Predicting the Winning Team with Machine Learning 

Thumbs up if you guys want a part 2 of this video (using deep learning + twitter sentiments)!

I just started learning python and this channel is great to see what I can do with the knowledge! Keep up the great work Siraj!

I'm an I.T. Architect, not a Data Scientist. You have no idea how quickly I am learning what I need to from you after jumping into this field. Keep up the amazing work. This isn't easy stuff, but I love it. Thank you for all you do.

thank you so much for posting such marvelous quality content out there on youtube so frequently! i'm lovin it! i'd really appreciate further and more in depth uploads to this topic (and all the others)! keep up the great work! greetings from switzerland :-)

Thanks so much for this video - learned more from this than any other football / modeling video on youtube. cheers!

This is one of my favourite videos from the longer 30 minute videos. The problem is easy to understand, the data was well described, nice simple live coding. Great job, looking forward to more videos related to sports.

This topic is awesome!! And it's a good example on what you can do with ML! :) Nice job!

Love your videos! Topic, explanation, and production. Thank you very much!

Really liked you share the preprocessing process. Great :D

I prefer using a time machine

This is my thing. Thank you Siraj for posting this 🙌🙌🙌

OMG thank you Siraj , I asked you to do a video about this and you did it , you truly deserve a medal Sir. Thank you ! :)

Thanks for the great video. Just one important point: you can not compare the algorithms by some default parameters and then pick the "best" one based on this naive experiment and then work on the winner's parameters to boost the results. In this video, you simply chose the XGBoost because, in the first experiment, its accuracy was higher than the others whereas the SVM is very sensitive to its parameters and one should spend a good amount of time on parameter selection for SVM before going through any conclusion.

This is one of the topics that always intrigued me. Please do make further videos related to Sports Analytics :) and thanks for this one by the way :)

Great content as always. Thank you!! Would love too see more soccer ml takes.

Hey siraj, You don't know me but I feel like you are my greatest friend..thank you so much for these videos even though I am beginner at ML but I watch each and every one of your videos.

Hey Siraj! Great Video Bro. Guess I am a bit late to the party. Would love to see more videos on Football Analytics. Keep Rocking!!

I have had this idea in my mind for some days now and now you do a video about it. Spooky.

Hello Siraj! I second Igor's opinion about splitting the data according to time and not randomly. Having said that, I thoroughly enjoyed this video. I would love to see another video on the same topic which includes sentiment analysis from twitter feeds as well on top of the baseline models. Perhaps we can use LSTM with a time distributed dense layer to capture the time information?

Hi, this is a great vide. Everything is well explaned. I see people are asking about deep learning approach. My suggestion is using a recurrent neutral networks such as LSTM. Also, for sentiment analysis, we can use a pre trained network as feature extraction layer. It can be pre trained separately on a Twitter datasets as warmup. In more detail, you can use another recurrent neural network with an embedding layer at the begining. For embedding layer, again, you can use pretrained word2vec models such as FastText, BERT, etc.

This is a very interesting topic! I'd be interested in seeing a further sports analytics video, like you mentioned at around 14:25 mins

Very nice and well explained! Definately interested in ML for sports bets!

That's brilliant to hear from you brother

28:35 5, I´m not sure that's accurate Siraj. Anyhow, great video as always, keep up the good work champ!

Please continue, this is most exciting topic I can imagine.

Hi Siraj, Interesting I'll been creating a machine learning model based on your random forest video which will predict players most likely to receive > 7 points based on recent form on the game Fantasy Premier League, which takes into account some of the features you talk about today but on a player level. Glad to see you doing a premier league machine learning video!

Im waiting so long for such a vidieo. Siraj please do more of them. Orange3 is a nice tool to quickly compare strategies. Could you also make a video about orange3. That would be nice. THX a lot ;)

Thanks Siraj, we spoke at Google Amsterdam, and I asked you this question (related to a different subject that needed multi-class classification). I want you to know that I got a pretty good accuracy now. Thanks!

This video is so informative, I hope you can make more videos like this.

always informative and fun... thank you, Siraj

Where do i get the same dataset ? Because my dataset dont have ['HTGD','ATGD','HTP','ATP','DiffLP'] and i have a lot more features then 12

Great video Siraj!

Hi Siraj, great video, ...inspiring.!! Just one question, ..after using Scatter_Matrix, we found out there are 'correlated features' ,,,,shouldn't we eliminate those correlated ones ? ..

