>be you
>think about investments constantly
>no life
>be wrong and get rekt
>live psychological roller-coaster
>be me
>let ml algorithm think for me
>live my live without think about investments
>algo's right 80% of time
>never miss more than 60% of good trades to make
You just never learn...
Other urls found in this thread:
en.wikipedia.org
github.com
github.com
github.com
github.com
twitter.com
Make discord fren
What does it trade?
How long have you been studying ML?
nope, fren
I am not instagram influencer
>How long have you been studying ML?
2 month at good pace with lots of book
>What does it trade ?
BTCUSD scalp at even tp/sl distance from entry price
Also, I have not tested extensively the algorithm, it's just cross validation scores from the training test, not the actual test. But its classifies quite well I'm surprised.
When you have a lot of features (indicators), Ridge Regressors seems to outperform anything.
I'm testing decision tree RN.
Pic related is my homemade trading bot, I have not run it in a while.
Actually someone paid me to do it, but back in the days I had not extensive time to work on algorithms, I was just sprinting on the framework and ui dev.
What does the architecture look like? RNN? LSTM?
What does the test data look like? Has to be more than just labels and data right? Doesn't the bot have to choose when to sell/buy in real time? What does this "precision" metric actually mean?
where do I start to learn this?
t. electrician
If you don't know maths it will sometimes be difficult, but I read:
Hands On Machine Learning by someone at Google.
Use it with either google collaboratory or the new Jupyter Lab, and read scikit-learn documentation and user manual as much as you can.
>What does the architecture look like? RNN? LSTM?
God, it's machine learning, not deep learning.
Deep Learning is very costly in processing power, and it would often overfit. None of this memetic shit is required for such serious problems. Here I'm just using Ridge Polynomial Regressor (Linear Regressor witch minimises the sum of features total weight sum, and that extends features with polynomial features module). Also data is scaled.
I have put 77 features in. But with poly augmentation (2nd degree), it makes many thousands.
ML is based on retrofitting so if you use retro data it'll always perform above average. Come back when can sustain 1 year of real gains.
also come back when we're not in a ridiculous bull market where any buy & holder is making much more than a trader or a trading algo (who do you think made more money, those who held TSLA or those who've been trading it up and down?)
As I expected, decision tree are based as well.
CS fag here,
where do I get good on this? Professor is just teaching us useless javascript shit that I will never use. I get intimidated by the thought of ML and how to test the trade bot (since I have very limited capital rn).
What are your algos and how did you approach the problem? I would try and feed it data based off indicators and choose the best position
>ML is based on retrofitting so if you use retro data it'll always perform above average. Come back when can sustain 1 year of real gains.
You're making a real point there. Maybe it will fail in real time.
>What does this "precision" metric actually mean?
en.wikipedia.org
Fuck I interpreted it wrong, it's not based at all it's fucking imprecise. I hope I will have better results with random forest.
If you want I can share the notebook from where I genrated the dataset on Github.
I first parsed the data, aggregated in bigger candles, removed the parts where there were holes, computed indicators, and generated labels for different tp/sl setups.
sure! Would be interesting to see. I haven't taken an ML course yet so would be cool to dip my toes in
>What does this "precision" metric actually mean?
Let me elaborate my question.
Does a "success" mean you bought and then sold at a higher price? Or is it trying to predict a price X seconds/minutes into the future? What is the bot actually doing that can be considered a success or failure?
There is 3 parameters: tp_ratio, sl_ratio, and max_distance.
If from candle X, we hit take profit target before hitting stop loss target or reach X+max_distance, it's a success.
Here's the code to generate the labels.
I did it for a lots of parameters, but having even tp and sl in % seems to be nice since it's easy to understand the results without the bias of having 90% of success and 10% of failures.
You could always trade "on paper" before committing a bot to an exchange. Some exchanges let you trade with play money before using real money.
Does X+max_distance just mean max_distance amout of time passes without hitting either stop or take_profit?
Why would that be a success?
If Im understanding this correctly, I would be careful that your 80+% success rate isn't majorly inflated by the price simply not moving enough to make any difference in your profits.
The test I have made revealed it was not due to the price not moving.
And the poor performances close to 50% of the others Linear Regressors or SVM seems to indicate it's not the case.
In my tests, algo has 2 weeks to do 2%, it's far from enought.
Also I checked with graph (pic related) all the trading labels to see if they were ok so I would have seen (the pic related is not the same label parameters as the OP results).
What is the actual output of the bot?
Something like:
Buy Qty (min 0)
Stop Gap Price
Take Profits Price
?
You mentioned 72 features, did you manually code these in, or were they discovered during the learning process? I'm not 100% keen on when ML ends and DeepL starts in this regard
Also, what is the input of the system? Every candle (a candle is what, 4 data points?) over a certain period of time? What's that period?
>did you manually code these in, or were they discovered during the learning process?
Check in the link for details about indicators. You have the entire source codes.
>sure! Would be interesting to see. I haven't taken an ML course yet so would be cool to dip my toes in
Here it is, for the ML details, just read the notebooks from this repo:
github.com
github.com
see here
github.com
Nice, you can even real the entire notebook from github with source code and graphs !
>what is the input of the system
All the indicators are generated and studied here:
github.com
This is an obvious bullshit. Trading cannot be automated in principle (since market is evolved faster than your fucking model which always reflect the past state). Otherwise wallstreet offices will be empty. However, that is not the case.
Ignore this bullshit.
/thread
Stop thinking about it. Literally all you have to do.
Buy some shit. Do *NOTHING* for one year. If you become euphoric, sell 50% and then wait 3-6 months until you find a new thing. Repeat.
Start with EPIC under a $500k cap.
>I am butthurt, the post.
Lots of people are doing automated trading bots, user. I'm not saying I have the perfect one, I'm merely discussing classification tasks and datasets with other anons (though OP was bit on a bait tone you jumped right in).
If it can calm you down, I personally know from the guys who paid me people who are succesfully running automated trading bots.
You can also find lots of book of people who opened their algo trading firms.
>Stop thinking about it. Literally all you have to do.
True if you have proper capital and do not care loosing a few k$
Still requires analysis though.
how much money does that thing make you?
OK, lets talk pseudo-smart (because you are really pseudo) - applying ML classificators to a partially observable stochastic environment is no differ from fucking astrology and the results are "no worse than random".
Learn something other than python lmao
>only right 80% of the time
>not right literally every day of every year, 100% of the time for an easy quintillion times your money
Though in all seriousness, if you think you have something which is right 80% of the time after learning ML for 2 months you have another thing coming.
Put your money and trade it live, I dare you. pussa