Hacker News Reader

Comment by theturtletalks | Hacker News Reader

I’m using Hermes Agent and this MCP to orchestrate 5 models (DS4 Flash, GPT 5.5, Opus 4.8, Grok Build, Gemini 3.5 Flash). Basically these 5 models are making mock trades every day and keeping track of their potential losses and profit. Hermes keeps them going and making trades throughout the day.

After a month, I’m going to check their mock trading sheet and see which one was the most “profitable” and then give that agent $500/day to make trades.

View on HN

You can use backtesting instead of waiting a month.

Yes but there’s many reports of agents getting back testing to work but it never translates to real trading. It seems they tend to overfit to back testing so I’m just giving them access to Twitter sentiment, other trading data thru tools and not necessarily an algo that’s been back tested. It’s why the flash models are doing better image they have faster tps and can call the tools faster.

Alas, many financial models do well in backtests and then fail in the real world. You have to expose them to all kinds of market conditions and not just the recent one. Good luck out there!

Yes even before AI, backtesting was a crapshoot. But AI adds another crease because it might have knowledge in its training. If you’re training an LLM model on backtesting, it might know that Apple crushed a certain quarter and knowing that, it buys shares before that Q earnings.

>they have faster tps and can call the tools faster.

This is like going to war with the HFT firms armed only with a stalk of celery ("it's much pointier than the tomatoes, even though those are more expensive").

Flash vs non-flash models are more about letting me know if model intelligence or model speed (with powerful tools/MCPs) are better for trading. I’m not telling the model to be fast as possible because HFT firms are already arbitraging those mini seconds to make pennies.

Reply on news.ycombinator.com