This video increased my interest in machine learning

is there a way to train the model on the full data set, with variables that we could only know for a past game, and then predict off the model using only known variables for a future game?

Great one..Please provide a video using deep learning for prediction.

Great video. For the next update you can replace 1 X 2 model to handicaping. I think this is the more popular way of betting.

Greatest Video I ve ever seen !

Awesome vid, would love to see more!

Loved it!

You've mad a one big mistake Siraj in 00:19, IT'S ABSOLUTELY LALIGA !! Great video (y)

I think this was one of your best videos. It was really easy to understand. But can you make sure you're typing on the screen? 15:53.

You just gained a new sub!

can you please, upload the video showing overall process of creating models of various variables choosing anyone prediction algorithm sir!!!

Thanks Siraj! Really good example! ⚽ However your scaling process is wrong. When you scale the data you need to "fit" your scaler only on the training set and then use it on the test set to avoid information leaking from test to train. You can also take advantage of the temporal information (as some other user pointed out) but I can understand that you didn't do it to keep it simple. Love your videos

Really great man

After 4 long months of requesting for this , You made it. Can't be more happy. You've got my money mate

This video is just awesome, I don't know what else I should say about it 😜

Everything is perfect in this video, the way things have explained and coded was amazing. But how about the end result.. can we get a table of variable with "H" or "A" win prediction that would be good and great.

Get more in depth with this please, maybe you could talk about voting classifiers or ensembles? Or get more in depth with the feature engineering? I've made odds prediction algorithms as well via ML, really interesting stuff! Don't tell the crowd they can make money with this though, the algorithms from betting companies are way better than anything you can make in just a couple of hours, even days.

Siraj, SVM tries to MAXimize the margin between hyperplane and closest points of different classes - not "MINinize" as you said. It was correctly written on your slide but you said the opposite. (22:53-23:15). Thank you for the video anyway. Very usefull.

@Igor agree with you. Along with this, there is problem of Data Leaking. Before match does start, how can one know no. of corner kicks, on shots, red cards etc. Very popular data mining problem. Great video otherwise BTW @Siraj Raval. Love from India :)

thank you for scikit-learn video

this was brilliant! can you do an advanced video for this topic ?

Very interesting, more of this pls

You're doing great!

Please do a video about deep learning on sports :)!! Thanks!

Where did you get these variables: 'HTGD', 'ATGD', 'HTP', 'ATP', 'DiffFormPTS', 'DiffLP'? They are not in the csv file.

Amazing, you are a genious Siraj, What if you have a system to capture cast data before the game starts, so know if any important player will play what can affect the outcome.

This was Awesome. Can you make one using Deep Learning ? Will be really helpful.

@Siraj Raval : Question, is your dataset randomly ordered? because, if you split your dataset randomly --> your training set will contain pre-knowledge (and that is something which isn't allowed ). To do it correctly, you should keep the chronological order of all the events. (in the training, validation & test sets). This means that you've to ensure that all the events of your training set, take place before an actual event of your test set [ sliding window might help ]. - It might sound confusing but, if you've trained your model with data from sept-2016 until February 2017 and from Match 2017 until June 2017 and you test your model with the data from February 2017 until march 2017. Chances are high that you've some good results. Example: you can use A B C D E as a training set for predicting game F. but not B D H E G for predicting F because then you've used some events which never happened in that timeline. Or an better example, when you've a Support vector machine which is trained on C D E G H I and your test value is F --> then the hyperplane of your support vector is pretty biased :) (even if you don't think about the previous 1-2-3 games, etc) :-)

As always great! video Siraj!!!!! I just think there are two problems with your approach: First: you cannot split data randomly using train_test_split function because you have a temporal information. For example: a team may be having either a winning streak or losing streak and this will be lost if u mess data. This is called lookahed bias. Second: u are using stats of a game to predict the result of the same game. What's the problem of that? It is really not a problem unless u really want to bet on the game because these stats are released only at the end of the game (too late for betting). I think a better approach is to use features from previous games. Like how the home team behaved in its last home game and how the away team did in its last away game.

Please make more videos on this topic Sir! :)

Hey Siraj, were the features you used to train the model the stats of the game you predicted? Or, were they stats from previous games that already happened and you used them to predict a game in the future?

It was great, Make more vidz on this topic.

