Founder: @Dither_Solana @tokens_terminal
Joined May 2021
- Tweets3.5K
- Following726
- Followers4.1K
- Likes24.4K
It was today. Why price go down? $SOL bad price
Looks like someone is trying to get at my X password.
If I'm tweeting about something other than $DITH then I've been compromised
I tend to have strong opinions loosely held (or I strive to).
I think this next report is the first time an AI has strongly influenced how I think about a few projects
This is a good read and something A LOT of people should bookmark. Most "I used chatgpt/claude/GLM to trade" are making huge mistakes.
Assuming they have data other people don't (big assumption) then they also need to NOT overfit. This is no small task
This is *not* how you use AI in trading. People actually doing this will get rekt.
AI has made backtesting fast and accessible to anyone with data. What it has not removed is your need to judge the output and catch the mistakes made along the way. And that's the trap, because not all friction is the same. Using AI well is mostly one skill: telling apart friction that only costs you time from friction that is actively training you. The key question is which friction you need to keep and go through yourself to be able to judge the output. This way you learn about lookahead bias, overfitting, etc. That's the intellectual debt you accumulate when you let AI handle the whole process.
People who outsource all of it make AI torture historical data until it confesses to anything. I've literally watched people backtest a strategy, find that it underperforms on certain days of the week, and simply delete those periods from the sample. To them, it looked like refinement, but it's quite the *opposite*.
Which is exactly what Step 4 tells you to do. "Analyse this trade history and tell me how to improve this strategy" is an instruction to overfit. There's very little fitting the average person can do before fucking things up big time. In practice, AI is just amplifying the same old biases with a lot more computational power and speed. You just make them at scale and get rekt quicker.
That being said, backtests are in themselves very tricky, mainly due to path dependence, which means your results aren't just about whether your strategy works.