I made an AI that can PREDICT your future RANK in Rocket League. Here's how I did it.

[@Bsjbsjgsjb]

I'm a ML developer in finance and I'm impressed that someone who sounds so young has been able to put something like this together. If you continue to pursue coding and statistics, you will be able to do great things. These skills can be applied to many things in the real world if you know what problems to look for. I hope that you continue this journey so that you can help make our country greater. What you made is a great start but there are some improvements that can be made. The most valuable information that can be extracted from this is the coefficients of the linear regression model. This will tell you how much the MMR is predicted to change based on the input variables. That way you know which variables positively correlate with MMR gain. The model summary will also include P-values, which will indicate the reliability of the predictive power of each input variable. The model appears to only use time as an input variable. This means that the prediction that is being made is MMR gained per day that passes. You need to be extremely mindful of what the model is predicting and what the input variables are. Adding more input variables that are relevant such as games_played_in_season might dramatically increase the predictive power of your code. Keep in mind this will require you to set an intended_games_to_play for the final prediction. Also, look into automating your input data collection. It's easier than you might expect. In addition, it may improve your model's predictive power if you exclude or otherwise handle periods of inactivity. These inactive days will mess with the weights of the statistical model. There are many ways to handle this. You could have a variable for days_played_in_rolling_week and/or month. You could also do days_played_consecutively or days_of_consecutive_MMR_drop or increase. You can infer when the player played by how the MMR changes. Letting the model 'know' about periods of inactivity or the player's history of habits may increase accuracy. These are changes in the input variables would allow you to extract more information from your data. That way you won't need to look for more raw input data. However you need to be very mindful of how you code what these input variables are actually measuring. The code results in a prediction that shows extreme growth in MMR a year from now because the model is trained only on the current progression rate of a beginner who is making good progress in improvement. This is like capturing the slope of a hill and assuming it goes to space. This is an inherit problem of machine learning models that interact with time. The predictions made close to the present will be decently accurate but not in the distant future. One potential fix could be to have only trained on a certain lookback_period worth of data. This will allow you to also do backtesting of the model's performance by using many models that are made every model_creation_interval. The issue with this fix is there would be less training data and it might be better to just create a graph of predicted distribution of outcomes. It would also be simpler to assume that predictions made far in the future are very unreliable. I see that your code is written by ChatGPT. I can tell just by looking at it. There's no shame in using ChatGPT to code, and I strongly believe it's the future of coding. The issue is ChatGPT has some weird habits. One of its habits is that it will usually make a train/test split of the data when you try to make statistical models from scratch. In some cases this is ideal for testing, but you usually want a testing method that is custom made for your use case. The train/test split may also be reducing the amount of data available to your model. I also encourage you to use ChatGPT to scrutinize your code, thinking, and strategies. It is extremely good at that if you tell it your code, plan, line of thinking, goal, and concerns. ChatGPT is also good at teaching statistics and coding. Tell it what you don't understand and repeat your understanding back to it. Also look to see if graphs of distributions would be viable for your use case. They give a much better indication of what could happen. Don't worry about your model being 100% accurate. The world is random. Even with all the data in the world, you can never be 100% accurate. I'm seeing some blatant misinformation in this comment section. I don't want you to be led astray by these brainrot Fortnite kids. Your mind is an organic machine learning model. If you put in garbage, you get out garbage "This has no predictive value" This does have predictive value, even as it is. The exact prediction being made is MMR change per day for the specific player that is being analyzed. The days passing correlates with your MMR increasing. The linear model also reveals the slope of the increase during the training data's time period. The issue is that the prediction is made purely on the MMR slope and the passage of time. In finance we call this a naïve prediction. The predictive power can be improved with more relevant input variables. "You need to know the player's gameplay behavior" You don't need this information at all. It might help to have a season of ballchasing com info on the player, but it 100% not necessary for your AI to work. All you need are relevant and correlated input variables. This may go against what you would expect, but the behavior of a player's gameplay might not be as useful as you might think. Let's say more goals correlate with MMR gain. You can't just choose to score more goals. It would be more practically useful to know things like how your playing habits- such as consecutive playing days- affect MMR change. These statistical models are most useful when you can make actionable change to optimize for the outcome you want. "You need to know the opponent's/teammate's gameplay behavior" You don't need this information. It would actually probably decrease the predictive power of the statistical model. Your MMR is what it is because of you. It's your journey. If you are improving significantly, you will rank up if you play ranked. You're not stuck in plat because of your teammates or smurfs. "This isn't AI, this is math!" Historically, AI has been defined as automated decision making. More recently it has included statistical prediction making. This is statistical model used for making predictions. This is AI. ChatGPT is a statistical model that predicts the next letter/word. You'd be amazed how the math can describe and predict things that we don't consider math. I've realized that I've accidentally written a book. Oops. I hope this comment is helpful for you.

[@chillhn]

Dude this is amazing. First of all thank you for taking the time to write all of this lol. Secondly your feedback is extremely helpful. I’m only a freshman in college and took on this project while I was bored at my on campus job (basically I just have to sit around for like 7 hours). We had just started learning about linear regression and ai in my Python class and as soon as I left the lecture hall the idea came to me. It was never meant to be something so serious so that’s primarily why I turned to Copilot for help, but now that I’m looking at it, if I take the time to really understand how to take this further it could turn into something remarkable. I can’t thank you enough, this comment made my day.

