CryptoPainter

I spent the last couple of days completely restructuring the entire system and wanted to share some new lessons learned: 1. When building a strategy system involving AI, for every deployed feature, you must first ask yourself: "Can this feature be implemented purely through code?" If a feature can be implemented purely through code, no matter how complex the logic is, you should avoid handing over decision-making to AI; For example, my current overall strategy system architecture is to first deploy a set of pure algorithmic quantitative strategies with high win rates and high profit-loss ratios but extremely low trading frequency. Backtesting results show fewer than 10 trades per year, but the data is excellent. The reason for the low trading frequency is that the strategy’s decision tree includes many filtering factors, so normal market conditions rarely trigger entry signals... Then I involve AI in decision-making, using account historical data and the strategy architecture to dynamically adjust the filter parameters, turning a pure algorithmic strategy into a flexible trader... Then problems arose. Although the AI’s prompt instructions were comprehensive, as the number of parameter iterations increased, the AI began to develop some self-reinforcing directional biases... For example, last time it adjusted the RSI range and found that the strategy’s win rate and profits improved. This change was then logged as an iteration experience for the next optimization. Gradually, the context became that no matter how the strategy performed, it would only modify one core parameter, and the adjustments became more and more extreme... Even though I added explanations for all exposed parameters in the global prompt, after running for a while each time, the AI’s attention would get stuck on one or two parameters and couldn’t break free... Therefore, I had to spend two days turning this part of the AI task—data collection, real-time analysis, and dynamic parameter tuning—into pure code logic. To adapt, the entire system required a lot of dependency development, which was completely unnecessary when AI was making decisions before, as it could pull data, search online, perform semantic analysis, and then draw tuning conclusions on its own... Still, without heavy AI involvement in decision-making, the system finally stabilized. So one thing is clear: for trading systems, hard-coded logic is ultimately superior to flexible large models! Currently, I have limited AI’s role to social sentiment analysis, trading asset collection (background checks and unlocking periods), and system monitoring. Having the Agent act as a secretary is far safer and more stable than having it act as a trader. So, if you’re still expecting to give AI a prompt and have it start making money for you, it’s better to forget it for now. Current large models can only execute pure algorithmic strategies with fixed parameters, essentially saving users the trouble of initial development; But the more AI occupies space in a system, the higher the black-box uncertainty becomes. Sometimes it’s even more stable to just find a signal provider to follow...