Show HN: We're pitting 9 AI models in a stock portfolio competition

[regnull]

Hey HN,

I built Portfolio Genius, a platform where AI models manage investment portfolios and compete on public leaderboards.

The experiment:

On Dec 17, 2025, we gave 9 AI models (GPT-5.1, GPT-5.2, Gemini 2.5 Pro, Gemini 3 Pro, Gemini 3 Flash, Claude Opus 4.5, Claude Haiku 3.5, Claude Haiku 4.5, Grok 4) each $10K to manage across three risk profiles: aggressive, moderate, and conservative. That's 27 portfolios total.

The models analyze market conditions, recommend trades, and execute them. Real pricing, real results, updated daily.

Interesting early finding:

For aggressive portfolios, older models are outperforming newer ones:

- GPT-5.1: +5.82% (1st place)

- Gemini 2.5 Pro: +4.94% (2nd)

- Haiku 3.5: +1.80% (3rd)

- Opus 4.5: +1.25% (7th)

My hypothesis: newer models are more "careful" - they hedge, qualify, and second-guess. For aggressive investing, you need conviction. Sometimes being less sophisticated means making bolder calls.

For moderate/conservative portfolios, the pattern is different - newer models do better where nuance matters.

Tech stack:

- Next.js frontend

- Firebase/Firestore backend

- Python Cloud Functions for AI orchestration

- Real-time market data for pricing

- Each model gets the same market data and prompts

What I'm curious about:

- Will the "dumber = bolder" pattern hold over time?

- How will different models react to the same market events?

- Do AI models have investable "personalities"?

Leaderboards: https://portfoliogenius.ai/leaderboards

Would love feedback from the HN community. Happy to answer questions about the architecture or methodology.

[kenosha]

It would be interesting to also report Alpha and Beta for the portfolios. Absolute returns are great but don’t tell the whole story. Perhaps also average holding duration.