I have run tons of tests, tons of different models and assembly of different models, i tried deep learning, but still i cant see that im anywhere near of finding alpha. I'm not here to be spoon feed, i'm here to vent some heat, i'm frustrated.

>>60769936 (OP)
Simple. You can't compete with huge players with million dollar algorithms. And then even they probably aren't doing great because algorithm trading is basically horoscope tier.

Making a profitable strategy is a full time job
Making a non shit algorithm is a full time job
Working to live is also another full time job

>>60769936 (OP)
My balls look like that.

>>60769936 (OP)
it doesnt (and cant) work because the market simply isnt based on probability its completely manipulated

>>60769936 (OP)
Try filter down all the crap somehow and get a handful of coins to review manually and risk a lil on those that look good. Still just luck in the end tho

>>60769936 (OP)
yes, it is

what's your tech stack, just curious

you are taking a midwit approach op

you are trading noise, you think there are patterns in noise, and you base trading decisions on parameters instead of simple psychology. you've been fooled by the indicator TA trading strategy astrology industry. buy low sell high. getting a 5% trade with 5x leverage is so fucking easy in crypto. All you have to do is survive drawdowns until the price rubberbands back to +1% above your entry.

i have a friend who literally is missing the bullmarket because he's developing a trading strategy

my leverage portfolio is up 18x since april I dont care if a nigger here believes me or not. but some of you retards are literally missing the most epic bullrun in history becauseyou dont have the balls to go 4X long with size RIGHT NOW, and you dont have the mindset to survive if things dont go your way. that's why you spend so much time developing 'strategies'. because you're risk adverse cucks. Just buy NASDAQ it's easier trust me.

>>60770187
>bro just go leveraged all-in on the top with an unhinged Cheeto in charge

>>60770180
i started using freqtrade, but looks like the framework is quite limited for multiple models use. So i started building a custom bot, where i will use DL to select features and determine optimal position size, and XGBoost + LSTM for my actual trading logic target values.

I also have not a fucking idea on what is a good target values for this kind of use case.

>>60770425
>building a custom bot
not sure what your bot is doing but a proper event driven backend and analytic / research layer over it is a ton of very error prone work. what issues were you having with freqtrade exactly?

>where i will use
have you done it yet? just throwing data into an ML pipeline tends to not be very rewarding.

>>60770449
>what issues were you having with freqtrade exactly?
Not exactly an issue, just really frustrated with my lack of "results". If something, the only issue i can think of is that i dont think freqtrade supports multiple models usage (Regression + NN) or does it?
>have you done it yet? just throwing data into an ML pipeline tends to not be very rewarding.
Yeah, but im also out of ideas, i need to move onto a new idea somehow, i hate being stuck.

>>60770456
have you watched the Quantopian lecture series, or read Michael Halls-Moore books (Successful Algorithmic Trading and Advanced Algorithmic Trading)?

what asset universe and timeframes are you working with? So basically you've just been feeding an ML model data and hoping it works out of sample? you are doing proper walk forward validation and out of sample testing at least, right?

Set a bot to watch the order book. If large buys = buy, if large sells = sell. Obviously more than that but basically yeah.

>>60770456
also, L1 or L2/L3 data? have you tried just reading arXiv and implementing papers? I'm making a tool you might be interested in btw

>>60770466
>have you watched the Quantopian lecture series, or read Michael Halls-Moore books (Successful Algorithmic Trading and Advanced Algorithmic Trading)?
Yeah, i have watched and read a LOT of Quant trading related stuff, but i dont know about the Quantopian lecture series you are referring to.
>what asset universe and timeframes are you working with?
I selected 10 pairs based on their correlation with BTC and ETH. I'm working with 5m as my main timeframe, and 15m and 1h complementary tf, and 7, 14 and 67 periods for my indicators, and a shift of 3 candles. My data split is 0.75:0.25 for training and testing respectively. I'm training on the last 180 days (due to hardware restrictions mainly).

Well, the results are terrible, even when the MRSE values oscillate between 0.009-0.018 for all training windows, which as far as i understand, its not terrible for trading. Ah, and my data to features ratio is 50:1 so overfitting chance is (in theory) really low.

>>60770485
>have you tried just reading arXiv and implementing papers
Yeah, the latest paper i read implied that some regression models + a Transformer works nicely for time series forecasting so i wanted to try that next, but looks like it will be troublesome to implement that with freqtrade.

no price and volume data is just random noise, you can't use it to predict future prices

Citadel gets satellite photos of Indonesian farms and Midwest Walmart parking lots to determine when to short agricultural products and Walmart's stock before/during earning calls but you want to make an algotrading bot?

