Nice post indeed, Mher!

I have studied your code a while and I have some comments.

Sorry if I misunderstood your code but here it goes...

As I ran your code I suspected the H2H calculation was not according the description you've provided.

H2H calculations should only take into account the previous games as one should expected and you said as well.

Your code gives us an single H2H value for every two teams confrontation.

That means H2H would be a wonderful predictor indeed since it contains information about all future confrontations of any two teams!

For a single game H2H calculation didn't take the date as a reference.

So the calculation takes all team1 against team2 and team2 against team1 games...

66% accuracy for classifier drops to something around 57%, as I have made the following changes:

headToHeadTotal[t1_, t2_, dt_] :=
 N[Mean[
   Join[
    Normal@
     outcomeAdded[
      Select[#Team1 == t1 && #Team2 == t2 && #MatchDate < dt &], 
      "Outcome"], -Normal@
      outcomeAdded[
       Select[#Team2 == t1 && #Team1 == t2 && #MatchDate < dt &], 
       "Outcome"]
    ]
   ]]

And:

h2hAdded = 
  outcomeAdded[All, 
   Append[#, 
     "H2H" -> (headToHeadTotal[#Team1, #Team2, #MatchDate])] &];

So the result is:

Would you please confirm my interpretation?

I have not built the predict of outcome yet so I do not have final numbers to show...

Please see a followup discussion here: http://community.wolfram.com/groups/-/m/t/1218215

Predicting Winning Odds of Italian Serie A Soccer 1934-2017

Thank you very much! Yes, I've used Python framework - Scrapy for obtaining the data for this project. It can be used for both small projects and also for building complex spiders. There are a lot of tools for web-crawling, but the ones that I find effective besides Scrapy, are Beautiful Soup, which is easy python library for pulling data from HTML/XML pages, and Selenium WebDriver, which provides Python API to use its powerful functionality. It was originally written for testing web pages, but is widely used for data crawling, especially in those cases where there are complex captchas and other barriers of accessing data.