ScorePredict

#DEVLOG 7#
Today I made the frontend sites with sample data.
It’s not the most beautiful website I’ve ever seen, but this will be enough for now. I’ll continue working on it, when I’ll have got a working backend.

What to do now?
Getting future matches’ data, write some code that puts together all of the neccesary information, etc.

#DEVLOG 6#
Check this out!
I quickly made a demo how the UI will look like!
(I didn’t add text, because it is not so easy to use this website😅)

Left side: It’s the home page, there will be the future matches shown. When clicking, it navigates to the match’s page, which is on the right side.
Right side: Match site. There is the xG (expected goals) and the outcome’s likelihood (home/draw/away).

What do you think? What would you do differently?

#DEVLOG 5#
I worked on adding another aspect to my AI model. I tried to somehow quantify a team’s (before-match) motivation. I successfully added this to the model, but it just made it worse.
I also recently found a bug, I used future data while calculating the strength of attacking and defending.
At this point:

1. I’ve tried almost everything to make my model more accurate
2. I lost my motivation to work more on the backend

SOOO, I’ll finally begin working on the frontend🥳🥳

#DEVLOG 4#
Okay, guys, I’ve finally found the cause of the issue.
It was not in the code, it was in my mind.🤦‍♂️If I wanted the model to predict a draw, it would lose by accuracy.
To understand this I made some calculations:

away: 7341  matches  - 31,6%
draw: 5522  matches  - 23,8%
home: 10368 matches  - 44,6%

SUM: 23231 matches

In a nutshell: It NEVER worth it to predict a draw!

Now that I “solved” this problem, I can focus on increasing accuracy: adding more aspects, finding bugs, etc.

#DEVLOG 3#
This is just a quick devlog from me.
I have some issues. The main thing is what I have already written about: the model NEVER EVER predicts a draw. (It is because of the mechanism of Poisson-model…). I tried different methods to solve this problem, but none of them worked.
There is 2 ways to solve this:

1. Get a bigger dataset. But the API that I’m using provides only 100 calls/day which is a really small number. So I would need to pay, which I don’t really want. (In the very unexpected case you want to fund my project, DM me @be.the.nike)
2. There is a model that corrigates the torsion of Poisson-model. BUT AI said to me that to achive this I will have to kind of write my own AI model training algorithm. And this sounds really scary to me

Cheer for me!

Some good news at the end: my model is a tiny little bit better in predicting the winner of a match than betting offices!

#DEVLOG 2#

I haven’t posted devlog for a while. But I was working hard.
I ran the secind evaluation of the Poisson regression model (the first model was so unreliable that it don’t desire to mention).

Results:

• Outcome (H/D/A) accuracy: 54.1% ± 2.1%
• Goal-count prediction (Poisson deviance): 1.13
• Feature importances look sensible: team_attack, opp_defense, is_home are the strongest signals
• Weakness: model almost never predicts draws I tried a “draw margin” trick, but it didn’t work

Next options (ideas only, none started yet):

• Add Elo rating (instead of current attack/defense strength method)
• Dixon-Coles correction to address draw underestimation
• Incorporate missing key players as a feature (it costs too many API calls)
• Tune alpha hyperparameter

This is my first project written in Python so I’m still learning pretty much. I’ll push my progress to Github as soon as I can produce accurate enough predictions. Be patient, I’m doing my best!

Arrivederci, Auf Wiedersehen, Goodbye, A viszontlátásra!!

#DEVLOG 1#

I started to work on a website that predicts the scores of a football match. I’ll make my own AI model to create it.
To do so my first task is to get a database. There are a lot of APIs on the internet but just a few of them provides a service without any limitations. I’ve finally find one on Github. So I started to work on it and format the data. BUT after ~2 hours I realized, it wouldn’t work.
The reason why it wouldn’t work: I wrote some criteria that this AI model should consider. And the Github repo did not fit to it.
Sooo I’ll start over with a new provider.
By the way, here are the aspects I was talking about: