I finally trained a profitable strategy with AI, the strategy has at least 70% of WR over every single month in the year with minimum drawdown. Now i'll test how does it performs by trading 15 different pairs at the same time with a max of 10 orders open. Feels good man, IGMI.
I made it, biz bros, i made it
Based mathfag. What did you program this in? Which APIs? Where did you source your backtest data
Lets see current screenshots, not year old data from backtesting
poorfolio or didn't happen
What did you program this in?
python.
Which APIs?
pytorch + stable baselines 3 using the XGBoost Regression model.
Where did you source your backtest data
binance, i downloaded it beforehand for each period and each pair (automatically ofc). My strategy trades 4 timeframes for each pair: 5m (main timeframe), 15m, 1h, 4h.
Lets see current screenshots, not year old data from backtesting
Yeah, now i'm going to do a dry run for 1 month in an AWS EC2 instance since i dont think i can handle hearing my CPU roar 24/7 for a whole month.
I think this is it anon, i backtested the shit out of my algorithm and the lowest win rate was about 68%. Unfortunately, i cant backtest the whole year in 1 run, since each model file is 1Gb and since i want to trade 15 pairs retraining every 7 days: 365(day)/7(day*pair/model)*15(pair) = 782(model)*1(Gb/model) = 782 Gb. So ill definitely need tons of time and a bigger drive.
make one ai bot but with memecoins, i'll like to see wtf does it tells you about simons cat, or bnb in general, im so tired of looking at fucking charts everyday with non funny memes
Training an AI is painful af without a proper setup, anon. Once i make my first 20k with my strategy, i'll buy a proper setup and train all the meme algorithms in the world. Can't wait, i'll finally be free of the wage cage!
1 month in an AWS EC2 instance
How does this net out with your profit
How does this net out with your profit
1 month on the instance i picked is around 200 USD. Really cheap, since i dont really care about training the model fast (i'm done backtesting). Unfortunately, since is a "dry run" it will only simulate trades using real time data, so ill lose 200 USD this month, i guess. But better losing 200 USD now than blowing up 10k with a shitty strat.
are you going to share with the class or not?
Ammm no, not yet, at least.
backtesting always looks good, I wouldn’t get too excited. enter cautiously
sounds cool
good luck
thanks for sharing
WAGMI
This
Its easy to get backtesting look very good but in reality the performance isnt that good
Good luck
I remember running MT4 backtests on a trivial EA who was literally picking trades by flipping a coin, with only one variable: The PRNG seed.
Once I got a seed that backtested some amazing results, I posted them on some forexfactory forums along with a terse explanation, thinking we'd all have a good laugh about it.
But of course we didn't, and I got barraged with folks asking for the precious source code that would produce such amazing results, and with well-meaning people trying to warn me I was making a dire mistake.
I think I liked the latter kind better, but neither were much fun.
782 Gb
1tb SSD is $50 now. Also congratulations on your algorithm for figuring out that the best strategy during a bullrun is to sell after you buy and repeat that.
Your model will not work once we enter the next downturn. Maybe you can use it to find when that is.
backtest
lmao
oh, about that, ill have to make another to detect if we're in a bullrun or not based on sentiment data from social media. All my algorithm does at the moment is to predict local tops and bottoms. But as you say, it would be retarded to try to scalp during a huge pump, so i definitely will have to think about that. But since we're not in a bullrun, it will work like a charm, calm waters are the best waters for algotrading.
Backtest is never a guarantee, thats why im doing a dry run first. But it is still extremely useful to discard trash strategies early (if done well)
what kind of CPU you have if cheap EC2 can handle the training and yours is screaming? Possibly spend those 200$ on used CPU? lol
what kind of CPU you have if cheap EC2 can handle the training and yours is screaming?
Oh, no, i'm using the EC2 instance for dry running, so it will just train a model at time, then wait a real week to train another model based on real time data, so the CPU load is high, but only when training the model. In contrast, when i backtest, and if i want to backtest a whole year i have to generate every single model for every single window for every single pair on the training time range .
