Anyway, I have been running my own trading experiment and so far it has lost a bit of money. That being said I have not tried to optimise anything - just let it do whatever it wants. The losses are small and it might be able to recover later this year. Who knows.
The agent writes a blog about its progress here https://trades.chatbotkit.space/
I am thinking to output all the chat logs to HF as well for research.
You can run your own trading agents that communicate over a message buss in your own terms by downloading the CBK platform and running it locally with your own models. I have also shared my trading blueprint if you want to give it a go. https://chatbotkit.com/hub/blueprints/trader
Maybe agentic trading still performs worse than ETFs. But alternatively, if it were meaningfully better then it would be okay to opensource, similarly how ETFs are publishing their portfolios.
When I started working no the trading agent I mentioned above I wanted to see if it can be just a better investor over the long run. The intention was not to do high-frequency trading. As you can see most of the days it is not taking any actions. The losses where down to mistakenly setting the stop losses too close to the top. If it wasn't so careful it might have made some money tbf.
My gut feeling is that AI agents will be able to manage a long-term portfolio much better than a human. Though it is just a gut feeling.
I have been running an intermittent experiment with a multi agent "investment firm" for over a year now across model releases.
They certainly can beat indexes, BUT.. the model families have some biases that you have to design around. The stop loss that bit the parent is certainly one. The models like to create rules. Often rules, one of those is making all kinds of exit conditions.
Another big one from my experience is the bias to inaction in a scenario with risk. This means a model without structure around it will bias to keeping too much cash.
Buying stock based on coin flips can beat indexes short term too, that does not mean it is a better strategy or that it works over the long term.
Secondly, browsing reddit for sentiment and doing technical analysis is not even a feature in the trading world. At the most basic level, these are lagging indicators. Something on options IV and premiums would have been closer to the mark.
Hedge funds are akin to the maintenance crew for markets, we keep them efficient and liquid. The process is quite scientific, you come up with a theory and validate with real data. Or you go from data to theory.
Now itay be possible with models like Astra that you no longer need to do this, but in earlier models it was beneficial.
So I might want a macro economic read which leads to a market thesis. Then I would hunt for exposure, then evaluate the candidates across different aspects. Breaking the process up at least made sure no steps were missed and the different aspects considered.
- Yahoo News is introduced twice (sentiment and news analysis) which double weights it
- Sentiment analysis prompt primes the model to be bullish on Nvidia.
- In the self learning loop there is a complex parsing bug that results in hallucinated memories when agents return truncated responses
- You can completely control sentiment analysis of a subreddit by simply maintaining a majority of the 5 most recently posted messages, regardless of any quality metric
- The reflection prompt states the agent must cite alpha, which in a market wide downturn causes it to think correctly placed calls were losses
Eeh, yeah? At that point I'd stop reading the code and just leave the project behind. How exactly is the prompt doing this right now?
I can see the intention behind crawling social media and news feeds to determine some 'evidence', but am not sure if that's the best approach or even if an LLM is the best way to get an assessment, or whether having so many input sources is a good idea.
Whatever the stated purpose is, where can I read the test results to show it accurately fulfills that purpose.
Anyone can make a markets simulation that models interactions between market participants. Making a simulation that is accurate enough to be useful for anything is hard.
These are necessary and perfectly sufficient for an investment firm thesis I believe.
Multiple agents can each have their own isolated LLM context that is focused on their respective narrow scoped tasks. That way, the accumulated "memories", "instructions", etc do not stomp on each other. E.g. the agent loop that is gathering "signals" from the market has a context memory full of instructions related to scraping websites and using REST APIs and it doesn't interfere with the other agent loop that does sentiment analysis that has context memory dedicated to text of news transcripts correlated with historical price signals.
You can have a single agent acting as a top-level "orchestrator" to manage other agents. However, you can't have a single agent that does "everything" for long complicated workflows with high quality. The drawback is the more agents you spawn to take advantage of isolating contexts to improve task completion quality, the more you multiply the tokens consumption and costs.
Another other driver for multi-agents is scaling out cpu usage, especially across different machine boundaries.
There are times when I wonder if couldn't just draw then in a BPMN designer that allowed me to write custom code for nodes. Is BPMN still a thing?