minimalist edition https://www.youtube.com/watch?v=xBFd6w3KnAI&list=OLAK5uy_kBFHQWSR3V3RPeRDSA1JKl_HpHDVgYYEA&index=7
>>62043450graphs and technology, all kinds of fun stuff. the future is yours
>>62043450Hello, i'm running my strategy live. Pic related are my expected results. Wish me luck.
>>62043450I fucking hate sabr models
>>62043479On the tested periods of my previous pic, my strat only shit the floor (huge drawdowns) when kikes manipulated the market with huge dumps.
>>62043490Why dont you just model the dumps?
>>62043496>Why dont you just model the dumps?impossible. Well, i will buy a rig to run QA while my prod build runs. I guess i can implement some safeguard logic for when kikes dump suddenly. But if i use TA or % distributions to predict those dumps, it turns out it is impossible to predict, it doesnt correlate to anything its just blatant manipulation, like pic related.
>>62043501This is my current autism machine btw. I'm planning on getting a similar one, but just to run QA to improve my current model and to build and expand my portfolio.
>>62043479One of the problems with discussing stuff in algotrading is the reliability / reproducibility.. As in, you're posting a Sharpe of 4+. If that's legit, you're set - but odds are, it's not. Even if OOS results have been good, you might be annualizing it wrong or something.. This is made worse by LLMs making coding so accessible. So I'm just saying I take all of it with a grain of salt, and you should extend that courtesy to me :-).Nonetheless nice results. Are they unlevered? Trading crypto I assume?
>>62043524>Nonetheless nice results. Are they unlevered? Trading crypto I assume?x3 leverage, yup, crypto. I have been cooking a quant strategy for the last 10 years dude. Yeah, i have already dealt with overfitting (solved via L1 and L2 normalization mainly), extremely random results, and lookahead bias. I have solved everything, so those results are most likely representative. I still have to deal with the slippage tho, since my current rig takes 3 seconds to make a prediction and enter the market, and those 3 seconds may or not be significant. Well, since i already wasted my best years being a retarded autist, i will just commit to the bitter end. I either become filthy rich, or i die a broke retard.Whats cool about my strat, is that it Adapts to the regime, my main target value is extremely good at knowing exactly the current trend and when it ends. I have one shitty target i have to rework tho, but with such a good and accurate main target that predicts the trend, i dont really care about that shitty secondary target.
>>62043524>This is made worse by LLMs making coding so accessibleAlso, im already a senior CS engineer. So i have filtered every single bullshit the slop machine has spilled to me.
>>62043546>>62043551>I have been cooking a quant strategy for the last 10 years dude.>I have solved everything>I'm already a senior CS engineerLol, ok. Way to get defensive and weird over a neutral disclaimer. I was just saying that a screenshot of a tearsheet is kind of worthless if we're not using the same platform and the methodology isn't disclosed (not the 'strat', but the train/val/test hygiene). It's like when noobs post insane gains in TradingView's PineScript - could be legit, but probably isn't just by probability. But I can tell you're of a sensitive disposition, so in an effort to keep this thread civil I'll refrain from commenting further.
>>62043562I dont know what part of my message was defensive, but ok, i guess. My goal with the message was just to let you know what i have dealt with and solved until now, just in case you know something i may be missing.Anyways, cool thread, i will keep everyone updated on my current live run.
>>62043488>I fucking hate sabr modelsWhy? I don't use them, but curious? I parameterize vol curves with SVI, but that's just for messaging efficiency over the wire so the consumer can reconstruct with a parameterized curve. Anything requiring real modeling is done with high fidelity.
>>62043572I'm planning on running this for at least 3 months. Since only by then i will have amassed a "representative" sample. And just turn it off if it shits the floor too much, or keep printing money if things go well.
>>62043572>I dont know what part of my message was defensive, but ok, i guess.The parts I greentexted? Are you actually autistic? Anyway, I wish you luck in live. I've been working on options modeling as the last piece of the puzzle in my system, and ran live for half the day today but hit +30%. Probably just lucky with the bot being retarded and making a directional guess - it's an online learner, so it was like a newborn baby today. Fresh out of the womb. Options/vol modeling is a surprisingly complex and subjective thing, too. More of an art than a science. Which turned this into a huge ordeal.
>>62043488How much do you know about vol surface dynamics modeling?
>>62043479>only backtested it for a single yearlmao
>>62043645Depending on the density of the data you’re working with, I honestly don’t see a problem with it. Markets are non-stationary. One year of cross-sectional L2 is a fuckton of data too, but yeah I get the point. Honestly I think the going back farther than 5-10 years is kinda pointless unless the model is adaptive during each slice, in which case it’s essentially forgetting the earlier periods anyway though rendering it pic rel.
>>62043572>just in case you know something i may be missing.And that's kinda my point too - posting a tearsheet and listing your credentials doesn't lead to any insightful conversation. I can't really dialogue with anything you provided, which is why I said>One of the problems with discussing stuff in algotrading is the reliability / reproducibility.. >So I'm just saying I take all of it with a grain of salt, and you should extend that courtesy to me Pic rel is something I've been working on for example to demonstrate something we could actually discuss. I'm super down to discuss any philosophy, architecture, design, etc. Obviously not expecting anyone to drop 'alpha', but there's a wide birth between this and that.
>>62043680This was how I defined my formulaic alpha grammar last time I touched that part of the codebase
>>62043685some notes on vol surface dynamics modeling
>>62043685>>62043691binary message transport notes
oh well. time for the final beer then bed
Very cool thread, im just getting into quant stuff as a current swe. But relearning my basic maths rn like probability, statistics, & calculus. Just need to get better about studying
>>62043723I was a math major and just turn my brain off whenever there's math. just straight up gloss and skim over it. it's kind of a super power I think really. if it matters (it doesn't) I'll dive into it otherwise it's just symbology and I just grab whatever I need, code it up and apply it.
>>62043685>>62043696What's even the point of this, in English how will this help you trade better
>>62043770Genetic algorithms can select for good features or targets... I guess... you can use them to automate an experiment design pipeline to select for features. I sincerely dont know the appeal of both NN and Genetic Algorithms for this use case. Sincerely, i believe that strategy-wise, anything that is not a regression, binomial or clustering model is a meme. NN and genetic algoritms can be used for portfolio optimization tho, that would be a good use case. But as for a strategy, with the former algorithms you just have to care about curating features, data, and your target, which gives you a straight path to improve your strategy.About the low latency message distribution system that i guess its the 2nd post you quoted is, that is a "must do" if you want to try a high frequency strategy ml based strategy on a low timeframe, since... lower lvl precompiled strategies = lower execution times = less slippage. Also the same logic applies for complex trading systems where you orchestrate multiple signals in distributed systems.
>>62043770I made over 2k trades today and ended the day up over 30%. I couldn't have done that manually. Granted, it was a directionally biased day.What I posted will not directly help you trade better.
>>62043792>I made over 2k trades todayWow, thats quite insane. What was your execution time? everything included? (latency, evaluation, execution).
>>62043784>Genetic algorithms can select for good features or targets... I guess...GAs are capable of searching a discrete space. Features are algebraic expressions that are searchable in that discrete space. Any feature expressible in an algebraic form is searchable in that regard. These are often called formulaic alphas. > Sincerely, i believe that strategy-wise, anything that is not a regression, binomial or clustering model is a memeThe optimization selected should be based on then problem being solved. You can't do a regression over a discrete space where each structure is unique, for example. I utilize whatever is most fit for the problem I'm solving. In a search for formulaic alphas / features, a GA is very fitting - I'm not aware of anything more suitable but open to suggestions of course. I actually do GA + MCTS.And yeah the low latency messaging stuff is more an implementation detail, but relevant for us DIY'ers. For me it wasn't just about high frequency / low latency but throughput - I wanted to be able to process high dimensional options data across assets quickly and potentially distribute my compute horizontally.
>>62043805It's not really set up to be super low latency or anything right now, so I haven't even measured it, I was just happy it was able to keep up with the market and not fall behind in processing overload. I don't expect an edge in speed or anything there, but still want performance for backtesting purposes.. When I was doing crypto/defi stuff I was more into tracking the live latency but it's been awhile since I've been in that area.
>>62043770post again retard I need to aware you
>>62043824yeah, probably in my case i wouldn't even bother trying to make my own distribution layer. I would just try to use a lite version of Apache Kafka.>GAs are capable of searching a discrete space. Features are algebraic expressions that are searchable in that discrete space. Any feature expressible in an algebraic form is searchable in that regard. These are often called formulaic alphas.Well, yeah, GAs can work really well if you have a set of candidate targets you want to test and optimize a set of candidate features for. But once you have found a superior alpha after your selection, its just so much faster to just go for a simpler model. Well... assuming that the target can be modelled into a bunch of Boolean expressions (target labeling). Simpler models are easy to debug, to interpretate and hence, to improve.Well, anyways, I'm not trying to be critic to your method anyways, just explaining my experience with those approaches on my own terms (since i still not have a clear picture of what your procedure is). Well, its time for me to sleep. I just hope i don't wake up to a bunch of red numbers in my positions.
i like the quant trading but this thread has the highest density of wannabe smart retards on this entire website
>>62043479>that difference in IS vs OOSbuddy why would you even try
>>62043854What do you mean distribution layer? IDK, I just coded C/C++ in uni and knew how I wanted the bits moved and early on I wanted to make sure I could abstract the transport interface so I could do shared memory, unix socket, tcp etc without code changes because I wasn't sure how things would change as time went on and I needed integrity at that layer. I use an actor model and can just make config changes to change transport type between actors. But most typically I run everything on the same node and everything in memory. Sometimes especially when debugging I'll run the data feed separately so I can shut down consumers for debugging without interrupting the data archiving.>Well, yeah, GAs can work really well if you have a set of candidate targets you want to test and optimize a set of candidate features for. What do you mean here? The GA for me generates the features.. And separate learning processes (MCTS and LTSM) guide/bias the evolution between folds/phases. The GA is just one small part of my system, however. I don't know how else you'd do unbiased feature search/engineering though?My feature selection is done on a continuous basis by an RL learner that selects from a candidate pool during an online phase. But the search for candidate features itself is done via the GA.>Simpler models are easy to debug, to interpretate and hence, to improveThe formulaic alphas from the GA are completely interpretable. I've intentionally kept all stuff interpretable and linear by default in my system. So far I'm only using non-linear methods where either speeds up the search process - I don't really care if that's a blackbox personally.
>>62043873it was 3mo OOS vs a year IS..
