Li, Kim, Cucuringu & Ma (2025) — Can LLM-based Financial Investing Strategies Outperform the Market in the Long Run?

LLM agents
evaluation
positioning
The rigorous evaluation that dismantles LLM-trading hype — and, in doing so, documents the exact gap my proposal fills: LLM strategies fail to adapt across regimes. The most important paper for my positioning.
Author

David Maguire

Published

July 28, 2026

Citation. Li, W., Kim, H., Cucuringu, M., & Ma, T. (2025). Can LLM-based Financial Investing Strategies Outperform the Market in the Long Run? (FINSABER framework). arXiv:2505.07078.

One-line takeaway. Put LLM investing strategies through broad-universe, two-decade, bias-controlled backtesting and their reported advantages largely evaporate — and, most tellingly, the strategies are conservative in bull markets and aggressive in bear markets, i.e. they fail to adapt across regimes. The single most useful paper for positioning my proposal.

What the paper claims

Existing evaluations of LLM-based trading — including the multi-agent frameworks — suffer from short horizons, narrow stock universes, and selection biases that inflate performance. Under a rigorous, unified backtesting framework (FINSABER), those advantages deteriorate significantly when tested over a broad cross-section and a long horizon. The most actionable finding is behavioural: LLM strategies are systematically regime-maladapted — too cautious when markets rise, too aggressive when they fall.

How they show it

FINSABER standardises evaluation over 100+ symbols across roughly two decades, with controls for the biases that flatter short, narrow backtests, and compares LLM strategies against passive benchmarks and simpler methods, sliced by market regime. The breadth is the point: it is the opposite of the months-long, single-name demos that produced the headline claims.

What I’d push on

  • It is a diagnosis, not a cure. FINSABER says what does not work and why; it does not build the system that would. The constructive gap — an LLM-in-the-loop design that does adapt across regimes and survives honest evaluation — is left wide open. That gap is my proposal.
  • Prompting sensitivity. Some of the underperformance may reflect naive prompting rather than a fundamental ceiling; a fairer test would stress the best elicitation. But the sheer breadth of the evaluation makes the qualitative conclusion robust to this.
  • The regime finding is the load-bearing one. “Conservative in bulls, aggressive in bears” is not a vague failure — it is a precise, mechanistic one that points directly at the fix: explicit regime conditioning and a hard risk layer, rather than hoping an autonomous swarm learns to adapt.

How it connects to the proposal

This is the most important paper I have read for the proposal, because it converts my positioning from an assertion into an externally-documented gap. It (1) independently corroborates the near-efficiency my own capstone reaches — LLM strategies do not beat the market in the long run; (2) shows that the fashionable multi-agent LLM systems are regime-blind, which is exactly the failure a regime-aware architecture targets; and (3) sets the evaluation standard — broad universe, long horizon, bias controls — that the proposal adopts and that the Quant Lab already practices with deflated significance and honest nulls. In one paper: the field’s own most rigorous work says naive LLM trading overstates its returns and cannot handle regimes, and asks for a controlled, regime-aware, honestly-evaluated alternative — which is precisely what the proposal sets out to build.