LLM Trading Agents Earn Beta, Not Alpha: A Look-Ahead-Free Benchmark
arXiv — cs.AI ·
What to know
A new study introduces rta-bench-1, a backtest benchmark for LLM trading agents across three US markets, finding that in 91 runs of six models from four vendors, no alpha was significant and market/style exposure accounted for a median 78% of returns. The agents' decisions performed no better than random stock picks from the same sector, and ablations showed they earned as much without news headlines. The study recommends builders supply their own edge and report exposures and alpha, and benchmarks should replay news by publication time.
Source and evidence
arXiv — cs.AIOriginal language · und ·Original author: Jackson Wang · Source: arXiv and the named paper authors. Descriptive metadata: CC0 1.0.Source licence
Why it matters by industry
- Artificial intelligence
This benchmark directly addresses the evaluation of LLM agents in financial trading, a key application of embodied intelligence and reasoning, providing evidence on the limitations of current models and guiding future development.
Reviewed summary. The complete article remains at its original source.