Please explain.

>>60770491
That’s a very small amount of data to train a model on.. Hardware isn’t a limitation, it’ll just take longer if the hardware is worse.

What are you indicators doing for you here? Generating signals in conjunction with the ML model or what? And why the random periods?

are you trying to do stat arb or pairs trading with those pairs or ..?

>>60770506
>what is capacity
brainless take

>>60770498
so are you just like never getting good results OOS or what? With ML it’s important to do walk forward during training btw. Otherwise I wouldn’t be surprised that it doesn’t generalize to OOS data. You have to do WFV.

>>60770491
check out the lecture series too, it’s interactive notebooks and teaches basic analysis. probably a lot review stuff but still good workflow practice

>>60770476

>>60770510
>hey please explain
>fuck you retard
Well fuck you too jeet boy. I hope you get rugged by another jeet in a windowless office in lower Manhattan.

>>60770507
>That’s a very small amount of data to train a model on.. Hardware isn’t a limitation, it’ll just take longer if the hardware is worse.
How much do you recommend? Hmmm its weird, i have 198GB RAM on my rig, but i always get oom when i try to train over more than 120 days of data, maybe it is a freqtrade limitation.
>What are you indicators doing for you here? Generating signals in conjunction with the ML model or what? And why the random periods?
Indicators are my features. I selected the periods based on a paper some Anon shared time ago, about the preferred periods for some trading and quant firms, i thought about running hyperopt to get better periods, but i first want to get some results that are worth improving (yeah, my current results are worse than flipping a coin).
>are you trying to do stat arb or pairs trading with those pairs or ..?
No, each pair is a different environment for my bot, i just want to identify consistently good performers on those 10 pairs.... Well, i use ETH and BTC as complementary informative pairs.

>>60770530
I thought you were being a ‘smart add’. At any rate, I did offer an explanation: capacity.

>>60770538
>198GB RAM
Oh my, that is nuts. I have 8 GB and several years of minute data I process with no problem. You have an issue somewhere. For minute data I’d want at least 8-12m personally, correspondingly
I’d want at least 4-5y for 5m data.

>Indicators are my features. I selected the periods based on a paper some Anon shared time ago, about the preferred periods for some trading and quant firms
If you features are the indicators, what exactly is ML doing for you? It sounds like you might just be running a TA algorithm at the core or something (nothing wrong with that, but it would explain the difficulty in finding alpha).

Anyway, I think watching the Quantopian series would serve you well. It’s way better than any of the popular books, e.g Ernie Chen’s.

>>60770538
And how did you select the indicators? Was this a statistically driven decision?

>>60770562
>If you features are the indicators, what exactly is ML doing for you?
My current target values are local maxima and minima and the daily price net movement. So the indicators are just data the model uses to predict the targets.
>It sounds like you might just be running a TA algorithm at the core or something
TA as Technical Analysis? Yeah, its exactly that. I have no idea on how to implement Fundamental Analysis features on my model.
>Anyway, I think watching the Quantopian series would serve you well
Do you have any link, anon? i tried to search for it, but i get multiple results.

>>60770590
Nvm, i found it, the channel name is "quantopian" lol.

>>60770590
>So the indicators are just data the model uses to predict the targets.
Yeah, I mean this seems like it could be a big problem. You're building a model off a model here. If the underlying model isn't great, it's putting a lot of expectation on the ML system. Why not just feed the ML raw market data and see if it can derive better features or predictive ability? I'd say this, the lack of walk forward validation and the limited amount of data are huge deterrents to finding alpha in your current workflow IMO.

>TA as Technical Analysis? Yeah, it's exactly that. I have no idea on how to implement Fundamental Analysis features on my model.
What I meant was to distinguish between an ML derived feature and yeah Technical Analysis indicator.

>>60770605
>What I meant was to distinguish between an ML derived feature and yeah Technical Analysis indicator.
hmm looks like my knowledge is really lacking here since i dont know how to reply. Well, thanks Anon, at least now i have a plan to move forward, i will watch the quantopian series.

>>60769947
why can't an AI based on quant trading succeed?