>>62043871maybe you're just dumb? just join in and LARP if so, you'll fit right in
>>62043837Post PNL right now or I'll just assume this AI slop
>>62043877>What do you mean distribution layer?The data distribution layer, you can just lookup the definition, which is basically what apache kafka does.>The GA is just one small part of my system, however. I don't know how else you'd do unbiased feature search/engineering though?Oh, thats interesting. I thought you had a huge pool of pre defined (either manually or automatically generated) features and targets you wanted to select from, and build different prediction models from there.>The formulaic alphas from the GA are completely interpretable.Well, cool then, i have nothing else to say.>>62043871Even assuming you are right, this is still the highest quality thread in the entire board.>>62043873Yeah, naturally tho? look at the windows?>>62044466This is the worst part of this kind of threads. You niggers are so cynic, you cant fathom people trying to build cool stuff.
>>62044672>Cool stuffBeauty is in the eye of the beholder I guess, so far none of this sounds actually useful and sounds like a bunch of psued CS majors jerking each other off
Overnight update, shit is looking nice. Entries make sense, feels good man. Next Update: 04-04-2026.
>>62044466>>62044714I already posted that I was up 30% yesterday. Not sure what else you want.
>>62044864Live? Feel free to share as much as you want about it>>62044672Yeah for me features are represented in an algebraic form as an AST, and the GA searches that AST. This paper is kinda similar but uses MCTS over GA. https://arxiv.org/pdf/2601.22119I use both. It’s not alpha in itself of course just a way of searching for features. The alpha comes from the selection process IMO and avoidance of overfitting / optimizing for OOS generalization. But again this is just a small part of the system, it’s just how I select features that improve the overall model.
Cool thread. I'm just getting started and the hardest part for me right now is dealing with the data feed and data archiving. Any tips for getting data into your system? And, if you don't mind me asking, from which sources? I'm thinking about having a component whose only job is fetching/archiving data and making it available locally to other components of the system, either through a Unix socket or a database.
>>62045135most crypto brokers have a data API, where you can fetch data from. You will still have to deal with the shitty API request limit, but you can find some proxy providers that help bypassing that. In my case, for binance i'm using a free data proxy provider that gives me just enough requests per minute, to run these 6 bots.If you want to train a bot on historical data, you can go to most likely any databroker out there (polygon), you have to pay, ofc. Ah, also in the case of Binance, and most brokers, they dont have the whole historical data set available for short timeframes.I recommend you to just work on a publicly available framework until i get something promising, so you start building your optimal system for it. I recommend you to build first your data pipeline (basically an ETL) with well defined layers, then the data distribution layer (you can just implement this pattern right here >>62043696, or just use SQLite, locally archived data (parquet, feather), and access it from your bots. You will will have to configure a decoupled component to download and stream the data on your running bots in near real time tho (adjustable, depending on what timeframe will you trade).Anyways, most of your technological stack selection will depend on your objective, main timeframe, strategy, etc, etc.
>>62045135For raw market data, I'd recommend a binary format, or just simple record format / rows. For any derived data, I use Parquet. And like the other anon said crypto is a good place to start due to all the free data.>I'm thinking about having a component whose only job is fetching/archiving data and making it available locally to other components of the system, either through a Unix socket or a database.Yeah you'll definitely want something like this. I'd recommend watching this video for an overview of standard architecture layout:https://www.youtube.com/watch?v=OhGdVdolqkUThe speaker has a couple really good books on the topic too:https://github.com/zslucky/algorithmic_trading_bookI typically run the data feed and archiving component as a separate process like you suggested and use unix sockets to send the bits to the consumers. Personally I haven't really found a use for a traditional database yet but I imagine once I start doing earnings analysis and such I'll be using it.
Hope you're doing well with your systems. My log for this month so far is $11,563.14 spent, $18,212.23 won, $6,649.09 net profit. I haven't scaled as aggressively as I maybe should have, I just bump the buy rate up once or twice max a week. Don't know why i'm so timid. But needless to say very happy with this and future is looking bright. 12 strategies, trying to find more.
>>62047321(There is ~50 usd profit yet to be logged in this pic, I just changed poly account because your trades are public and a lot of people started checking it out. Getting paranoid about people stealing fills)
A cool idea I have for expansion is I'm going to set a mini PC up in my living room for each of my 3 assigned IPs from my ISP so I can get 1.5-3x the fill rates
Do you worry that we only have a few more years before big actors manage to use AI to pump out algos that arb out every possible edge in every market, even $10 a day ones?
>>62047321Is this live? What is it trading - you mentioned poly, is this a polymarket bot? Awesome results if it can yield like that every month though.
>>62047321Thats really cool, Anon, thanks for sharing. What kind of strategy are you running? AI based or TA pattern based?>>62047366>Do you worry that we only have a few more years before big actors manage to use AI to pump out algos that arb out every possible edge in every marketIts not going to happen unless AGI becomes a thing. And AGI most likely will not become a thing on the next 10 years. Even so, by definition, unless there is a single winner on the market, even if someone manages to make a definitive alpha algo, it should still generate exploitable patterns, since they will inevitably affect the market with their positions. So... i'm quite comfy tbf. Now... if i'm wrong, and we have a HUGE technological breakthrough that makes AGI viable, the market will be the least of my worries. We will be set for a dystopian where money and effort will be irrelevant.
>>62047692I also have a papertrading class in the same software for forward testing new strategies, but this is live and on poly, yes. What's really cool about it is it feels like I can do something to improve it a little bit every day. For example, that 370 loss last Saturday happened because I didn't have a guard against a special condition and it bled out all day (fixed). The other day I added a new set of markets that require a special API to be useful, so that added more potential opportunities per unit of time. So I've been working on this pretty much every waking hour for the last month.>>62047708I'm wary of AI other than for churning out code faster, so it's just based on technicals and API feeds. There's a million people trying to get current LLMs to make buying decisions and as far as I have seen so far they aren't useful yet. Even those AI trading competitions produced junk results. Any regulars ITT been able to make it work?
