Hey HN,
For the past few years, I've been obsessed with trying to get AI to produce accurate trading signals. The main challenges I found with AI trading models are lack of consistency, context window bottlenecks, hard to backtest, and high cost.
Asking ChatGPT "Should I buy Bitcoin today?" doesn't work well because the LLM doesnt have a set trading strategy to opperate from. In addition, the small context window makes it challenging to fit enough historical data into. Not to mention it gets very expensive as well.
My solution is a hybrid approach. Instead of having the LLM make direct predictions, I use AI as the "conductor" in the system that has the ability to run a series of backtesting simulations on high thread count AMD EPYC servers.
The AI then processes the results, identifies optimal parameter changes, and can self-optimize after every trade (as needed) to adapt to changing market conditions. This setup uses LLMs for high-level reasoning and CPU cores for what they do best.
I should mention that TrendFi is designed as a trend based model. It's not for day trading or catching every small market fluctuation. Its goal is to accurately identify major trend changes and provide clear alerts for those key moments.
Check it out: https://trend.fi
– Michael
The market is a battlefield of adversaries. For someone to cash out on a signal profitably, the others have to get less. A signal system is then easily abused by the signal maker who takes a position before hustling a bunch of people to take it. Or it is abused by a hedge fund or market maker who does the same and raises spreads on you when you are trying to make trades, etc.
The usual response applies: if you have profitable trading signals, why would you sell it? Every time you sell a copy, that makes it less effective.