this stage, we can look at how each group is doing>>>print(t.groups["A"])Position|Team|Points|GS|GA1stNetherlands6842ndQatar4213rdEcuador4154thSenegal345>>># or, we can go ahead and play the knockout stages>>>t.play_knockout_stages(model)>>>print(t.winner)

Our top two favourites to win the competition, Brazil (25%) and Belgium (19%) got knocked out early on in the competition, getting eliminated in the quarter-final and group stages, respectively. However, AIrgentina gave the eventual winners, Argentina, a 13% chance of winning the tournament.

For England and Wales, our model suggested their most likely outcomes were being knocked out at the quarter-finals and group stages respectively which unfortunately were their actual outcomes in Qatar.

We entered our pre-tournament predictions to@Futbolmetrix1's WC2022 Sophisticated Prediction Contest and in the end came6th!

We'll see 😉