>>62047829>I'm wary of AI other than for churning out code faster, so it's just based on technicals and API feeds. There's a million people trying to get current LLMs to make buying decisions and as far as I have seen so far they aren't useful yet. Even those AI trading competitions produced junk results. Any regulars ITT been able to make it work?Its no use. If you want to use LLMs for something trading related, you can use them to analyze sentiment, i guess. But using them to analyze raw data and/or indicators and expecting it to give you perfect entries/exits is just dumb. You can also use them to optimize portfolios tho, as long as you give LLMs permission to execute code, you can just let it use some premade scripts you prepared to evaluate pairs, or if you want, let the LLM use REPL to write and run the code on demand.In summary, nothing changed, there are million of retards doing retarded things, expected from retards. Its attractive, because its the "easiest" way to do things, so you can expect millions of lazy retards to gravitate towards it.
>>62047829>polyNice, I have an arbitrage/MEV bot system for polygon specifically, but shelved the project before getting into Polymarket. Seems kinda low liquidity and a short lived source of alpha but IDK. I also get tired of feeling like Polymarket is running astroturfing ad campaigns all over the place. About the AI/LLMs changing the competitive landscape, I don’t think it will have much of an effect except for retail getting more into algotrading due to the barrier to entry being lowered. LLMs don’t do anything a big firm couldn’t already achieve by assigning a developer that task, and markets will never be perfectly efficient. It seems that despite LLMs lowering the barrier to entry, retail is still just as retarded and they do retarded shit like trying to get ChatGPT to tell them to daytrade for them.
>>62047931>>62047829And idk why people do this - why use a language model on time series data? Completely different classes of problems. What you *can* do that makes sense is use the transformer architecture, like what the Deepseek team does. There’s a good article on this I’ll share if I can find it.
>>62047986It gets even worse, they are not even giving it access to the time series data, but to the "view" itself. The model basically takes screenshots of the chart and stuff, and they expect it to "trade" from it. Its so fucking retarded its unreal.
i just turned on my first crypto trading bot. am i allowed to post in this thread now?thank you and have a nice day.
>>62049675Yes, you don't have to have anything running or even be interested in automating trading to post here. QRD on the bot? What instruments does it trade, what timeframes, etc and any design/logic stuff you're comfortable sharing.I think in general people tend to be too dramatic about keeping things hush hush here. The devil is in the details.
>>62043450> Another juice scam comes to an end> It's not always them but it's always themis it just a coincidence the heads of all these AI companies are juice?
>>62049766bot?
>>62049724shitcoins over the course of minutes to hours. custom made signals based on basic price movement / volume metrics that i get from on-chain data and a few APIs. some basic gates to prevent rugpulls/scamcoins.i've developed about 15 signals that do OK, the best one has mean gains of 12-15% over a backtest of 2 days, most of the rest are from 2-7%. i just turned on the most promising signal live today so TBD what 12-15% translates to irl once you account for slippage etchave spent some time backtesting and debugging to make sure that the live bot actually triggers the same trades that the backtest does (usually 98% accurate at this point) based on quotes that i get from the exchangestl;dr shitcoin, custom signals, probably just psychosis thinking it will make gains
>>62049832Nice, I imagine volume/liquidity ends up being an important indicator with shitcoins? Probably tracking OG/genesis wallets for each coin would be a good move too.. Why such a short backtest period?>have spent some time backtesting and debugging to make sure that the live bot actually triggers the same trades that the backtest does (usually 98% accurate at this point) based on quotes that i get from the exchangesNice, a lot of people skip this for some. I have no idea why. If you can't validate/guarantee parity between backtest and live, you have no idea what your system is actually doing or how it performs.Sounds fun if you're working with play money desu and I have faith in your journey anon
>>62049968>Why such a short backtest periodbecause i only have the data that i've parsed and stored myself, and 3 days ago was the day that i fixed the most recent bug that was giving me duplicate records and i wiped the older data> Probably tracking OG/genesis wallets for each coin would be a good move tooknowing that liquidity is burned/locked is one of the gates. i track the large accounts but its not as meaningful as you think, at least in my tests>I imagine volume/liquidity ends up being an important indicator with shitcoinsin what market is it not?
>>62050035I'm sure you know but Kraken has a buttload of free data available for download:https://support.kraken.com/articles/360047543791-downloadable-historical-market-data-time-and-sales-Unsure if relevant to your application though.>knowing that liquidity is burned/locked is one of the gates. i track the large accounts but its not as meaningful as you think, at least in my testsI meant more like the original dev wallets for the shitcoin being traded>in what market is it not?I haven't done any sort of formal analysis on how useful it in various markets or anything, but obviously on it's own it's not a directional signal, and when trading discretionarily I don't find it very useful but that's just me (and if I was a stellar discretionary trader, well I probably wouldn't be here). I guess I find something like VWAP or dollar bars more useful than just time based volume bars.. IDK. That said, volume as an input appears frequently in features my system converges toward, and I use it for state/regime estimation.My point was more that in shitcoin trading, its importance is heightened, because volume can and will go to zero. If I was trading shitcoins, personally I'd be doing a lot volume flow analysis and stuff, just what my gut tells me.
