Introduction

The algorithm made 192 wrong predictions and 188 correct predictions. This is an accuracy of 49.5%. If you placed a 1 euro bet on every predicted result over the course of the seaon you would have an outlay of 380 euro and a return of 349.49 euro. As such you would have lost 30.51 euro. The biggest win was selecting Spurs to beat City at home at odds of 4.1. The most unexpected result was City’s draw with West Brom. City were 1.12 but the game was a draw (odds of 9.96).

Accuracy by Team

Accuracy was highest with Man City and lowest with Everton.

Predicted Points vs Actual

The model was overconfident for the top teams, expecting them to pick up more points over the course of the season but overall the model performed well. At the bottom end of the table Crystal Palace were predicted to perform much worse than they did in reality

Predicted Table

If the model had been accurate for all games, Liverpool would have been league winners. The predicted top four are identical to the actual top four, although in a different order. The inflated points finish is due to the model aversion to picking draws. Two of the three predicted relegated teams are correct. Crystal Palace performed better than expected and were not relegated. Sheffield United performed worse than expected and were relegated.

Predicted Premier League Table
Premier League 2020/2021
Position	Team	Points
1	Liverpool	111
2	Man City	105
3	Chelsea	93
4	Man United	90
5	Tottenham	84
6	Leicester	75
7	Arsenal	66
8	Wolves	60
9	Brighton	54
10	Everton	54
11	West Ham	54
12	Leeds	48
13	Southampton	48
14	Aston Villa	40
15	Sheffield United	37
16	Burnley	36
17	Newcastle	27
18	Crystal Palace	24
19	Fulham	24
20	West Brom	9

Comparison to other betting strategies

Whaat would have happened if we placed our 1 euro bet on the favourite in ever game? Our outlay would be 380 euro and our return would be 327.02 euro, an overall loss of 52.98 euro. As we saw earlier the random forest model performs slighly better than this. Popular broadcaster Mark Lawrenson does a weekly prediction of premier league fixtures on the BBC. This data was collated to compare to our algorithm and to the bet on favourite strategy. While Lawro predicted fewer accurate results than the machine learning algorithm, Lawro’s tips actually made a profit of 9.65 euro from a return of 389.65 euro.

What happens if we selected a result at random? Of course if we did this for one season and compared it to a second season we may achieve very different returns. However if we do this over the course of 1000 “seasons” (380000 games), how often would we expect to see a profit versus a loss? A simulation was carried out to determine this. A profit was returned in 470 of the 1000 seasons. In comparison to the betting strategies above, we would outperfom the random forest algorithm length 823 times out of the 1000 seasons and the favoruite betting strategy 930 times out of the 1000 seasons. However given that our machine learning algorithm and favourite betting strategy occured for only one season we do not know how these stratagies would perform long term.

Conclusion