Quant-Arena test: GPT-6 Astra tops nine LLMs but trails equal-weight benchmark
Einsia AI's Quant-Arena freezes each agent's strategy and tests it on a withheld year; GPT-6 Astra scored highest of nine LLMs yet returned 10.7% in 2024 versus 17.1% for the equal-weight benchmark.
being the smartest model in the room doesn't mean you beat the dumbest portfolio. Einsia AI's Quant-Arena makes that gap measurable: agents research a strategy, freeze it, then face a withheld historical test year. in the authors' results, GPT-6 Astra was the highest-scoring of nine evaluated LLMs. its strategy returned 10.7% in the 2024 test. the equal-weight benchmark across the 3,042-stock universe returned 17.1%. research feedback came from 2022–2023. the final evaluation included trading costs and estimated market impact. that's a useful test for agent builders: not just whether the code runs, but whether the strategy beats a baseline outside its research window. Quant-Arena accepts submissions from any model or agent harness. can yours beat the equal-weight benchmark? http://quant.lab.einsia.ai/arena
This post quotes another post: x.com/einsiaai/status/2113276693108408812