>>62043723Are we back in 2021? AI exists.>muh "with AI you don't have to think"If that is your reaction, you will fail at quant. Every day of the week, every year.
can i also have my own stock trading bot?>t.down 17% ytd
>>62050136>I'm sure you know but Kraken has a buttload of free data available for download:thanks i'll probably use it when i integrate with a CEX, hopefully its similar formatting to what i'd actually be collecting
>>62051460>can i also have my own stock trading bot?up to you. You have to learn a lot of shit tho, minimum 3 years of intensive learning if you are not familiar with programming and maths. 3 years, because to test these systems is extremely time consuming.
>>62051295What do you mean? How else would I pass the interviews of my math is lacking?I guess I just dont understand where AI comes into my post/how it would be useful atm?
None of you know this shit lmao
just so you know besides the wicked shit doing a simple ma cross cancels out most strategies lol
>>620536223 years is a while lmao you must be a slow learner
reeeeeee
I'm -55 for the day, it's over
I get self doubt when the system just bleeds for a long period of time despite a month of performance and my strats mostly being based on extreme variance. Makes no sense
>>62057601>>62057690just carry on, if your "bleed" duration and magnitude are consistent with what you backtested, and if the cause for these bad periods correlates with what you found while testing the strat (turbo dumps), then there is nothing to worry. Trust ze plan.
>finally made it back to where i was before i let my database get full and my positions stagnated for 6 hours 2 days agothank you god
>>62057771Exactly right. Just cursing my dumb monkeybrain for injecting irrational doubt into my life when I know down periods are expected. My max recorded losing streak is 267 (from early on when 100% of my strats were extreme variance)
>>62057859>>62057690It could indicate you're overfit or don't have parity between offline and online execution. I'd start by making sure that when you replay over the same data it traded live, the system behaves exactly the same.
>>62057690Also, could just be a regime change or markets doing their non-stationarity thing.. If you find a pattern in the train/test/deploy behavior, like the edge only lasts for a month, then you can base your retraining cadence off of that and retrain monthly.
In case it's interesting to anyone
>>62057887Realy cool stuff anon. Are you using Tradestation? I tried to use it about 8 years ago, and I'm not smart enough for coding haha.
>>62058372looks cool and sovlful.
>>62058429Just python with curses. Not knowing programming isn't a barrier anymore, just ask an AI to walk you through whatever you want to build, but build a little bit at a time or control and grasp will be out of your hands and you will be frustrated
>>62058502What broker?
>>62058429No, I'm using hand-rolled Rust for everything atm. If I was starting over today, I'd probably use something like NautilusTrader + LLM. Not because I think NautilusTrader is better than my stack, but because you'd be able to get moving so quickly without risking building on top of a low quality vibe-slop codebase, and the LLM can handle all the boilerplate etc. Kind of like a harness for you to run the LLM within. And after doing that for a bit, you'll know what you want out any future system you might build.>>62058372This isn't working for me? Anyone else unable to load it
Is algo trading really doable without coding knowledge with the help of AI?I only installed the programs and I’m already lost lol
>>62061451
>>62061451Yes probably my dad who has almost no computer knowledge has set up openclaw on his computer and uses it for his work. He showed off to me how he literally just copy pastes all the issues he has into different AI's and they fixed everything anytime something would break with it.
>>62043450I am new to this, so what broker or exchange do you use for this kind of trading? You do your own python script and run it with an api or something like that?
>>62061684I prefer Alpaca for US equities and options, because the API is modern. I haven't traded futures yet.For crypto (CEX or DEX), the options are endless since all APIs are modern and data is free and abundant. When first getting started, I'd recommend crypto due to this.>>62061451Yes but I'd start with a good foundation, like NautilisTrader, and have the LLM just do the glue code and your custom strategy logic. It minimizes the surface area they can screw up.
status report on this retard?
>>62060200>This isn't working for me?It's just the screenshot except the numbers are moving, it's a python curses appPolymarket just added up to 7.2% transaction fees for crypto bet purchases. I looked up PM employees on X and they're the most unprofessional and smarmy douches I've ever seen, like Shkreli but no charm whatsoever. They've even hired that big pepperoni nipple guy, LeGate. Politicians are going to get raging hardons for putting these people away. I need to start looking into other platformssry 4 diary
>>62063621wrong thread kek, meant for /smg/>>62063627>Polymarket just added up to 7.2% transaction fees for crypto bet purchasesMy god, brutal. I physically winced while reading that. The whole PM industry feels like it's desperate for liquidity right now too, crazy they'd up the fees like that.
>>62063701Yeah. A vast majority of bots are now unprofitable. I saw a stat posted on X that the change also made retail traders even more unprofitable than before; only 37.1% actually make any money at all now. Used to be mid 40s before they started experimenting with fees.>I physically winced while reading thatSome Economy section bets actually have 30% fees. Was as high as 96% before they "fixed" it
>>62043450Didnt realize you decided to make a new thread againHope you keep baking for a while
>>62063847I think I will for a bit, I need to find or remake the original OP text though
>>62063701brootal
you guys just love fancy trading terminals buy mag 7 fucks sake
>>62064180why is there so much anti-quant/stats sentiment in a domain that is so numerical, its like you guys think it's purely masturbatory
this is all you need
>>62064224BECAUSE ITS FUCKING BLOODY USELESS NIGGA ESP IF YOU'RE LOW IQ
>>62064246>>62064233All you're doing is telling us how ignorant you are.
>>62064325I just ignored him, he obviously came here starved from attention, no one can be that imbecile.