>>60769936 (OP)
>sack of a shepherd

>>60770615
Yeah no problem, if you get a moment please give me some feedback on this site I've been working on:
https://alphapulse.org

home page is quant news feed, brief market recap
develop page is full IDE and terminal for power users
research page is like a UI around Jupiter notebooks, with a button driven version for easy pre-baked analysis
monitor page you can deploy and view live or replay (make sure it's working as expected, trace data through the full stack)

most the site is just a mockup currently but I'm trying to get real functionality now. lmk if the UI is intuitive pl0x

>>60770643
I like the aesthetic, Anon, but i'm autistic. If you intend to make a product out of it, the aesthetic has to be more inline to what normies like. Also, i understood the idea behind the modules, but an interactive tutorial for the research module will be needed anyways. As for the "monitor" module, i like it, but i would also put a button above the chart to switch to some other charting tools providers (tradingview for example) unless you want to implement every charting tool that these other providers have available (there is nothing wrong with this i guess).
Anyways, that's a very cool project, Anon, good luck with that.

>>60769947
>>60770045
>>60770456
sounds like these anons havent found the right bot QQ

>>60770803
>Also, i understood the idea behind the modules, but an interactive tutorial for the research module will be needed anyways.
Yeah, it's not done, just a mockup. I have to prioritize certain things, right now I sorta just play pretend and imagine my workflow when it's done. I just want to be able to work with any dataset (including a pre-computed signal database) and combine any strategies post-hoc with a suite of tested, robust code at the click of a button.

>i would also put a button above the chart to switch to some other charting tools providers (tradingview for example) unless you want to implement every charting tool that these other providers have available (there is nothing wrong with this i guess).
I agree the monitor tab also needs work, I think it will 'figure itself out' as I begin using it, so I haven't budgeted much time to it. the charting software is actually from tradangview's lightweight chart library.

I just hope the notebook driven UI makes sense and is appealing enough some people, at worse I'll have a nice UI for myself. you can export the button driven UI to Jupyter notebook code, so I thought that was kinda cool. then users could make their own UI in that manner in the IDE, too. anyway, thx for the feedback. good luck on your journey m8

>>60770499
>Human psychology doesn't have patterns
>Bots built by humans don't have patterns
HA. HAHA. HAHAHAHAHAHAHAHAHAHA.

>>60770506
anyone who disagrees with this post needs to order a Midwit + Dunning kruger test kit

>>60770972
peak dunning Kruger

>>60770562
>>60770605

I messed around with this stuff for a few weeks recently. It feels like there are much better uses of intellectual energy. You pour so much effort and resources into constructing this bot. You can never be entirely sure it's going to work beyond the moment at which you give it real money. Because no algo works forever. And if it does work, for some short period of time, 20-30% gains is all that's realistic to look forward to. It's the definition of trying to squeeze water from a stone.

>>60771009
I got lucky and had a working one pretty quickly, was pulling 5-12% gains per day for awhile, without a red day. I stopped because I noticed some bugs and unpredictable behavior, and am rewriting my whole system currently.

If you get lucky enough to get a taste of that nectar, it's all worth it.

>>60771009
And what else are you doing that's so intellectually rewarding? IMHO, it's not a bad idea to spend time on this as at best you get rich, at worst you have a valuable pet project on your resume.

>>60771009
>It feels like there are much better uses of intellectual energy
I’ve made decent money over the past few years building games and websites, but the work is grueling, and in the end it’s just taking crap from bosses or clients in different forms.
Trading apps haven’t made me anything yet, and blockchain development has been nothing but frustrating still I wont give up. This is the best shit for loner devs.

>>60769936 (OP)
>If price is below 1%/2.5%/5% recent average, and stochastic RSI blue line just passed the orange line from below, buy
>If above average by the same amount, and blue line passed the orange one from above, sell
You're welcome. >>60770476 has the right idea.

>>60769936 (OP)
i made one that buys btc everyday, it has been very profitable so far

>>60771488
When does it sell?

>>60771498
it only buys

>>60769936 (OP)
no

>>60769936 (OP)
Most TA/"Quant" strategies are bogus and don't make money, the vast majority of money is in arbitrage and market making (in of itself a form of arbitrage). If you can predict the movements of SPX with any degree of accuracy in the short term, you can get rich very quickly, that's why it is so hard to predict it.

>>60773643
>>60769936 (OP)
Once you read "A Random Walk Down Wall Street" you realize that none of this shit works unelss you can "cheat." And the only ones who can do that are employing increasingly ludicrous amounts of capital to scrape the bottom of the barrel. Even worse than that reall. More like adding water to the bottom of the shampoo bottle then shaking it up to get one last use out of it.

>>60773643
>>60774936
get off my board filthy casuals

>>60775019
you left out the PnL tab

>>60775037
was from a few months ago before finding bugs and rewriting system. any it was proof of concept that profitable strategies are out there.

>>60775093
This could be as little as 4 trades. Which is not statistically significant whatsoever. Need something more tantalizing than that, man

>>60775110
it was averaging 12-15/day, and was in alignment with my backtest performance (~130k bars). It was confirming the statistical significance of the backtest was holding up on live data.