This is great. Am now fully committed to learn machine learning next year! thanks for the great work Suraj!🙌👌✔

Great video, it's nice to see the model comparisons. Btw there are 11 players in each team in premier league (until people are sent off or injured)

Awsome, when you say sentiment analysis of twitter, does that mean the positive or negative sentiment from twitter is nothing but one more feature adding into the existing data. Looking forward to see full game analytics and stock market prediction(with twitter sentiment analysis integrated) videos, You are awsome...

hi, very good job..i have not had chat with u long time. good to c u ur doing good ,

Hi siraj, Can u make a bunch of video on data exploration and data preparation for different algo?

SIRAJ, english subs in MoI #7 is finished (by Chris), and also MoI #6 in spanish (: greetings

this video will be so popular I can see

Great video Siraj ! But I have a question, Where can I see the predict matches ?

i like both football and machine learning😂😂 and much more than that...i like your videos which are truly awesome and knowledge. And also i have a question in my mind related to probabilistic approach of computer science and i wanted to tell you. i am very excited. how can i do so??😃😃😊😊

When I was 18yo I built a football pools forecasting program which was 30% better than chance - but not good enough to overcome the percentage of the pools money that was taken out in costs instead of being fed back as prizes. It really needed to be 50% better than chance to break even. But then I hadn't learnt about optimisation, it was just a formula I made up. However, I had different formulae for each type of result. So I had one for Score draws, one for home wins and one for away wins and I would pick the highest ranked matches from each list and they were my selections on the pools coupon. I think you'll find that better results can be had by using each model to only achieve one result (score draw or not score draw, home win or not home win, away win or not away win) rather than one model to try to classify each type of result.

Yes please make a deep learning version as well, as I understand it, this would result in automatic feature selection right?

FTR at the "final_dataset.csv" contains values (H / NH) and not (H = home win, D = draw, A = away win). At end of dataset "Scraping and Cleaning.ipynb" it converts FTR variable to H / NH. What makes the solution binary and not multivariate. Compromising this result presented. How do you explain it? @ Siraj Raval

wow great video!!!!

Great video..been working on various football models, you learn the thing is to try and optimize the model to specific niche not seen by the bookie. The goal is to beat the odds, not just correct predictions. Say you predict 80% chance for a team to win,bookies know this too and reflect it in their odds plus sum vig. It seems easy but not quite

Nice tutorial indeed. very informative. What software do you use for making the video siraj? If i may know

hello suraj, im working on the same project which is been shown by you in this tutorial, but i m confused with data cleaning, the video shows you just started with jupirter and then some commands and boom, you just got the required results. But how to do this ? will you elaborate the procedure for this please, would be really glad to know.

Please, talk about XGBoost!! Great video :)

Hi Siraj, could you explain a bit, what "Difference in last years prediction" is and how do you calculate it?

hey Siraj, I'm having a lot of troubles trying to install xgboost for mac. Do you have any guide? I tried the official from xgboost site without having a result.. Cheers

I'm trying to build the same for a tennis game from where I can get the relevant data ? Thankyou.

Hi Siraj, at 27:36 when u show the performance of the best estimator u got out of grid search, the accuracy on your training set is far worse than the one on your testing set. That seems weird, can you explain it? Thanks for your vids, they're great!

can u upload the database you're using? Because i can't see where u got 'HTGD', 'ATGD', 'HTP', 'ATP', 'DiffFormPTS', 'DiffLP', i don't see them in the csv file

Hi siraj, thanks for the great video. do you have another video where i can see hoe the model finds th FTR for a future matches based on what the model is trained of ? Thank you.

You are awesome!

Okay wait , did we talk about predicting future ?, that's so freakin cool

so, can I use the model built ti predict a match if those two teams have never played each other again?

AWESOMENESS SHOULD BE YOUR NAME

Great Video

well done

I think Over under goals and BTTS (both team to score) are more easier to predict. And 1 element , I think the most important is the value to define how important the match for the team. example for the some match when the midtable team that cannot qualify for EL/UCL against another team that need the win the most to escape the relegation. One team probably play more nothing to lose and another play with 200% of their heart and mind

When are you going to do a detailed chatbot tutorial? One with formating the corpus??

Hi Siraj, a little off-topic, but, can you suggest how can I do depth prediction from single monocular images? Thanks!

Okay 1 question though, we say that xg boost has the best accuracy. But what is accuracy here, i mean how does the system determine if the result is accurate or no?

what type of software should i use to code for machine learning ?

This is fun !