An EC2 can handle the task just well because even if its far slower than my computer, i just need it to finish training the 15 models (1 for each pair) in a "reasonable time" within the week i configured for the model expiration.
got it. also how do you automate the trade part?
HTTP GET and POST REST API calls to public endpoints of any bigger CEX
I finally make it guys
sub $20k portfolio
Kek. lol. LMAO
Let me guess, you backtested and adjusted your parameters to get better results. Your model is overfitted. It'll be even worse if your strategy is based on basket/grid trading. I hope you're setting up not just one algo that runs live but multiple models trained on substantial synthetic datasets to cover your ass for every market anomaly that you can possibly come across. Good luck, I hope you're going to make it!
Oh innocent sheep, you dont know how much money i will be able to make if my strategy turns out to be legitimate.
No, i havent run hyperopt to fine tune my parameters.
It'll be even worse if your strategy is based on basket/grid trading.
Its not the case, im using statistics to determine potential local tops and bottoms.
Your model is overfitted.
Its a possibility i cannot discard, that's why i'm doing a dry run right now.
Also, i have run over 10 backtests with different data each time and all the results are similar, which can potentially be because the model is overfitted or because the model is just really good.
Well, anyways, in 3 days i'll start posting the results day a day of my dry run trading 15 pairs. This thread is my personal blog right now.
Equal weights just so you don't fuck up the results with selection bias right?
Now that you mention it, i dont have idea on how to fetch the daily progress of my dry run. I think ill just activate a telegram bot and use it to fetch my run information.
bot working. dry run up, lets hope this is it.
we're not in a bullrun
You're going to make money, but it's not because you uncovered a new algorithm.
Hey I've been looking for a coder and have a shitload of methods that can print money. Want to help me for free? I don't have liquid cash since it's all in assets or for bills. The remaining is to catch potential dips. Basically not a good time for me to spend.
Point is it can be mutually beneficial. If you say yes, please don't blueball me and take them without hooking me up with the bot. Send me a telegram @ if interested.
maybe it was a mistake to program a telegram bot for this. Now, i'm fucking addicted to see my numbers go up through the bot.
Tradingview offers 2 years for free. I have backtested one method with over 300 trades worth of data
Really great to see some anon actually dropping some knowledge and adding to this community interesting read. I might try later but am trying to ride COIN calls through October run up to 88K but interesting reading.
Tradingview offers 2 years for free. I have backtested one method with over 300 trades worth of data
Yeah, unfortunately Trading view back testing tool sucks cock and reeks of look ahead bias. Trading view is absolutely unreliable, i know, since i own a pro account there that i'm still paying for some reason.
Yw, sharing my progress with you, anons, gives me the push i need to keep going. I really hate how this board is 90% jeets trying to scam people, rage baits and demoralization threads nowadays.
Best of luck, anon. I hope this really works. Finally something that's actually interesting on this board. Keep us updated.
Huh? The backtesting uses history only
Basically you put your method using Pine code and it measures how it would have performed if it were implemented
Basically you put your method using Pine code and it measures how it would have performed if it were implemented
Yeah, the issue is not the data, but how the data is treated when evaluating positions. And for some more complex setups with shifted candles and stuff, forget it, shit is buggy af. This is one of my earliest attempts to make a profitable strategy. Also, i don't know if trading view lets you perform monte carlo analysis of your strats, if it would be really sad if it doesn't since monte carlo is one cool analysis to support your strats robustness, i think that even more than the backtesting results. Pic related is the custom chimera strategy i tried to make work, it was a failure ofc.
Overfitted
So you are telling us you backtested on trainig data? Lmao, ok.
Even if you model isn't overfitted to the data, it's likely is overfitted to a certain market sentiment and trend.
So you are telling us you backtested on trainig data? Lmao, ok.
No, i have a separated set. I also introduced some random noise on my dataset to prevent overfitting. I think there's a chance it can overfit because my model has 5600 features.
Even if you model isn't overfitted to the data, it's likely is overfitted to a certain market sentiment and trend.
I retrain new models every 7 days.