CryptoPainter

Spent 2 hours buying a domain and made a Dashboard on the website with all Agent's trading data... Today is a bit interesting, the strategy positions are all short, let's see if there's hope to dump some altcoins on Friday. Currently, Agent's overall account has already lost 8%, sigh... If there's anyone to blame, it's me for constantly changing features and fixing bugs every day. Previously, I optimized strategies using backtest data, but this time I want to try pre-testing with live trading data... The former has perfect backtest data but performs poorly in live trading; the latter performs poorly in live trading but with continuous optimization, the strategy is getting better and better... I won't disclose the website domain for now, I'm a bit worried about security measures. I'll wait until Claude audits it for me and the strategy starts making profits before saying more...

Someone asked me why I haven't updated that "genetic algorithm" engine? I actually want to update it! After setting up the framework and perfecting the features, the next step is full-scale training. But the problem is, look at the log below: training one tree requires 80 iterations, nearly 1 hour. There are 6 trees in one Forward window, and the entire sample set is divided into 6 windows for Walk Forward... 1h x 6 x 6 = 36h, and this is only the first layer of volume-price data training. After that, it moves on to the macro data layer, and finally the futures, on-chain, and other Alt-Data layers. Running all of this probably takes about 3 days... And the most, most, most critical issue is that, when luck is bad, the factors randomly injected into the factor library fail to hybridize into excellent expressions, resulting in the current screen full of "Failed" statuses. This means the trained factors perform well in-sample but fail out-of-sample... These are basically overfitted factors and get directly PASSED... For a whole week, I've been stuck in this tedious loop, feeling like my alien laptop is going to be wrecked sooner or later... So, it's really not that I don't want to update, it's just that I honestly have nothing to update...

Is it true this time, or is it different?

While researching new strategies for the Agent and configuring a pure algorithmic state machine, I suddenly realized this mechanism could also be applied to the previous ASR strategy. So I hurriedly started modifying the code, and then I was moved to tears... An old strategy that hadn’t been successfully optimized forward for a whole year suddenly came back to life! As for the specific changes, it’s just using the state machine to record market volatility in real time, then fine-tuning the volatility into the original strategy’s parameters. Altogether, it’s less than 20 lines of code, but it made the original ASR channel differ in many subtle details... The overall return of the 5-year BTC strategy improved by over 75%, while the maximum drawdown decreased by 14%!!! Previously, when researching pure algorithmic quantification, I looked down on state machines. Only when it truly produced positive optimization effects did I realize how great it is... I guess an updated version can be released soon! It’s really been so tough...