ML
MoneyLion Inc.
stockNYSE

A lot of things, like order getting squared off by some strategy, are reflected in some other strategy. Some scripts might hang on the platform because of infinite loops running. But I have tried to cover every corner case and make it rock solid and architectured in a way that it can't go bad in most of everything that you see in vibe-coded apps. I do not claim that it's bug-free, but if you consider algo trading, then I consider this to be the best tool I can see so far. Here you have no limitations from basic buy-sell strategy to have an ML model trading for you. For anything you can think of, you can convert your idea in seconds on the platform.

I don't do ML or QIS anymore, but yeah back to the drawing board and look for better ideas/something with a better rationale otherwise if you keep tuning you'll just overfit the backtest on the test set

In all seriousness, there is nothing special traders offer that someone can’t just hack together and figure out. I continue to maintain a position on the Kalshi leaderboard without having any prior trading / betting experience. It has been over a month, and around 2m in volume. ML has dominated trading for decades now, so stomping your feet about this and saying “nah uh” to reality is not the best look.

I guess I just disagree, or maybe think that this perspective has to be one from a person who didn't understand the ML landscape before this. Compared to what was available when the term AGI was coined, basically every LLM we have is generalized. The models that were around when AGI was coined were able to do things like identify an object in a picture, then chatgpt came along and was generalized to a whole bunch of different tasks. At that point, I'd say we have reached AGI. Any frontier model can easily use the same tooling a movie producer would use, and we already know they are basically superhuman at math. It just seems like a really tiring and repetitive argument where the "never AGI" will keep moving the goalposts until they actually see the face of god. I actually think the trump term fits what you are talking about now - superintelligence. I think mostly people have a hard time coming to terms with the fact that intelligence can emerge from language alone. It cheapens our own intelligence, which people don't like.

developed by me and top researchers at tier-one firms. We’ve run several mathematical and ML models to develop the perfect formula to calculate YOUR odds

At least from my experience (midfreq macro/statarb) we find that adding ML into the final signal aggregation layer maybe gives a 15/20% boost in performance. But TBh you can throw massive complexity at a set of dogshit features and not really get anywhere. At least that is our philosophy; we put most focus into a set of high quality signals to start with. Some ppl will take a more bottom-up approach

I mean not necessarily, I personally know at least two principal engineers (not ML/AI researcher) at Meta/Amazon that are clearing well over $800k (the original salary in the posting)

AI/ML

Hey op thanks for the info. Can you tell about your background? I am a quant dev at a hedge fund with 4Y exp but want to pivot to research engineer roles in ML so maybe I can work in AI labs too

LLMs don’t drift, and neither does ML inherently. Drift only happens in ML when the data you are inferring against changes relative to the training data. If the underlying data never changes, then there’s no drift. Drift for an LLM would be like if we never asked it a question in Chinese, and then we started to. They’re burning so much money on R&D because they’re in a space race to constantly have the best model. At some point the base training will be good enough that a specific model will be good for a long time. That’s basically the whole business model. They’re hoping they can get to that point before they go bankrupt.

Hang in there, bud!! Clean energy will come roaring back, including a significant AI/ML component too. #:')

I used it in the early 2000s. Lots and lots and lots of OpenCL programming. The few friends I chat with from those days say that the best money comes from ML, and the most consistent from HFT and its related ilk.

Just curious why is NN so popular in hft? Why not other form of ML

Those GPUs are because they all build large transformer models ex like time series foundation models on tick data.. Every modern HFT/MFT/LFT alpha model is a ML model, tree or deep learning… OLS is just one of your statistical tools to get your methods correct

As the title mentions, I wanted to get more insights into the use of machine learning in quant finance, especially in alpha research/signal generation. I used to read on this forum that the vast majority of systematic trading strategies were based on linear regression. However, for years now, some major players (XTX, JS, ...) have been investing massively in GPUs. My question is: Why are these top-tier firms spending hundreds of millions on chips if they are mainly using OLS/Ridge/LASSO? Are these chips only for internal LLMs? Are DL and ML really a myth when it comes to alpha generation?

I see, is the OMSCS worth it too? Not super interested in the Computer Engineering aspect, more so just the AI and ML part. And is the QCF program still as good as your comment above? I didn’t get in for spring so I’d be applying again for next Fall

Why do ML engineers care about whether mounts can fly?

I work in an area where we need compute. We cannot find availability anywhere. I've checked every AWS/GCP/Azure region over and over for it. Regardless if its for AI or other ML workloads, the demand is there

I don't know exactly how the market will shake out: there will be winners and losers and people have been placing bets all over the place. But as someone who's been doing AI/ML implementation since before the LLM era, these tools are not going away and will be baked into everyone's workflows the way email and slack/teams are. People who think this is a fad are going to be made obsolete.

from what I've seen of past top writeups, the winners mostly use simple stuff, not ML. the core loop is: estimate a fair value for each product, then quote buys a bit below and sells a bit above it, that's market making. other rounds usually have a basket vs its components (trade the spread when it drifts) and options, where plain black-scholes gets you most of the way. probability is basically how you get the fair value and decide if a fill is worth it. CP helps less than you'd think, clean python and your own local backtester that replays their logs matter way more. I'd start with the dumbest market maker on one product and get it running end to end. a lot of top teams posted full writeups and code on github after each round, search the competition name plus "writeup", those are the best tutorial there is

You're surely aware of the long history of IP theft, at astronomic economic levels, by the Chinese. Making decisions regarding the power of AI/ML in the hands of the Chinese is, I believe, way above either of our pay grades especially when it pertains to potential threats from them across a wide range of points of leverage. At least the Biden administration had real expertise, not the corrupt clowns and Luddites this regime has brought into play. And their rise to expertise in AI/ML is kinda like their penchant for ripping off brands. I started out in speech recognition in the 70's and have continued to work in ML since then. It was a very long time before you saw any contributions from Chinese R&D, so I'm kinda biased against "late comers to the game". I also believe our multi-cultural heritage has real benefits in R&D, and I'm not just saying that to be PC. In the short term, Nvidia doesn't even need their business because they can't fulfill the orders they have now (or so I'm told). We have Europe, India, and so many other potential customers coming online too, not to mention the rapid expansion of enterprise adoption here and a fork ton of startups that is hard to imagine, any number of them could be huge new applications and businesses. We don't have to apologize for bad uses of a new technology -- it would be the first time in history that this didn't happen. A very close cousin of mine, and a doctor too, was just diagnosed with metastatic pancreatic cancer. There are now new developments in personalized cancer vaccines that rely on ML for optimization of their efficacy, and would literally not be possible with selection of the optimum set of factors to be part of those vaccines. For every "Flock camera" I'm sure we could easily identify 10 positive uses. There's no going back, of course, as this technology has been in development in the 1940's (I'd developed a variant of Hebbian learning myself at one point).

I am first year bachelor student can you tell me which topics should I study for quant ML

Fair enough. But for QR internships, how much ML are undergrads actually expected to know during recruitment? Would basic ML knowledge and a few projects be enough, or do they expect considerably more?

I used zero ML or Neural Networks btw

What work have you don't to try and make the ML interpretable so the non quant folks can engage with it? Assuming its ML and not AI there are a lot of interpretability packages that are designed to help with this. But yeah, if you sell someone your black box I don't see how you can be surprised that they surprised by estimating the generative function by sampling from it? And treat you as the generative function of black boxes and evaluate you based on their estimate of the black boxes you are generating.