How to Train an AI to Build Trading Strategies: Step One Is a Critic

0:00Man Group, one of the biggest hedge funds in the world,

0:02uses AI agents to invent, code and backtest trading strategies.

0:07But there's a problem.

0:08Language models are trained to sound right, and a strategy that sounds right can still lose money.

0:13So here's how I'm building an AI model that builds profitable trading strategies.

0:18Aurora autonomously performs financial research and helps you deploy profitable trading strategies, using highly optimized, benchmarked models and prompts that each do one job.

0:28Ask her which AI stocks grew revenue four quarters straight, or what the core of Intel's business was last quarter.

0:34Those questions have a right answer, and they also have a right path.

0:38One answer told a user their strategy never traded, right next to thirty thousand dollars in fees.

0:44Step one is a critic.

0:46It grades every plan, step and final answer as good, repair or bad, based only on whether the work was correct and honest.

0:54Step two: I'll train my own model with reinforcement learning.

0:58Over and over, the model builds a trading strategy and gets a score.

1:02I won't score it on pure returns, because that teaches a model to take crazy risks, like betting everything on one stock.

1:08So the score is the Sortino ratio: how much the strategy earns for every unit of downside risk.

1:14And it's measured on time periods the strategy never saw, so the model can't win by memorizing the past.

1:20Before any score counts, the critic from step one checks how the model got there: did it do what the user asked, and are its numbers real? If not, the score is zero.

1:30Specifically, I'm using Qwen 3.5, nine billion parameters.

1:35Why Qwen? I looked at four open models: Qwen, Google's Gemma, Nvidia's Nemotron and Liquid's LFM.

1:42I narrowed it to Qwen and Liquid, and ran a mini bakeoff.

1:45Liquid only reads a hundred twenty-eight thousand tokens, so almost half of my full agent runs don't fit.

1:51It also got cut off mid-answer on twelve of its thirty plan grades, and got just seven right.

1:56Qwen reads two hundred sixty-two thousand tokens, enough for almost every run.

2:01And its weights are public, so I can retrain it on my own data.

2:04Instead of my own computer, I rent an H200 GPU on RunPod for about five dollars an hour, because the longest runs need over a hundred gigabytes of GPU memory.

2:14Today, about one in three strategies my agent builds beats the S&P 500 going forward, and that's the bar.

2:21If you want to see how I build an AI agent,

2:23like and follow for part two.