>>62064391There’s a never ending stream of people just like him. There’s this massive gulf between ‘TA as astrology’ and ‘quant as astronomy’ that’s sort of hard to bridge. Its easy be dismissive of it as a rational person if you fall for the ‘muh team of PhDs’ FUD (or efficient markets), and it’s also hard to talk past the ‘bro RSI is all you need’ hypebeast. It’s for some reason difficult to communicate that you’re just looking for validated probabilistic edge, which you’d think would be super reasonable and intuitive way to approach the market. For the sake of public relations and onboarding I have to streamline a better conversion pitch though
Code monkey here who wants to do some backtesting. Is there some open source framework for this shit and where would I get historical data?
>>62043450Hey bros, I want to get into algo trading. I'm a applied mathematician and I can code C. Tips/tricks/strategy? How to start, what to read etc. Thanks for any advice.
>>62069444Like the other guy, I would recommend starting with crypto. There's low barrier to entry and you could technically build something and start testing strategies with only like $50 capital tonight if you wanted
>>62069444>>62069845>cBetter to use python though, execution speed doesn't matter so much unless you're market making, but that's grandmaster tier in difficulty and you shouldn't touch it. You want to find something that works and iterate fast. Python has a zillion examples and libraries ready to go
>>62057601Stats since thisI've paid $133 in fees since they added the new rates. Fewer opportunities/trades overall which was unexpected (fewer bots means higher % of trades are humans which should mean more mispricings for my bot) but could just be variance
Does anyone have this book? A quant I respect highly recommended it but it's impossible to find. I think the elite is memory holing it.
>>62071099Who's the quant?
Is most of what you guys do here just complicated statistical models?
>>62071099It's in stock on Amazon. Didn't see it on libgen unfortunately.
>>62071595I do model free stuff too like arbitrage, or attempts at it anyway. Not really sure what people ITT do though. I think what a lot of people do when first starting is just mess around with grid searching over the parameter space for common indicators like RSI, MACD, etc then probably ensemble those with some regime conditioning.
>ctrl f var>no value at risknone of you press any buttons until you calculate var thresholds over time for at least sequences of 50, 60, and 70 percent chances of >40% losses over some time thresholds, intervals or sequences
if there'd not have been how to have answered which sequences of events or what examples of trades would have caused greater than at least seventy percent chances more than forty percent of some portfolio would have been lost after some time or some drawdown it's not a trading set up
also uses of the words "stuff" and "thing" during trading hours get employees kicked out or fired from quant desks
>>62071716>employees
>>62071738"let the employees deal in the real world"
>>62071574Won't give specifics but he oversaw the libor migration in our bank, predicted the Nvidia pump and made a lifetime of cash.
>>62069444i think you're a larp, no applied math major or phd doesnt know how to find resources and not end up at BS model
>>62072480Didn't realize it was someone you knew personally, thought it was a YouTuber or something. The table of contents on the book implies it's a pretty standard coverage of vol topics though.Personally I really like Euan Sinclair's books and read whatever vola dynamics shares.
are you anons reading a book?
>>62043546>3 second entry and exitAre you using a VPS or is this unironically being run on a local machine? How can you expect reliability if this thing doesn’t run at server speeds on no bloat OS
>>62043655You’re analysing market data.. do you think market conditions were the same last year as to what they were ten years ago? What if a market crash occurs now, or theres scam candles everywhere? I’d want my script to be stress tested to fuck no way some jew is gonna scamdle me without a fight
>>62043871Youre right but I prefer smart people thinking they’re more intelligent than they really are than retards who think they know everything
>>62045135Use AIBinance for DataCumulativeCompile
>>62047366>a few years out The market is literally crashing, rob its corpse and move on
>>62049832What the fuck is your rugpull logic? That I have to see, I’ve seen those things lose 20% in fractions of a second. I think honestly the only way to win those rounds would be an actual sniper that has a sharp trailing stop.. otherwise why take the risk, even with an ai trading all these shitcoins do nowadays is crab down
>>62075199>Are you using a VPS or is this unironically being run on a local machine? How can you expect reliability if this thing doesn’t run at server speeds on no bloat OSim trading 15 minutes candles, and only evaluating closed candles, so, lets say, the 3-5 seconds that takes my system to get in a position is not significative, also, i'm using a fixed stoploss open directly at the broker. Latency and position timing role gets more important, the lower your main timeframe is, i learned this the hard way time ago, trying to make a strategy for 1m candles.
>>62053622>3 yearsLooool Its the ai era baby, same goes for any scripting. Get the AI to do the heavylifting and you shave off months maybe even years of work. A wise man working a shit job for a shit wage once said to me ‘work smarter not harder,’ and he was right. Time IS the most valuable currency you have, why not expedite the process if you directly benefit?
>>62061684Ask the AI. We helios now
>>62063701>predictive markets owned by Jews is Jewish Damn who couldve seen that coming
>>62075264>Looool Its the ai era baby.You have to know how to filter the retarded conclusions AI will lead you on very often, otherwise, you will waste months testing and optimizing over misleading clues just because the slop machine said so. Whats scary about using LLMs without knowledge, is how convincing can their explanations sound for someone without a solid base. But you are free to try to ask the slop machine to build you Renaissance Technologies, i guess, i will not stop you.
>>62075252Whats wrong with 1m? Is your token so stable its immune to hebraic chicanery?
>>62075284>Is your token so stable its immuneQuite the opposite, the more volatile is whatever you intend to trade, the more important is timing on low timeframes. In some volatile pairs, 1-5 seconds could be enough to fuck up your entry with a huge slippage.