>Need something more tantalizing than that, man
right, I'm working on it, but I can't just pull it out of my ass. this is a complex domain. hoping to be back up in a couple months with a much more efficient research and deployment platform.

>>60775110
Maybe it was more than 12-15, here's a sample from Alpaca.

>>60770506
That's what they want you to think when in reality they are just buying your order Dat from your brokerage and their market making algo moves in the most optimal path to at everyone's lunch

>>60775110
>>60775126
>>60775142
yeah looks like it was closer to 40-50/day, which would be around 1k trades while it was live, that's statistically signifiant on the live trading portion, imo. but I understand your point from the info I supplied.

>>60775142
I noticed that these were almost all losses, that day was one of the dips in the equity curve, picked a bad day to screenshot. the -17% in relation to current equity value as well btw, I think the worst day I had was like -5%. it only lost 1.15k that day.

>>60769936 (OP)
Just have the bot sell during new local highs and buy during local lows. It should be possible to program that.

>>60775241
>Just buy the bottom and sell the top, bro.
i keked.
>>60775019
What text editor is that, anon? it looks gorgeous.

>>60775507
The editor itself is Monaco, the open source version of VS Code, but I styled everything, added tabs, sidepane, tiling, etc. I'm using it for a project (https://alphapulse.org), it's mostly in a mockup state but I got the Jupyter Notebook backend working on my local deployment. At some point I'll extract the CSS styles and whatnot into a framework or something and post it on GitHub

>>60769936 (OP)
yes, but:
1. I would not try to focus on highly traded instruments like SPY/SPX/ES, etc. Tons of attention on them, and even if you find a strategy with a theoretical edge you won't be able to outcompete prop funds running colo'd FPGAs for execution.
2. I'd try to look for multi week or longer term investments using leverage in smaller market cap companies via options, or at crypto

>>60774936
>Once you read "A Random Walk Down Wall Street" you realize that none of this shit works unelss you can "cheat.
I've never read that book (I don't believe that reading books helps you accomplish things, it just biases you to whatever the bookwriter is writing) but I disagree with this. As I stated, the money is in arbitrage and market making and while market making may be a bit difficult for a smaller firm (outside of less liquid markets like obscure options), arbitrage isn't quite the same. Anyone who tells you the market is efficient is retarded, there are plenty of pennies that smaller firms can pick up that the bigger firms haven't either because they haven't found it or the potential profits aren't enough to justify finding it.

>>60770616
Sure, it can. There's 2 key issues I see. First, if your algorithm is time sensitive in any way, the big HFT shops are going to beat you. They literally just have faster connections to the exchanges; no self-hosted or AWS/Azure solution will win in a race. Second, if your algorithm isn't that time sensitive then you still need to beat all of the other quants that are doing this professionally. You might, but based on how readily AQR and other places publish papers on strategies that they reveal after they've lost their edge, many algos have lifespans measured in months or quarters at best. They're constantly tweaking weights and improving so anything you develop still likely comes with an expiration date.

I am not a quant and have no knowledge, but I did work on FPGA and ASIC design for a company that sold semi-custom chips to trading HFT firms.

>>60775909
What an otherwise reasonable analysis and viewpoint like this misses is the concept of strategy capacity.

>>60775789
Are you doing MM? Drop some knowledge bombs if so please. Is it worth diving into MM on dexes?

I find great edge easily, but its typically overfitted, but I don't care I use it until it falls below profitability metrics, I use blofin to very easily automate from a tradingview alert. there is edge everywhere just take whatever indicator you want and run through the timeframes minute by minute, you'll find something.

>>60769936 (OP)
The first step in sustained success is realizing that trading is literally actually a videogame. Not metaphorically, literally. The second thing is realizing that it's PvP not PvE. Understand that there are other humans (at the big banks and institutions) on the other side of the screen that are playing to win. The third thing is realizing that any strategy or "algorithm" that you come up with 'is only going to work for a limited amount of time if at all, because the humans that are playing against you are playing to win by taking your trading money and they will adapt and adjust if your "system"'has too much success playing against them.

If you are not treating active trading as a PvP videogame everything you do is dooned to fail eventually.

>>60770506
thats actually fucking fascinating. Anymore sources?

>>60775789
Yea and the “market makers” in crypto are literally cheating, getting free supply from devs to MM and then just selling. So his point actually applies completely.

>>60770181
>>60769997
these

>>60769936 (OP)
.i could give you a lead but i won't because all of you niggers use chatbots and i don't like AI