[@skecpg]

The cool part about this is that although it looks convincing, it has absolutely no value. This is similar to course examples of models that predict stock value. it looks legit because it shows trends and variations, that makes sense compared to training data, but the thing is, it's only good at that, (constructing values based on values). the reality is that just like for stock, a player's MMR won't vary according to the players past values, it will vary because of how the player is doing compared to his opponents during games (and that data is not accessible to you because its in the future...). Cool exercise though, I've done a similar one while learning ML.

[@chillhn]

Yea, of course it is all theoretical data but that is very true. What does on in game is what really matters, but it was cool to get a theoretical visual of future mmr based on past data.

[@Otiyyy]

Yes pls make this public

[@pidojaspdpaidipashdisao572]

I don't need AI to know that I will be a diamond forever.

[@chillhn]

Nah gotta be positive bro 💀

[@MrBubbyG_Official]

I think you need to have 50 games of ranked replay data uploaded, then compared to the ranks above you, and based on the exact data on how your car moves around the field, boost management, positioning, etc, you could give a rough estimate. Still not totally accurate, but if it seems like you play the same way the rank above you does, but you are in a lower rank, then it indicates that you may be good enough for the next one

[@chillhn]

That’s a solid approach, I like it.

[@TrueBangers]

i have not understood a single grain of sand, but i must agree that this is cool.

[@JonLake]

The real prediction is you'll still be diamond a year from now 😂

[@chillhn]

nah fr 😭

[@TheFormerTeam]

1500 is gc1 and dang if I gained rating that fast I'd be so happy. I've been gc1 since old season 13 and I'm just now breaking into gc2, currently at 1552 in 2v2. I do take long breaks and I don't play consistently that much so thats probably why I'm still where I'm at. Ganna try and grind the game more in 2025 see if I can get gc2 in 1s.

[@Rough.]

Who do you use for coaching?

[@jakobskouv]

This has nothing to do with AI tho. It's just math...

[@chillhn]

@@jakobskouv no it does. If you look closely there are lines of code where you can see training and testing variables. This is training and testing the ai to get more of an accurate prediction. (Although it’s not even that accurate anyways lol) The imported libraries from sklearn deal with machine learning.

[@dahuday]

1. 1500 is mid gc1 2. the highest mmr anyone's gotten is around 2700 I believe 3. do you plan to release this?

[@chillhn]

Oh yea I think I mixed that up with 1s mmr. To answer your question - short answer: yes. But I’m going to need to make it more streamlined and easier to use before I release it fully.

[@itsjndo4574]

here before every single youtuber use it

[@marvemes]

As a Rocket League Coach who Coached players from Champ to SSL i can say that this Graph does not look realistic at all. Your algorithm probably takes data of the curve from other players, the problem with this is that most players around Champ 1/2 get tilted after deranking and get their highrank friends to boost them. Wich is a more Common thing then people might think. According to your Graph u go from 900 MMR to 1450 MMR within 2 months, wich is a 550 MMR differenz. So basically from Having a hard time Hitting the Ball Consistently to being able to Make Quick decisions with consistent touches wich 90% of the GC players took years to Learn, in 2 Months. Even with a Pro eSport Coach (with which i worked with alot), this will be tough especially if a player is Hardstuck in Diamond, who builds up a lot of Muscle Memory and bad Habbits grinding Games in this Elo that much. I dont say its Impossible, to reach C3 untill May, but if a Tracker looks like this, its much more Likely that a Player is Boosted instead of being a legit GC. if there wouldnt be a hard drop from 1170 to 900 from January to April then it would make much more sense even tho i think going from 1170 to 1450 ist still pretty tough.

[@chillhn]

No I definitely agree with you. Like I said in the video though it’s not 100% accurate because it uses a linear regression model. Meaning no matter what the ai is going to predict improvement at a positive constant rate, regardless of how much noise or variation there is. The ai uses my past data specifically, no one else’s. That's why the prediction looks so similar to my past data/it follows almost the exact same trend. Also like I said in the video, the code is extremely simple for something of this magnitude, so accuracy is hindered at the cost of simple code. To be more accurate is my goal for improvement with this ai, which would require me to get more data from other sources. I'm not sure which graph ur looking at but for the graph at 6:55 it predicts my rank to go from 900 MMR to 1450 MMR in around 7 months, not 2. But I completely agree with you, this is a good point.

[@ZxnicRL]

i want to download this

[@mellowhub.]

This is cool asf!

[@FootClob]

we can all see the chatgt comments hahahha (im a proffesional and use chatgpt everywhere) but good job

[@chillhn]

@@FootClob yea xD

[@wesupportfrance]

yoo W please more about it

[@blackeye2312]

9:56 My man what is this sin(ax) + sin(bx) + cx aah graph 💀

[@chillhn]

Man I just work here ahh graph 💀

[@pikaturno069]

nope 1500 is gc

[@chillhn]

yea I realized that I was thinking 1s mmr for some reason.

[@jaygoddi7680]

yooo hit me up ill contrib if its cool