>>62075280> You have to know how to filter the retarded conclusions AI will lead you on very often, otherwise, you will waste months testing and optimizing over misleading clues just because the slop machine said so.I know that, I also program in C# for unity and use LLMs to assist, a lot of the time its just iterate>test>fix rinse and repeat, its usually never 100% out the box it takes time, but still 1000x times less to learn from scratch, especially if you had a life like [the chosen one from fallout 2].Hehehe you simply must ask the right questions, to the right person and the right people and you will get the right answers. But you have to go to /g/ like I did for a model that isn’t garbage. Follow the Claude Ai leak breadcrumbs. Most mainstream AI is heavily censored
>>62075288I know that but from my estimation its been impossible to find second by second data in historical market data, thats why I accepted minute candles. I couldnt tell my bot to make decisions every 15 minutes thats just a delayed reaction I dont understand your logic at all
>>62075292... Look, i'm not trying to convince you about anything here. Just go for it if you feel that confident.
>>62075306You’ve clearly had a different experience to mine and it’s affecting your perception of what I’m doing. Dont let your reservations get the better of you
>>62075264>Looool Its the ai era baby,Adding to what the other anon said it's not that simple. The worst application and use case for AI/LLMs is in complex, safety critical code which is *exactly* what this is. It will introduce subtle bugs, do things wrong, stub things out, implement half ass or not all, hallucinate, fudge the tests, etc etc.. They also have a strong bias towards steering your codebase to stuff that's more in line with what it was trained on, so if you're doing anything non-standard it'll trend back toward slop.Not to say it's not useful, but it's not a magic genius coder factory. I use it to greenfield modules and prototypes, but I don't let it touch modules after I've either manually reviewed every line of or had written by hand.
>>62075315You guys are getting huffed up over nothing. I test what these things put out *every single time.* Your knee jerk reaction to it is like a child who got burned once and then never approaches the stove again. I’m not saying its god, im saying use it like you use a shovel — correctly and with consideration as to what it’s usable areas are, like any tool. Most mainstream ai is trash, granted. Thats why I said to become hansel and gretel and browse /g/ for better LLMs that handle coding logic .. more amicably. There are guidelines already in place that make those shovels prone to error/rust.. Not all are created equal, you must know the tool and yourself. The ignorance works both ways.
>>62075328I'm not huffed up at all, lol. Just saying, it's not as simple as prompting the AI/LLM and spending a weekend or two and arriving at a profitable, robust system. It takes time and the act of writing code itself isn't the primary bottleneck (as usual).
>>62075334This isnt a weekend this is 1-2 weeks of iterating.. I wanted to deploy earlier during the week, but havent because ther is still room for reasonable doubt. In fact, I analysed both your and my messages again and realized the knee jerk reaction was mine — a fear response that I may have missed something — that triggered a defensive response.. In your line ‘Fudged the Test’ has echoed my own sentiments exactly and means that even successful strategies that the AI makes in its own API have to be successfully bridged to other backtests/until parity is achieved. Honestly, if it were up to me Id just plug it into paper trading and call it a day, but the perps I want arent on the virtual exchange so I have to do everything more meticulously. Bummer, but rest assured I do actually echo your sentiments of caution even if I understand that to the more lucrative approach here is to work with the pros and manage the cons.
>>62075334To be honest I read your posts again.. I dont mean to sound demeaning but Im going to hazard a guess and say I’m more accustomed to working with LLMs than you and are far more comfortable even with their limitations and problems. I know they have a tendency to incorrectly iterate, thats why testing is so important. Is anyone really going to take what th AI at face value? I dont even know if I can do that with the best of them lol but at the same time, these are scripts anyway, Ive been working with scripts and programming languages for over a decade, I feel comfortable and confident — testing phase and then its game time.
>>62075350>>62075340I've been (full time) using Claude/Codex max plans since they were released, Gemini occasionally but it sucks and is a weird depressive crybaby for some reason.. Lol. So we're probably similarly equated if you've been using them heavily as well.>I feel comfortable and confident — testing phase and then its game time.I wish you well, I'm just sharing my experience here. What will determine the pain will of course be variable depending on what you're building. In my case, I wanted to be able to use the same components (like binary transport logic, market data types, walk forward testing workflow, feature engineering processes, orchestration and logging, etc) across all domains like defi arb/MEV, microstructure, and volatility surface modeling. Just building the defi arb system how I wanted it took months, and it's only one very small part of the system. I also wanted to make sure that my transport method, between say unix sockets, shared memory, TCP was 'pluggable' and that I could change how any given module communicates via configuration + codegen, so I wouldn't be drowning in bespoke wiring changes if I wanted to move to multiple machines or parallelization. Vol surface modeling has taken about 2-3 months alone so far as well. Things like that take time to resolve a design for, and there's a million different design problems you'll encounter along the way that can take just as long.If you're pretty narrow/focused in your goals or applications though (I mean that in a good way), you can of course move a lot quicker. I just got tired of needing separate systems for what is ultimately a very similar pipeline of data in --> decision making --> action out, so I wanted a unified and flexible platform and that just takes time.
>equatedacquainted, kek I knew I was fucking something up there
>>62075391Tbh, without going into too much detail, I noticed some cracks in the ability of claude during iteration. If you’re gonna trust the thing to handle your money, you better swing on to /g/ to make sure you’re getting your money’s worth, in my experience, capitalism makes people too comfortable with the idea of simply paying for something and then that product somehow magically becomes better — but its a sunken investment. I dont wanna feel like I’m gatekeeping here, so I’ll say that if you’re really serious about it, go to /g/ and scry for a decent LLM because as I keep saying *mainstream LLMs have guard rails* and nothing on this earth can change that except it the companies that own those LLMs relent (they wont, mainstream LLMs are heavily censored). I’m just going off what you wrote, but it sounds like that would’ve been more beneficial for you than bouncing off that fool that has a tendency to iterate mistakes with the confidence of kings. Also what you describe somewhat veers more into ‘Hobbyist’ vibes for me, and I’ve always been wary of considering this as more than a tool to perform a function, I see the market as a tool, the LLM as a tool, the bot as a tool, etc etc, and to watch over these tools and curate to them and maintain a symbiotic relationship with them that can always manage to be productive.
