This is very cool, but who is the ideal customer here? I used to work at one of the top tier shops and we had multiple teams whose entire responsibility was building and maintaining our simulation environments.
We are selling to frontier labs, e.g. OAI and Ant, and fintech companies that are trying to build agentic trading system but do not have those internal quant setups, e.g. Coinbase, Kalshi etc.
I'm skeptical frontier LLMs can actually do well (e.g. alpha 5%+) without fine-tuning, especially on historical market data. Presumably you support fine-tuned models?
cofounder here - LLMs can do some model training, they train on ML competition data after all. But they do struggle with low signal to noise ratio of market data. But that’s exactly what our environments will teach.
>Quant Trading RL Envs to Teach LLMs Research
Oh my Current Thing. This this enough current things?
languagelearner, I think you need to spend more time learning languages
One thing I always think about whenever someone talks about solving investment is "and then what?"
Say there's a crystal ball, wouldn't everyone use such crystal ball? Wouldn't crystal ball become illegal? Wouldn't crystal ball nullify the effects of things?
What am I missing, can someone from this field educate me on how this stuff scales?
this seems to be a common misconception, our envs use market data, but the goal is not (only) trading. Market data just happens to be a good source of hard data science tasks.
Re trading: I’d argue there is no such thing as solving investment nor is there “the one profitable strategy”. Every decision from personal risk appetite to trading horizon changes what is the optimal strategy for you and there are multiple strategies that make money.
Also note that even the most profitable alphas are no crystal balls. Someone else mentioned 5% correlation to future return - depending on horizon and data such level of correlation can make 9 figure PnL and is by no means easy to achieve
What's the margins that makes this worth chasing then? That's the part I maybe don't quite understand, why would you pour a lot of money and resources into something that is stochastic at best?
Quant trading is an extremely high margin business itself. Quant shops are printing billions and have on average much higher profits per employee than tech companies. So it is definitely a business worth doing.
On the other hand, you could also view quant research as some very hard research problems, so training LLMs on these problems could also enhance their general research capabilities.
Here is an example rollout trace with gpt 5.6 luna. Checkout if you are interested what kind alphas the agent found XD https://hub.harborframework.com/jobs/af0299f9-a3bb-44ea-8ced...
I looked through the transcript/output of the model/run linked but didn't find anything that showed much, if any, alpha. Maybe I missed it?
not sure about your background, the trace shows the feature engineering the LLMs did
It is the raw trace, so it is the most complete records but hard for human to read. We showcased some features they found in this research blog post: https://edotenv.com/blog/alpha-autoresearch
I'm not fully clear on this. Is this a quant trading benchmark for LLMs or a RL env?
It is both. We can use the same setup for both RL and Benchmarking.
Is it STRICTLY for LLMs & quant research or does it do generic trading simulation?
It is not only quant research. We also have live trading: https://edotenv.com/blog/long-horizon-planning
For now it is only facing LLM/Agent.
if the data is not synthetic, how do you ensure that the LLM hasn't learnt about this data for example from training on the Financial Times.
We do a 2 step anonymisation: 1. Mask all symbols, timestamps etc. So the agents cannot infer the assets/time periods. 2. Mathematically transform numerical values and returns. E.g. the market return targets are not the raw market returns, but neutralised and manipulated. So even the agents have certain bullish/bearish biases, it cannot make use of it, as we use the transformed values.
In addition, we did not observe such behaviour in our traces. An example: https://hub.harborframework.com/jobs/af0299f9-a3bb-44ea-8ced...
ah i thought so, interesting. I think the challenge is to do 2 while still keeping it realistic, which actually gets very close to synthetic data generation.
we do affine transformations of the data, so all return/ pnl measures are still the same as with untransformed data. The transformation doesn’t change the conditional distribution of the data, which is what alphas ultimately measure