FBPredictor

Scrape football stats and train models to predict the outcome.

Scrape matches

Example scraping command:

python3 scripts/scrape.py -c configs/scrape_cfg.yaml

Train model

Example (GBDT) training command:

python3 scripts/train.py -c configs/train_cfg.yaml (--feature_select) (--hp_opt) (-a)

There are many features to train with - if you want to reduce them, try running with a modified Boruta-SHAP algo by adding the option --feature_select. To optimise the model hyperparameters, add the --hp_opt option. To add info on betting odds from football-data.co.uk, add the -a option.

The current best test accuracy is: 67.5%. For thee class models that predict draws (see branch: three_class, the best accuracy is around 53%).

predict upcoming matches and back test

predicting upcoming matches

To make a prediction on each match in a given league e.g. the EPL in 2022:

python3 scripts/predict.py --config configs/train_cfg.yaml -l "Premier League" -y 2022 --model models/model.json 

Back testing

Adding the -b [ammount] option backtests with the Kelly Critereon used as the betting strategy. There is also the option to use this combo with the -f option to fix a bet at the ammount specified by -b, rather than updating the total i.e. assume same bank balance of -b [ammount] before each bet. This is less risky if the updated pot gets larger.

Using KC on the realistic three class scenarios gives positive but small returns over bookies.

To-Do's: scraping

• weigh up scraping more data v.s. not having expected-goals etc. for older years -[x] -> no real improvemnt (see more_years branch)
• check if BundesLiga matches are somehow being scraped in EPL

To-Do's: model training

• Try and 3 class model for predicting draws too
• remove hard coding on train/test split date
• Plot inputs and make correlation map (separate script?)
• Try different models
• debug why some rows are lost in training join

To-Do's: betting strategy

• consider how to quantify an uncertainty on the game probs

To-Do's: general

• go through all FIXME's
• think about whether to drop either home/away match to prevent double same info being fed twice (FYI done autoamticaly if joinining odds info in)
• fix issue with having to scrape data and re-run workflow each time you want to predict
• add docstrings
• scrape live odds for upcoming matches if adding these features to models