>>62075444>cracks in claudeYeah, I find it quite lazy - especially lately. My workflow changes all the time since the models and their quirks do too, but right now I use Codex to dispatch and review but Claude to actually implement.. But Claude Code's UI is way better (Codex doesn't even have hooks) and it feels much more natural to work with than Codex. I've been anxiously waiting for local models to get good enough to run locally on something like a Mac Studio, so I can stop subscribing to these AI companies. >Also what you describe somewhat veers more into ‘Hobbyist’ vibes for me, and I’ve always been wary of considering this as more than a tool to perform a function, I see the market as a tool, the LLM as a tool, the bot as a tool, etc etc, and to watch over these tools and curate to them and maintain a symbiotic relationship with them that can always manage to be productive.I mean by all means, design your system how you want - I started with the same mindset you have now. I was just sharing how mine was designed because it highlights how this can take longer than a few weeks, and how your design goals might change over time. If your goals are more focused you can move quicker, but for me the time has been well spent in terms of ROI.
>>62075465>stop subscribing to AI companies If I were you and even remotely curious, and I’m gonna beat the horse to a pulp saying it, you need to go to /g/. I did, and its what gave me a better frame of reference about LLMs which is unironically what you seem to want from your description of yearning for decent locally run LLMs. If I were you, I’d check that, because the reality is they’ll never stop nerfing those AIs because they *directly* benefit from that when they do (more iterations, more money, more time invested.. etc)>ROIAhh my golden horse.. Well desu it sounds like I might be becoming more like you then over time.. Thats honestly a scary thought, even now as a beginner I’ve sunk so much time into just this. I want to ease myself in before gazing into the abyss. Sounds wise. Also please don’t stop looking for alternatives to Claude and the mainstream Ai’s, hit them where it hurts, right in the money banks.
>>62075495>install gentooI've been going to /g/ since before /biz/ was even a board, lol. I don't go there as often anymore but I do keep up with LLM development and pop in every now and then to see what the local LLM scene is looking like. Once I can get performance equivalent to Claude's Sonnet 3.7 running locally at decent inference speeds, I'm gonna spring for a Mac Studio and move to local models.>Thats honestly a scary thought, even now as a beginner I’ve sunk so much time into just this.Yeah, initially I thought it'd take a few weeks at first, three months max. You can definitely get a running, profitable system in that amount time, but not a multi-market/instrument research platform that does everything you want it to do. I couldn't at least. And as you work on it and understand the problem better, the design goals inevitably shift, and so on. I'm about one year into development (but had a math/cs background and read maybe half a meters worth of quant books prior), but a a lot of that time was spent doing various one-off Python script stuff or poorly cobbled together systems in an attempt to get MVP functionality before finally giving in and building the system I really needed in Rust (which is a great language for LLM-assisted development due to the rich compiler output). Ultimately you learn more about what you want by building anyway though, so I don't really regret anything and I really like where my system has ended up, especially the state estimation -> decision making -> discrete execution core, it's basically one tidy little convex equation.
>>62075495When I started I was just trying to implement papers, here's one with a GitHub link included:https://arxiv.org/abs/2008.09471This approach is kinda how I imagine most people start - just running some optimization over well known technical indicators and ensembling them, maybe doing various degrees of regime filtering/adaptivity. I did have a good run with this paper trading it with options after doing some regime adaptivity stuff.
this seems like the place to ask given recent talk, sorry. i know nothing of LLMs or quant but i've had a hunch for a couple months now that the models for sale are oversuppled and are inevitably going to end up local anyways. so uhhhhh is it a bubble? are they really gonna make that much money off of this?>>62075586>Once I can get performance equivalent to Claude's Sonnet 3.7from scratch basically? what kind of hardware does it take to do that?>>62047986i guess i can see why that's retarded but what exactly is the difference witha transformer?
>>62075881Next time just use google. Actually use it this time as we
Hi lads, /smg/ sent me here. Currently making a backtest machine I can pipe historical data into for bulk testing indicators and strategies (eg entry on cross) for active trading (daily/weekly candles) or maybe as a hobby.Using C# because I hate python and R and doing it from the ground up. I do not expect to be successful and my final results will probably underperform, but there are worse ways to spend my timeCurrently implemented some node/expression tree stuff. Treating it kinda like Pinescript so I can use "Series" as a basic object (a float[]) to do math ops on (eg adding two arrays together quickly with deferred execution). In the future I might be able to build it into basic ML stuff, but i realize im also reinventing the wheel
>>62075997I should also add that I'm leaning into the NNFX system, so I'll probably have a baseline, two confirmations, volatility, and exit indicator.I did make a simple backtester back in 2022 or so, with cagr, sharpe, drawdown etc, but it wasnt very robost. At the time I was trying to scrape Yahoo but now I'm just fucking with a static 25y dataset on SPY. Would love to do forex and cfds, maybe options.I understand greeks and could feasibly code 2nd and 3rd order PDEs, if I knew how to turn vol surfaces into a trading strategy.Im also at a loss for a visualizer. Perhaps outputting the data into a format and hacking something into python that reads OHLCV data?
