Thread has been deleted
Last comment
As I told before in this thread hltv.org/forums/threads/2371424/neural-n..
I wanted to continue researching machine learning and csgo predictions with it.
My first tries was simple back propagation neural network, it didnt work, very bad winrate.
Then i was trying another stuff and didnt found anything interesting..
Before today...
I tried one technology (dont wanna tell right now) that uses some very smart stuff to do predictions.
There are 10k matches marked with atleast 1 star on hltv. I used 5250 of them (tomorrow gonna download more). I trained my model with 5000 matches, it took me just 30 minutes on "Google colab". Then I tested my model on 250 matches that I didnt put in train... and model predicted 234/250 of them correctly..
I still can't believe this, going to download 5k more matches tomorrow, retrain my model and check again...
Then i will need several days to test it on practice and if it will work i will be so happy
Good luck man
Sounds interesting
6 replies
5 replies
Just don't train on test data not to disappoint yourself lol
3 replies
1 reply
good))
Awesome!
1 reply
You're webscarping data and analysing statistics or download demos?
Easy money goes just as easily, if there's no emotional attachment to it you cannot hold it and it will perish rather quickly. A moment's joy at best.
1 reply
-1
That's bullshit!
Good luck, but as a pessimist and someone experienced in both gambling and ML, I think it is very tough. As I understand, all the bookmakers purchase their CS odds from a very small number of sources whose whole business model is employing multiple engineers that use ML, fed with a lot more comprehensive stats that aren't available to us on HLTV (it's also worth being wary of data provided for free by a gambling-driven site like this). So you are up against tough competition, especially when you add the bookies' juice advantage.
That said, there are hypothetical reasons you could come up with a more accurate algorithm, e.g. you've adapted some recent paper, vs them maybe having organizational/bureaucratic bottlenecks slowing them down from updating their stuff constantly to the new state of the art
Good m8, best of luck with it.
Did you use Python for machine learning?
9 replies
8 replies
5 replies
4 replies
3 replies
2 replies
I minored in Applied AI and also worked a lot with ML. I'm willing to bet you're predicting the matches based on a statistic like "won_rounds", "total_rounds" or something to that extent (everyone can predict which team wins when you know the rounds). I know because I've done the same research a while back with all top 30 pro games played from 2015-2020. I was able to accurately predict a matches outcome with just the average player rating and KDR and got upto a ~73% success rate with a RandomForestClassifier (DecisionTree ensemble) and a LogisticRegression model, both yielded similar results with I think the RFC just edging the LR out. Can you let me know which stats and models you are using?
9 replies
1 reply
I'm just saying that a 95% accurate model is hardly possible with a predictive dataset let alone a pseudo-random dataset like CS:GO matches. I've done similar research during my minor so that's why I'm questioning the results. Also, he doesn't show us any models or any data he uses to base our assumptions off so I take his results with a grain of salt, no offense to him. If he manages to back up his research and ML model, I'll believe it.
6 replies
Interesting, I hadn't tried head to head stats yet. Are the head to head stats the stats when that match happened or the total head to head? And are they lineup dependent or not?
5 replies
3 replies
Did you calculate head to head yourself or did you get it from HLTV in some way?
2 replies
1 reply
Oh okay, those are lineup dependent I think. Seems good man.
If you go on the match page of any given HLTV game, at the bottom are HtH-stats, which are lineuo based. For example, a game of fnatic vs Virtus.Pro wouldn't show the games from 2015, rather it would just show the games of the Ex-AVANGAR-lineup vs this current Fnatic team. It all works with the same 3-man-core system as the rankings do. So, they should be up-to-date.
Sounds interesting please update us about your projet
if it rly works u would be millionaire from betting
Very nice
woah gj men maybe you can make money with this later
Sounds very cool!
I would say if it works go bet haha
But keep in mind that if you download all matches from the past 7 years you might face a problem.
The % will be useless if the teams changed their lineup. For example Virtus Pro, they were amazing years ago and they won a lot against other teams. The scores from the past wont match up with the current state of the team though.
Maybe you should just focus on scores/game which were in the past year so you can avoid that.
1 reply
Very awesome! I'm in software myself, so I'm crazy about this stuff.
What are you basing the victory odds on? Teams? Players? I feel like both of these metrics would be skewed. For example, there are loads of games that FaZe was dominating in 2018, so the AI might predict FaZe winning a game in 2020 which is more unlikely. Also players, in 2015 you'd be pretty safe assuming olofmeister would win any given game, but he had a role shift in FaZe from star player to support-player, and his stats dropped remarkably.
I would understand you wouldn't wanna release the entire thing but I'm just wondering what input-metrics you're using.
6 replies
5 replies
4 replies
3 replies
1 reply
2 replies
Model was trained on matches ranked with 1+ star on hltv.
I just tried to see how much winrate is on 0stars matches. It was same as on tests before: 90%
Model saw all these 0star matches first time, i did not train it on them, and it gave 90% winrate...
what if it really does predictions with this winrate? i still cant believe myself even if i rechecked everything like 4 times alraedy..
Interesing
1 reply
Thanks!
Hardest thing is to collect all data about past matches. Training model is easy part because you just import some library and call functions.
Just try to find good site with alot of stats like HLTV but for soccer. Then you will need to look at the html code of pages with stats for each match and see what do you need to collect from there. Then just do request, get html response, find there needed pattern and collect data!
Every site require its own way to scrap data so i can't tell you 100% method
Hi _g00s3_. It is good that you can forecast it well. But your challenge is to forecast it better than forecasters since you want good performance in a market for predictions. You should try to calculate the implicit probabilities within multiples paid by betting sites before the start of the match. Then sample a decision-maker that bets drawing from these probabilities and compare its forecasting ability with yours (do this enough times to be confident you are better than the average market bettor...). Otherwise, you can let your strategy bet on these multiples and backtest it.
Hello man.
The last thing you should do is to post sources and/or tutorials how to achieve your results (if you're not lying).
Use it for your own profit. Bet by yourself or try to sell the data to the bettors (e-sport prognoses), or betting companies (analytics).
If your claims about winrate are true, you can make a shitload of money. You shouldn't share this secret with anyone.
Now playing
FISSURE
5619 viewers
Top streams
All(24)
Casters(11)
Streamers(5)
Organizers(8)
FISSURE
(5619)
ESL TV
(2024)
Elisa Esports
(1406)
BetBoom
(1268)
ESL TV
(1188)
EU Pro League
(452)
juho
(435)
the real fer
(339)
n0rb3r7
(250)
mch
(244)
CCT
(214)
EU Pro League
(213)
bt0
(202)
wonderbai
(197)
eSports BLINK
(184)
Bubzkji
(63)
mantuu
(51)
leszklareN
(36)
gaules TV
(17)
REPUBLEAGUE A
(16)
Polen
(11)
BTSBrasilFPS
(3)
CCT (YouTube)
(3)
BTSBrasilCSGO
(2)
