I Built a Free LLM Benchmark for Portfolio Decisions. The First Results Surprised Me.

Most LLM benchmarks ask models to answer questions, write code, or solve puzzles. I wanted something messier: can a language model make repeatable investment decisions when given the same market context as its peers? So I built a benchmark that turns LLMs into portfolio construction engines. QUINETICS - LLM Investment Benchmark The setup is simple on the surface. Each model receives the same market snapshot, selects a long-only equity portfolio from the S&P 500 universe, and the resulting allocations are passed into a deterministic backtester. The leaderboard then ranks model/prompt/rebalance combinations by Sharpe ratio, with return, excess return, benchmark performance, and max drawdown shown beside it. The benchmark is not trying to prove that any model can “beat the market” in production. Backtests are fragile. They are useful, but dangerous when treated as prophecy. What I wanted was a controlled arena: same rules, same universe, same rebalance cadence, same scoring system. Here a